AI in Education VIP Research Exchange
A collaborative knowledge base for NYU's AI in Education VIP team.
Search 0 research entries
VIP Subteams & Research Areas
Understanding, predicting, and guiding the impact of generative AI in education through four complementary research pathways:
Evaluating Tools 0
Understanding Users 0
Assessing Impacts 0
Innovating EdTech 0
Background Resources
Learning Science 0
Research Methods & Resources 0
Understanding Users Attitudes & Behaviors
These resources help us to answer questions like: How, and how much, are students and teachers using AI tools? What values, principles, and assumptions shape their engagement with AI tools?
General 0
General perspectives on how people perceive and interact with AI systems.
Administrators 0
Administrative perspectives on AI policy, implementation, and institutional change.
Students 0
Understanding how students perceive and use AI tools in their learning processes.
Faculty 0
Faculty attitudes toward AI integration and their teaching practice adaptations.
General Perspectives on AI
General perspectives on how people perceive and interact with AI systems.
Sundar, S. S. "Rise of Machine Agency: A Framework for Studying the Psychology of Human–AI Interaction (HAII)." J. Comput.-Mediat. Commun. 2020, 25, 74–88.
This paper is useful for framing the psychological impact of increasingly powerful AI agents. The paper proposes a framework that measures two dimensions of agency in human AI interaction: "cognizance", or the perceived sentience and consciousness of the AI agent, and "autonomy", or the extend to which the agent can carry out tasks without prompts from a human user.
Alexander Landfair 12/2/2024
Faculty Perspectives on AI
Faculty attitudes toward AI integration and their teaching practice adaptations.
Celik, I. "Towards Intelligent-TPACK: An empirical study on teachers' professional knowledge to ethically integrate artificial intelligence (AI)-based tools into education." Comput. Hum. Behav. 2023, 138, 107468.
This paper applies the well-established TPACK framework to the problem of AI integration in education. TPACK is a framework to describe teachers' knowledge needs in terms of their content knowledge (CK), their pedagogical knowledge (PK) and their technological knowledge (TK). This paper proposes a new AI-TPACK framework, and surveys 327 teachers to establish baseline measurements.
Alexander Landfair 12/2/2024
This article focuses on the role of teachers in the educational integration of AI-based tools. The study explores the technological, pedagogical, and content knowledge (TPACK) framework in order to identify what knowledge is necessary for teachers to participate effectively in this integration. The team conducted a study where they created a survey and identified 5 factors and 27 questions that had the most bearing on judging how effectively a teacher would be able to integrate AI-based tools into their classrooms. The 5 factors were technical knowledge (TK), technical pedagogical knowledge (TPK), technical content knowledge (TCK), TPACK, and ethics. These questions didn't just focus on their knowledge of how the technology works; they also focused on their understanding of when and how these tools could be used to enhance their already established pedagogical methods (TPK), how the tools are being used by professionals in their field and how they can use them to further their own knowledge in their field (TCK) and finally their awareness of the transparency, inclusiveness, fairness and responsibility concerns of the tools. They then gave this study to a different set of teachers to find the relationship between the 5 factors.
One of the most interesting findings was that the more teachers recognize ethical issues in AI, such as fairness, the more they become knowledgeable about the integration of AI-based tools in education. This caused them to expand the TPACK model into Intelligent-TPACK, highlighting the importance of ethics in a teacher's role in AI integration. I thought this was fascinating because I think teachers want what's best for their students, and learning about these technologies to learn how to decide if they can be used fairly in their classrooms seems like such a great angle for raising interest. They also found that offering professional development that taught teachers how to use AI would help them come up with ways that AI could enhance their own established pedagogical practices. These tools offer a lot of room for innovation, and I think giving teachers the knowledge they need to be able to innovate on their own methods offers a much less scary relationship to AI than being told to use it in a specific way. At one point, the paper says that Intelligent Technologies should be regarded as more of a partner than a tool, and I think a lot of the TPACK principles can help teachers to learn to see it that way. The study concludes by hoping for future work that tests whether the Intelligent-TPACK framework can be used as a predictor of the quality of teachers' use of AI tools. There is a lot of focus on testing on the student side, so it was very interesting to see the development of a test for the teacher side.
One of the most interesting findings was that the more teachers recognize ethical issues in AI, such as fairness, the more they become knowledgeable about the integration of AI-based tools in education. This caused them to expand the TPACK model into Intelligent-TPACK, highlighting the importance of ethics in a teacher's role in AI integration. I thought this was fascinating because I think teachers want what's best for their students, and learning about these technologies to learn how to decide if they can be used fairly in their classrooms seems like such a great angle for raising interest. They also found that offering professional development that taught teachers how to use AI would help them come up with ways that AI could enhance their own established pedagogical practices. These tools offer a lot of room for innovation, and I think giving teachers the knowledge they need to be able to innovate on their own methods offers a much less scary relationship to AI than being told to use it in a specific way. At one point, the paper says that Intelligent Technologies should be regarded as more of a partner than a tool, and I think a lot of the TPACK principles can help teachers to learn to see it that way. The study concludes by hoping for future work that tests whether the Intelligent-TPACK framework can be used as a predictor of the quality of teachers' use of AI tools. There is a lot of focus on testing on the student side, so it was very interesting to see the development of a test for the teacher side.
Jesse Noppe-Brandon 09/30/2025
Fitria, T. N. (2021). "Artificial intelligence (AI) in education: Using AI tools for teaching and learning process." Prosiding Seminar Nasional Pendidikan, 3, 67-76.
This text came right before / at the forefront of AI being widely adopted and in populous common use in educational settings. It begins by outlining several use cases and existing technologies that are seen as 'innovations' that help save time and are efficient. The argument and clarity gained surrounds the philosophy that human teacher skills - delivering new knowledge and creativity is what should be valued, protected and continuously harped on. While systematic procedures that have historically been manual and necessary to do, might be what is significantly automated.
The research ultimately takes a deeper dive/ inspection into the various forms of currently emerging edutech, to create philosophical commentary on its advantages and disadvantages, especially in its form / medium. Fitria offers commentary especially under the lens of, again, helpfulness in systematic procedure, vs what might create a distance from interacting with lucrative human teacher value. It mentions several somewhat niche technologies too for educational settings like voice assistants, and language translation between teachers and students that are intriguing.
Its method and materials are mostly concerned with library research. That's a limitation.
Cate Hackett 10/9/2025
Park, Y., & Choo, S. (2025). "Prompt engineering strategies for educators: Optimizing AI-assisted teaching tools." Education and Information Technologies, 30(1), 145-168.
This paper explores different prompt engineering strategies designed to benefit educators and how they interact with Large Language Models (LLMs). It introduces a number of different frameworks designed for different steps of prompt engineering, with the most important being CLEAR, PARTS, and REFINE. Together, these provide a comprehensive approach for educators to engage with LLMs.
The PARTS framework stands for Persona, Aim, Recipients, and Theme. These terms are what you should follow when creating context for the LLM. Persona refers to setting the context of your prompt. Aim is setting the goal of the prompt and what you want the outcome to be. Recipients are who you want the outcome to be for as the AI can tailor its response for that audience. It's important to note to be very specific with this as the more specific it is, the more tailored the response can be. The next is Theme, where you describe the style, tone and any restrictions. These could include which voice to use, how the tone of the response should sound, and a word count. Finally, there's structure. This represents the desired format for the output. This can include, but not limited to, bullet points, a table, code, and a graph.
The next framework, CLEAR, stands for Concise, Logical, Explicit, Adaptive, and Reflective. This is how your prompts should be structured, as creating clear and concise prompts help the LLM follow precise instructions and allows it to follow a clear flow and order of ideas. While some of these overlap with PARTS, keeping prompts concise and explicit are definitely important to keep in mind.
The next framework is REFINE. This stands for Rephrase key words, Experiment with context and examples, Feedback loop, Inquiry questions, Navigate by iterations, and Evaluate and verify outputs. This comes after the initial prompt and output from the LLM. This framework is used to develop a new prompt to modify the output.
Together, these frameworks create IDEA, which stands for Include essential PARTS, Develop with CLEAR prompts, Evaluate outputs and REFINE prompts, and Apply with accountability. The last part refers to making sure you are aware of the potential limitations and to make sure to use LLMs with responsibility.
This research ultimately brings together different prompt engineering ideas to create an ultimate framework which I think should be taught to educators. Teaching these principles of prompt engineering can enable them to become designers of AI-Assisted learning, rather than regular consumers. By being able to adopt these frameworks, educators can develop somewhat of an AI literacy, the ability to understand and shape how AI produces information.
Rohan Kapur 10/23/2025
Student Perspectives on AI
Understanding how students perceive and use AI tools in their learning processes.
Li, M.; Hu, Y.; Hu, W.; et al. "An explanatory study of factors influencing engagement in AI education at the K-12 Level: from the perspective of students." Educ. Inf. Technol. 2024.
This paper aims to understand what external and internal factors might influence a K-12 student's engagement with AI. To do so, the researchers surveyed 1,022 K-12 students across more than 100 different Chinese schools using the UTAUT2 model, and complemented these surveys with interviews of 15 K-12 students. The paper identifies 8 factors that significantly affect K-12 students' willingness to engage with AI.
Alexander Landfair 12/2/2024
Nguyen, S. "Enhancing Student Engagement Through Artificial Intelligence: A Case Study in Higher Education." ArXiv 2024, abs/2408.08551.
This paper describes a case study of students using an AI Teaching Assistant (AITA) in the author's class on organizational behavior. The author notes that students found the AITA very helpful with some limitations. However, the paper does not control for the novelty effect, or provide much data on how or how much the students used the AITA.
Alexander Landfair 12/2/2024
Marrone, R., Taddeo, V., & Hill, G. (2022). "Creativity and Artificial Intelligence—A Student Perspective." Journal of Intelligence, 10(3), 65.
This article explores how AI is involved with education by looking at focus groups of secondary school children. Their findings are similar to our current class discussions about AI, but it is especially insightful because these focus on younger students. The results point to 4 buckets: technological, social, affective, and learning. Students believed that AI could harm social skills, making people less capable of real-world communication, while others saw potential for AI as connection if used collaboratively. Their affective responses varied, students who understood AI felt comfortable and trusting toward it, whereas those unfamiliar with it felt anxious or fearful. They often equated AI with robots and futuristic machines, showing a limited understanding of its everyday uses. Overall, they viewed AI as technically advanced but not truly creative, though it could inspire or "spark" human creativity. The researchers also proposed a 4AI model that mirrors the 4C models of learning.
Sabria Islam 10/9/2025
Administrator Perspectives on AI
Administrative perspectives on AI policy, implementation, and institutional change.
Schiff, D. (2025). "Education for AI, not AI for Education: The Role of Education and Ethics in National AI Policy Strategies."
The paper analyzes 24 national AI policy strategies to explore the role education plays in policymaking, stating that the strategies focus more on "education for AI" rather than "AI for education" (AIED). This means that the strategies mainly emphasize on preparing a workforce skilled in AI technologies through training AI experts, reskilling workers, and promoting public AI literacy, but lack discussions about the ethical and social implications of AI applied within educational settings. The author argues for bridging the current policy gap by encouraging policymakers to integrate ethical considerations explicitly into the design, deployment, and governance of AI in education. This article pushes for a dual approach of investing in human capital for AI and harnessing AI responsively to improve education itself.
Nasrin Alanna Aidi 10/23/2025
Tadimalla, S. Y., & Maher, M. L. (2025). AI Literacy as a Core Component of AI Education. AI Magazine, 46(2). https://doi.org/10.1002/aaai.70007
The authors of this research frame what it would look like to facilitate comprehensive teaching about AI, which I found to be a rare approach, but come to think of it extremely helpful. Why wouldn't we starting thinking about how we need to scaffold the education on these tools (and thus implementing)?! This framework would significantly educate the younger generations not only about all surrounding AI infrastructures, i.e. understanding their inner workings when it comes to computer science and machine learning, as well as "literacy" (literally how to effectively use it), ethics (think ownership and responsibility), cognitive learning and problem solving. What I found this research to be most helpful in was its defining and labeling of individuals' knowledge extent, and "well-roundedness", if you will. They strategically chart such labels for example "naive AI user", "AI ethicist", "AI engineer", "uninformed AI consumer" - just to name a few - based on the factors 1. Knowledge 2. Skills 3. Attitudes and 4. Social awareness ranging from technical to socio education. These definitions and frameworks get impressively granular and extensive in topics and detail, even proposing a 16 module lesson plan with at least 3 subtopics in each to cover.
I started liking the concept of the learning module (7) on prompt engineering - this is where I thought it would be leaning more toward what I was hoping for which is more about effectively using AI in learning and especially not just asking for a generation of body of work. This topic I felt this proposal failed at. Though, the rest of the modules, to me, felt to extensively and adequately cover policy, ethics, basic computer science, understanding models and engineering, and other more meta socio-perceptions and projections of AI.
CONCLUSION: Overall great opening to the conversation and foundation to being thinking of how to structure learning about AI - OPPORTUNITY to skew and expand upon it even more toward education and creating - - aka how can you teach about the learning and creating process with AI 😁?
I started liking the concept of the learning module (7) on prompt engineering - this is where I thought it would be leaning more toward what I was hoping for which is more about effectively using AI in learning and especially not just asking for a generation of body of work. This topic I felt this proposal failed at. Though, the rest of the modules, to me, felt to extensively and adequately cover policy, ethics, basic computer science, understanding models and engineering, and other more meta socio-perceptions and projections of AI.
CONCLUSION: Overall great opening to the conversation and foundation to being thinking of how to structure learning about AI - OPPORTUNITY to skew and expand upon it even more toward education and creating - - aka how can you teach about the learning and creating process with AI 😁?
Cate Hackett 11/5/2025
More resources on administrative perspectives coming soon!
Student Attitudes & Behaviors
Understanding how students perceive and use AI tools in their learning processes.
Cotton, D. R., Cotton, P. A., & Shipway, J. R. (2024). "Chatting and cheating: Ensuring academic integrity in the era of ChatGPT." Innovations in Education and Teaching International, 61(2), 228-241.
This study surveyed 2,500+ students across multiple universities to understand their perceptions and usage patterns of ChatGPT. Key findings include widespread adoption for academic tasks, concerns about academic integrity, and varying levels of AI literacy. The research reveals generational differences in AI comfort levels and highlights the need for institutional guidance.
Alexander Landfair, January 2025
Evaluating AI Tools
These resources help us to answer questions about the capabilities of generative AI tools, which tools perform better on certain tasks, and how well AI-detection works.
Coding and Math 0
Evaluating AI tools' capabilities in programming, code generation, and debugging tasks.
Creativity 0
Assessing AI's ability to generate creative content and support creative thinking processes.
Detection Tools 0
Evaluation of AI detection technologies and their effectiveness in educational contexts.
Ethics / Safety 0
Examining ethical considerations and safety concerns in AI tool deployment.
Equity / Social Justice 0
Analyzing how AI tools impact educational equity and social justice issues.
Error / Hallucination 0
Understanding AI errors, hallucinations, and accuracy limitations.
Reasoning 0
Testing AI's logical reasoning, problem-solving, and critical thinking capabilities.
Usability 0
Assessing user experience, interface design, and ease of use in AI tools.
Translation 0
Evaluating AI translation capabilities and multilingual support quality.
Tutoring / Instruction 0
Evaluating AI tutoring systems, instructional capabilities, and educational feedback quality.
Detection Tools
Evaluation of AI detection technologies and their effectiveness in educational contexts.
Ardito, C. G. (2025). Generative AI detection in higher education assessments. New Directions for Teaching and Learning, 2025, 11–28.
This paper explores the pitfalls of AI Detection software and offers remedies to the problem. It explains that AI Detection is both fragile and misaligned with how education is today. It states that detection with these softwares is unreliable, as students are able to avoid it by paraphrasing AI generated work, while non-fluent English writers are being flagged as their writing uses common English sentences. These inconsistencies make it almost impossible to prove whether AI is used or not, and often lead to false positives, and leave students anxious about their writing and left to deal with proving themselves to their professors that AI was not used. The author also disproves watermarking, a technique experts have suggested to help prove AI was used in writing by having GenAI tools leave a hidden signature in their writing. An example of a signature could be by first dividing a dictionary of words into 2 sets, and then alternating between the two when generating text. AI detection tools would then search for this signature and give a score based on that. However, watermarking isn't immune to paraphrasing, and additionally, there are already models released without watermarking, and implementing it would just lead to a new market of non-watermarked models, essentially making it hard to regulate.
This paper recommends shifting school work to in-person centered assessments, presentations, and having on-going check-ins for writing. This could limit the use of AI and lead to students writing on their own. Another approach to this issue, the paper suggests having transparent disclosures on the proper uses of AI. Having these set rules will help properly integrate AI into learning. As the paper mentions, AI mirrors the invention of the calculator, as like AI, it was at first looked down upon and considered cheating and banned in schools. However, a gradual shift in perception occurred where calculators were later accepted in schools and they began regulating its use and now see it as a tool to make mathematics more efficient. I think once we are past this period of AI slander and begin to see its potentials, creating clear and transparent guidelines for its use will bring down the uses for cheating and bring down our reliance on unreliable AI Detection tools.
This paper recommends shifting school work to in-person centered assessments, presentations, and having on-going check-ins for writing. This could limit the use of AI and lead to students writing on their own. Another approach to this issue, the paper suggests having transparent disclosures on the proper uses of AI. Having these set rules will help properly integrate AI into learning. As the paper mentions, AI mirrors the invention of the calculator, as like AI, it was at first looked down upon and considered cheating and banned in schools. However, a gradual shift in perception occurred where calculators were later accepted in schools and they began regulating its use and now see it as a tool to make mathematics more efficient. I think once we are past this period of AI slander and begin to see its potentials, creating clear and transparent guidelines for its use will bring down the uses for cheating and bring down our reliance on unreliable AI Detection tools.
Rohan Kapur 11/13/2025
Casu, M.; Guarnera, L.; Caponnetto, P.; Battiato, S. "GenAI Mirage: The Impostor Bias and the Deepfake Detection Challenge in the Era of Artificial Illusions." Forensic Sci. Int. Digit. Investig. 2024, 50, 301795.
This paper considers the impacts of "imposter bias," an emerging cognitive bias arising from the rapid proliferation of AI-generated media. Imposter bias describes the tendency to misidentify human-generated media as AI-generated. Implicitly, the essay asks us to recognize distrust of factual evidence as a problem as serious as the unquestioned acceptance of false evidence. The problem isn't (only) that we may be fooled by doctored photos or texts, but that we'll increasingly mistrust the genuine.
Alexander Landfair 12/2/2024
Chiusano, F. "Two minutes NLP: Perplexity explained with simple probabilities." Medium dot com. 2022.
A popular blogger explains "perplexity," one of the core concepts underlying most AI-detection tools. In short, perplexity describes a system's ability to predict the next word in a string of words. Underlying most AI-detection tools is the assumption that human-generated writing is somewhat less predictable than AI-generated writing. It's important to note that the concept of perplexity arose as a way to measure the quality of large-language models (because humans write with low-perplexity). (In my opinion, this necessarily disqualifies perplexity as a useful measure for detecting AI-writing).
Alexander Landfair 12/2/2024
The way this is written may be slightly misleading. Perplexity measures the model's literal "top predicted words" and checks how likely the actual next word is to be generated by the model. Lower perplexity means the text has lower "generation entropy". That is, the model has a high probability of generating that sequence. Humans inject noise that models don't naturally predict well, so they get higher perplexity. Modern models may deliberately raise entropy via RLHF. The current implementations of perplexity are miscalibrated, since they likely rely on GPT-2 perplexity. A differential perplexity test, such as a pure comparison of a non-chatbot transformer's perplexity and a modern chatbot's perplexity, is probably one of the strongest detection models possible.
Srinivas Harish 9/19/2025
Elek, A., Yildiz, H. S., Akca, B., Oren, N. C., & Gundogdu, B. (2025). Evaluating the efficacy of perplexity scores in distinguishing AI-generated and human-written abstracts. Academic Radiology, 32(4), 1785–1790. https://doi.org/10.1016/j.acra.2025.01.017
This study investigates the effectiveness of using perplexity scores (metric for text predictability) to differentiate between human-written and AI-generated (ChatGPT-3.5) academic abstracts.
The researchers also measured the performance of three AI detection tools (Google's Bard, ZeroGPT, and GPTZero) on this task. The research found that human-written abstracts had significantly higher (less predictable) perplexity scores than their AI-generated counterparts (AUC=0.7794).
The primary utility of this research is its direct comparison of commercial detection tools, which revealed a massive disparity in performance. GPTZero (an AI detector that also uses perplexity) was found to be highly accurate (95%), while ZeroGPT was mediocre (69%) and Google's Bard performed worse than chance (36% accuracy). This finding is important for educators, administrators, and journal editors, as it demonstrates that while AI detection is possible, the reliability of a given tool cannot be assumed (this also supports the findings of some of the other articles under the Evaluating Tools section).
Furthermore, this study provides a good empirical test case for the other articles. It applies the core concept of perplexity that Chiusano explains, but its findings are immediately complicated by other research (like Lindberg & Nobel's and Liang's) showing this exact metric is unreliable, biased against non-native speakers, and insufficient without other features like "burstiness." This research also directly benchmarks the same tools (GPTZero and ZeroGPT) that Pratama, Weichert, and Elkhatat evaluate. While Elek finds GPTZero is highly accurate (95%), these other papers provide the crucial counterpoint that it is also biased, easily fooled by prompt engineering, and inconsistent against newer models. This highlights the real-world consequences of imperfect detection, such as the "imposter bias" and the genre-specific false positives.
The researchers also measured the performance of three AI detection tools (Google's Bard, ZeroGPT, and GPTZero) on this task. The research found that human-written abstracts had significantly higher (less predictable) perplexity scores than their AI-generated counterparts (AUC=0.7794).
The primary utility of this research is its direct comparison of commercial detection tools, which revealed a massive disparity in performance. GPTZero (an AI detector that also uses perplexity) was found to be highly accurate (95%), while ZeroGPT was mediocre (69%) and Google's Bard performed worse than chance (36% accuracy). This finding is important for educators, administrators, and journal editors, as it demonstrates that while AI detection is possible, the reliability of a given tool cannot be assumed (this also supports the findings of some of the other articles under the Evaluating Tools section).
Furthermore, this study provides a good empirical test case for the other articles. It applies the core concept of perplexity that Chiusano explains, but its findings are immediately complicated by other research (like Lindberg & Nobel's and Liang's) showing this exact metric is unreliable, biased against non-native speakers, and insufficient without other features like "burstiness." This research also directly benchmarks the same tools (GPTZero and ZeroGPT) that Pratama, Weichert, and Elkhatat evaluate. While Elek finds GPTZero is highly accurate (95%), these other papers provide the crucial counterpoint that it is also biased, easily fooled by prompt engineering, and inconsistent against newer models. This highlights the real-world consequences of imperfect detection, such as the "imposter bias" and the genre-specific false positives.
Josephine Li 11/20/2025
Liang, W.; Yuksekgonul, M.; Mao, Y.; Wu, E.; Zou, J. "GPT detectors are biased against non-native English writers." arXiv:2304.02819. 2023.
This study evaluates the fairness of seven widely used GPT detectors by testing their performance on writing samples from both native and non-native English writers. The researchers' core finding is that these detectors are significantly biased against non-native English speakers. While the detectors are nearly perfect at identifying essays written by native speakers, they consistently misclassify the writing of non-native speakers as AI-generated. For instance, over half of the TOEFL essays written by non-native speakers were incorrectly flagged as AI-generated. The study suggests this bias occurs because detectors often rely on a metric called "perplexity," which measures linguistic complexity. Non-native writers tend to use less varied vocabulary and sentence structures, resulting in lower perplexity text that the detectors mistake for AI-generated content. The authors caution against using these tools in educational settings, as they risk unfairly penalizing non-native English speakers for academic misconduct.
Josephine Li 10/13/2025
Dalalh, D. "The false positives and false negatives of generative AI detection tools in education and academic research: The case of ChatGPT." The International Journal of Management Education. 2023, Volume 21 Issue 2.
The study investigates false positives (human-written text incorrectly flagged as AI-generated) and false negatives (AI-generated text missed by detectors) in AI Detection, with a particular focus on its impact on academic research and education, where integrity is crucial. Running tests on the abstract and literature parts of research papers before 2022 (before the release of ChatGPT), researchers found that the literature section of papers was more likely to be flagged as AI-generated compared to abstracts. This shows that the "formulaic" style of literature reviews resembles AI writing more closely. Beyond the results of the experiment, this paper argues that the current definitions of plagiarism and authorship are outdated in the AI era, calling for combining plagiarism detection with AI-detection tools.
Nasrin Alanna Aidi 9/25/2025
Elkhatat, A.M.; Elsaid, K.; Almeer, S. "Evaluating the efficacy of AI content detection tools in differentiating between human and AI‑generated text." International Journal for Educational Integrity. 2023, pp. 1-17.
This study evaluates how the top 5 used AI content detection tools (OpenAI, Writer, Copyleaks, GPTZero, and CrossPlag) distinguish between human-authored and AI-generated contents. The tools were evaluated on both paragraphs produced by ChatGPT models 3.5 and 4, as well as on human-written controls. Results shows that all the detection tools were better at predicting GPT 3.5 than GPT 4.0 and with inconsistent performance on human-written samples, which includes false positives too. This concludes ongoing challenges in AI Detection reliability with the advancement of text generation technology. The authors recommend combining AI-detection tools with Manual Review to keep the Academic Integrity intact.
Unmesh Achar 10/15/2025
This empirical study investigates how well several commercially available AI-authorship-detection tools (for text) perform at distinguishing human- vs AI-written content. Specifically: 15 paragraphs each from ChatGPT 3.5 and 4 on a technical topic plus five human-written control paragraphs; tools by e.g., OpenAI, Writer, Copyleaks, GPTZero, CrossPlag were used. Results: the detectors were more accurate for GPT-3.5 than GPT-4, but when applied to the human-written control responses they yielded inconsistent results, including false positives. They conclude that detection tools still have substantial limitations as AI-generated content becomes more sophisticated.
Yanze Wu 10/15/2025
Lan Z. et al. "ALBERT: A Lite BERT for Self-supervised Learning of Language Representations" ArXiv. 2019.
Transformer models like BERT and RoBERTa, used for AI text detection, achieve SOTA results by stacking dozens of "attention" layers together. However, prior work (ALBERT) has shown significant redundancy across these layers. Many parameters do not contribute meaningfully to representation or expressivity of the model. Furthermore, information flows strictly forward with no mechanism for later layers to talk to earlier layers or revisit information from earlier representations. The team should aim to fix this "weakness" with an improved architecture.
Srinivas Harish 9/19/2025
Pratama, A. R. "The accuracy-bias trade-offs in AI text detection tools and their impact on fairness in scholarly publication." PeerJ Comput. Sci. 2025, e2953.
This study evaluates three AI text detection tools (GPTZero, ZeroGPT, and DetectGPT) and finds a significant trade-off between accuracy and bias. The tools flag text from non-native English speakers and certain academic disciplines as AI-generated, raising concerns about fairness in scholarly publication. All tools struggled with AI-assisted text and showed systemic bias. Abstracts from social sciences and those written by non-native English speakers had a significantly higher false positive rate, meaning that they were more likely to be incorrectly identified as AI-generated.
Josephine Li 10/09/2025
Lu, N., Liu, S., He, R., Ong, Y.-S., Wang, Q., & Tang, K. (2023). Large language models can be guided to evade AI-generated text detection [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2305.10847
The paper introduces a concept named SICO (Substitution-based In-Context example Optimization). It is a novel and cost-efficient method for constructing prompts that guide large language models (LLMs) to evade AI-generated text detectors. The method automatically optimizes a few human-written examples to be less "AI-like", which are then used as in-context demonstrations for the LLM.
The research's strength is in demonstrating a significant vulnerability in existing detection systems. The study shows that prompts generated by SICO enabled GPT-3.5 to successfully bypass six different detectors, significantly outperforming baseline paraphrasing methods. A comprehensive human evaluation further confirmed that the SICO-generated text maintained human-level readability and task completion rates.
This work is valuable because it highlights that evasion can be achieved through simple prompt-guiding rather than external tools, and it provides a new, effective benchmark (SICO) for evaluating the robustness of future detectors.
It also connects well with several past papers I've found that looks into prompt engineering and using that to bypass AI-detection tools (through targeting perplexity and burstiness).
The research's strength is in demonstrating a significant vulnerability in existing detection systems. The study shows that prompts generated by SICO enabled GPT-3.5 to successfully bypass six different detectors, significantly outperforming baseline paraphrasing methods. A comprehensive human evaluation further confirmed that the SICO-generated text maintained human-level readability and task completion rates.
This work is valuable because it highlights that evasion can be achieved through simple prompt-guiding rather than external tools, and it provides a new, effective benchmark (SICO) for evaluating the robustness of future detectors.
It also connects well with several past papers I've found that looks into prompt engineering and using that to bypass AI-detection tools (through targeting perplexity and burstiness).
Josephine Li 11/13/2025
Mitchell, E.; Lee, Y.; Khazatsky, A.; Manning, C. D.; Finn, C. (2023). "DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature."
This paper proposes DetectGPT, a detector that flags AI-generated text by checking whether a passage sits in a negative curvature region of a language model's log-probability function. Unlike classifier-based detectors, DetectGPT needs no extra training data—it uses only the target model's log-probs plus small perturbations from a secondary model. In evaluations (e.g., fake-news detection with GPT-NeoX-20B), the method improves zero-shot AUROC over prior baselines and provides a model-agnostic way to test text authenticity.
Yanze Wu 10/15/2025
Lindberg, K., & Nobel, I. (2025). AI or Human? Detecting Machine-Generated Text Through Perplexity (Dissertation).
This paper investigates the effectiveness of using perplexity (a metric for how predictable text is to an AI model) as a method for detecting whether a piece of writing was generated by AI or a human. The core idea is that AI-generated text should be highly predictable (low perplexity) to its own model, while human writing is more varied and less predictable (high perplexity). The findings show this evaluation method is only reliable in highly controlled scenarios. For example, when the same AI model is used to both generate and evaluate the text. The paper notes that the separation between AI and human writing "becomes less clear" as soon as more varied sources are introduced, revealing the method's limitations.
Josephine Li 10/22/2025
Weichert, J.; Dimobi, C. "Detection Undermining via Prompt Engineering for Deepfake Text." arXiv. 2024.
This article rigorously evaluates three AI text detectors (Watermarking, ZeroGPT, and GPTZero) on their ability to distinguish human-written from AI-generated essays. Results show that paraphrasing attacks, including prompts designed to mimic "college student" style or intentionally increase perplexity, can dramatically raise the false negative rates of all detectors, making AI-generated writing much more likely to evade detection. For instance, ZeroGPT's false negative rate rose to nearly 90% with targeted paraphrasing. The paper warns educators against relying solely on AI detectors for plagiarism judgments, highlighting the ease with which detection can be bypassed and the risk of harm from false accusations.
Josephine Li 10/22/2025
Equity / Social Justice
Analyzing how AI tools impact educational equity and social justice issues.
Liang, W.; Yuksekgonul, M.; Mao, Y.; Wu, E.; Zou, J. "GPT detectors are biased against non-native English writers." arXiv:2304.02819. 2023.
This study evaluates the fairness of seven widely used GPT detectors by testing their performance on writing samples from both native and non-native English writers. The researchers' core finding is that these detectors are significantly biased against non-native English speakers. While the detectors are nearly perfect at identifying essays written by native speakers, they consistently misclassify the writing of non-native speakers as AI-generated. For instance, over half of the TOEFL essays written by non-native speakers were incorrectly flagged as AI-generated. The study suggests this bias occurs because detectors often rely on a metric called "perplexity," which measures linguistic complexity. Non-native writers tend to use less varied vocabulary and sentence structures, resulting in lower perplexity text that the detectors mistake for AI-generated content. The authors caution against using these tools in educational settings, as they risk unfairly penalizing non-native English speakers for academic misconduct.
Josephine Li 10/13/2025
Reasoning
Testing AI's logical reasoning, problem-solving, and critical thinking capabilities.
Hussain, S. M.; Malik, A.; Ahmad, N.; Ahmed, S. "ChatGPT vs. Students: A Comparative Analysis of Performance in Corporate Finance Exams." Journal of Educational Technology Systems. 2025.
This article focuses on a direct performance comparison between ChatGPT-4 and university students on corporate finance exams. One of the most striking findings was the stark contrast in performance based on question type. ChatGPT significantly outperformed students in descriptive questions (87.1% vs. 74.6%), demonstrating a clear strength in processing and summarizing conceptual, text-based information. However, this advantage completely reversed in numerical questions, where students vastly outperformed ChatGPT (81.6% vs. 31.9%). The authors conclude that ChatGPT relies on predicting likely answers from language patterns rather than performing actual calculations, leading to a fundamental failure in quantitative reasoning.
Meiling Zhang 10/9/2025
Ethics / Safety
Examining ethical considerations and safety concerns in AI tool deployment.
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2024). Generative AI Can Harm Learning. The Wharton School Research Paper. https://doi.org/10.2139/ssrn.4895486
A large field experiment in U.S. high-school math classes deployed two GPT-4 tutors for ~1,000 students: a general "GPT Base" chat interface and a guardrailed "GPT Tutor" designed with prompts to encourage learning behaviors. Results show a tradeoff: unguided assistance can boost practice performance yet undermine actual learning/retention, while structured guardrails can mitigate harms and better support learning objectives—evidence that design choices (not just model quality) matter in education.
Yanze Wu 10/15/2025
Yang, Z., Wu, J. G., & Xie, H. (2024). Taming Frankenstein's monster: Ethical considerations relating to generative artificial intelligence in education. Asia Pacific Journal of Education, 45(4), 1330–1343.
This paper provides an overview of the ethical concerns that arise with the increased use of Generative AI (GAI) in higher education, and outlines a specific set of guidelines for mitigating the unintended consequences of AI in learning environments. The introduction of the paper grounds the analysis in the history of Science Fiction, and utilizes Frankenstein's monster as an allegory for the consequences of unregulated innovation. Mary Shelly's monster is quoted: "Beware; for I am fearless, and therefore powerful" - as a means of further driving the notion that GAI is currently unbounded and unregulated. From this initial literary reference the paper goes on to first define consciousness and thought - and relate this back to our understandings of GAI. The paper notes that it is imperative to distinguish between human thought and GAI and that as tech continues to evolve the line can become increasingly blurred. Following this definition the varying ethical concerns that arise with GAI use, within and outside of educational environments, are discussed. The central concerns outlined relate to sourcing and informational bias that can arise with the use of GAI. The paper uses the example of corporations intentionally creating bias in AI to sway public opinion as a means of exploring possible malicious uses of GAI, while other primary concerns focus on the unintended consequences of GAI use. The paper then poses general questions about regulation in the face of ethical concerns - who's responsible for regulating AI? The paper responds to these questions by outlining a set of guidelines to be implemented by both students and educators as a means of combating the consequences of AI use:
Submodel for educators:
- Equity concerns
- Digital literacy
- Enhance vigilance of bias
- Pedagogy and assessment concerns
- Encourage higher-order thinking
- Incentivize the learning process
- Strike a balance
- Optimizing potential of technology
- Establish a general policy
- Set specific boundaries
- Update policy regularly
Submodel for students:
- Preparation for using GAI
- Improve awareness of limitations
- Carry-out critical evaluation
- Applying GAI
- Seek out additional resources
- Incorporate creativity and originality
- Disclosure
- Acknowledge and site sources properly
I found that the specific breakdown of responsibility for both educators and students to be well organized and thought out - I felt that the formatting gave a clear understanding of intention and a real attempt to find solutions and answers to questions which remain fairly open ended. I don't know if the literary introduction is as indicative of what the paper really seeks to argue - however I appreciate the reference and the greater thought framework it represents.
Submodel for educators:
- Equity concerns
- Digital literacy
- Enhance vigilance of bias
- Pedagogy and assessment concerns
- Encourage higher-order thinking
- Incentivize the learning process
- Strike a balance
- Optimizing potential of technology
- Establish a general policy
- Set specific boundaries
- Update policy regularly
Submodel for students:
- Preparation for using GAI
- Improve awareness of limitations
- Carry-out critical evaluation
- Applying GAI
- Seek out additional resources
- Incorporate creativity and originality
- Disclosure
- Acknowledge and site sources properly
I found that the specific breakdown of responsibility for both educators and students to be well organized and thought out - I felt that the formatting gave a clear understanding of intention and a real attempt to find solutions and answers to questions which remain fairly open ended. I don't know if the literary introduction is as indicative of what the paper really seeks to argue - however I appreciate the reference and the greater thought framework it represents.
Elan Hebert 11/13/2025
Usability
Assessing user experience, interface design, and ease of use in AI tools.
Sundar, S. S. (2020). Rise of Machine Agency: A Framework for Studying the Psychology of Human-AI Interaction (HAII). Journal of Computer-Mediated Communication, 25(1), 74-88. https://doi.org/10.1093/jcmc/zmz026
This article focuses on the shift in communication scholarship from "communication via" technology to "communicating with" technology, marking the transition to Human-AI Interaction (HAII). Media technologies have evolved from passive channels to active, intelligent agents with autonomous decision-making capabilities. The paper examines the fundamental tension between machine agency (AI's ability to personalize, recommend, and curate content independently) and human agency (users' desire for control and meaningful participation). It proposes adapting the Theory of Interactive Media Effects (TIME) to study HAII through two mechanisms: the "cue route" (how users perceive AI as a symbol—machine heuristics, transparency, interface cues) and the "action route" (how users engage with AI through collaboration, customization, and control exchanges).
The article critiques how increasing machine agency threatens human autonomy through privacy erosion, algorithmic opacity, automation bias (overtrusting machines), and algorithm aversion (rejecting algorithms even when optimal). Users often feel helpless against incomprehensible algorithms, experiencing "algorithmic anxiety" (Airbnb hosts) or unawareness of curation (Facebook feeds). To address this, the author advocates for human-AI collaboration: seeking user assent before automated decisions, increasing algorithmic transparency, enabling users to train and direct algorithms ("ground-truthing"), and pursuing mutual augmentation where humans and AI extend each other's capabilities. Trust built through deeper engagement (action route) proves more robust than perception alone (cue route), ultimately calling for co-creation that respects human agency while leveraging machine capabilities.
The article critiques how increasing machine agency threatens human autonomy through privacy erosion, algorithmic opacity, automation bias (overtrusting machines), and algorithm aversion (rejecting algorithms even when optimal). Users often feel helpless against incomprehensible algorithms, experiencing "algorithmic anxiety" (Airbnb hosts) or unawareness of curation (Facebook feeds). To address this, the author advocates for human-AI collaboration: seeking user assent before automated decisions, increasing algorithmic transparency, enabling users to train and direct algorithms ("ground-truthing"), and pursuing mutual augmentation where humans and AI extend each other's capabilities. Trust built through deeper engagement (action route) proves more robust than perception alone (cue route), ultimately calling for co-creation that respects human agency while leveraging machine capabilities.
Joseph Jiminian 10/16/2025
Testini, I., Pacchiardi, L., & Hernandez-Orallo, J. (2025). Measuring Data Science Automation: A Survey of Evaluation Tools for AI Assistants and Agents. Transactions on Machine Learning Research. https://openreview.net/forum?id=MB0TCLfLn1
This survey paper examines the landscape of evaluation methods and tools for large-language-model-based assistants and agents (especially in data-science contexts). Rather than detecting generation per se, it focuses on how we evaluate, benchmark, and characterise AI assistant performance (e.g., suggestion of methods, code generation, interpretation), including human-AI collaboration, levels of autonomy, and the tasks involved. It finds (1) a heavy focus on a small subset of goal-oriented activities, (2) limited study of intermediate human-AI collaboration modes, and (3) limited attention on automation via task transformation rather than substitution.
Yanze Wu 10/15/2025
Xu, W., Dainoff, M. J., Ge, L., & Gao, Z. (2021). From Human-Computer Interaction to Human-AI Interaction: New Challenges and Opportunities for Enabling Human-Centered AI. ArXiv. https://arxiv.org/abs/2105.05424v1
This paper summarizes what HCI professionals can contribute to human-AI interaction (HAII) including research, design, practice and policy. The authors argue that HCI professionals should act as advocates for human-centered values in the development and deployment of AI, particularly concerning data privacy and algorithmic equity.
Alexander Landfair 12/2/2024
This article focuses on the role Human-Computer Interaction (HCI) professionals must play in developing human centered AI systems. The paper examines major challenges that distinguish AI system development from traditional computing systems: managing unique machine behaviors in AI, transitioning from simple interaction to human-AI collaboration and teaming, bridging human intelligence augmentation with AI, developing human machine hybrid intelligence, moving from UI usability to AI explainability and interpretability, shifting from automation to human-controlled autonomy, integrating ethical design principles, and creating intelligent interaction paradigms. Throughout these topics, the authors argue that AI systems possess fundamentally different characteristics, such as non deterministic behavior, learning capabilities, and autonomous decision making that require new HCI approaches beyond traditional design thinking.
The article critiques the AI and computer science communities for developing systems in isolation without adequate input from HCI professionals and behavioral scientists. It criticizes how machine behavior is primarily studied by those without formal behavioral science training, how "explainable AI" is often built for AI professionals rather than end users, how ML engineers are positioned as the "human at the center" instead of actual users, and how ethical guidelines lack practical implementation methods. The authors advocate for a Human-Centered AI (HCAI) approach where humans remain the ultimate decision makers, calling for HCI professionals to: leverage interdisciplinary expertise to translate psychological theories into computational models, develop human-in-the-loop systems that combine human and machine strengths, create meaningful human control mechanisms for autonomous systems, ensure AI explainability leads to true user comprehension, and apply iterative design methods with end users throughout AI development rather than as an afterthought. The paper emphasizes that effective AI systems require deep collaboration between HCI and AI professionals from the earliest stages of development.
The article critiques the AI and computer science communities for developing systems in isolation without adequate input from HCI professionals and behavioral scientists. It criticizes how machine behavior is primarily studied by those without formal behavioral science training, how "explainable AI" is often built for AI professionals rather than end users, how ML engineers are positioned as the "human at the center" instead of actual users, and how ethical guidelines lack practical implementation methods. The authors advocate for a Human-Centered AI (HCAI) approach where humans remain the ultimate decision makers, calling for HCI professionals to: leverage interdisciplinary expertise to translate psychological theories into computational models, develop human-in-the-loop systems that combine human and machine strengths, create meaningful human control mechanisms for autonomous systems, ensure AI explainability leads to true user comprehension, and apply iterative design methods with end users throughout AI development rather than as an afterthought. The paper emphasizes that effective AI systems require deep collaboration between HCI and AI professionals from the earliest stages of development.
Joseph Jiminian 10/16/2025
Coding and Math
Evaluating AI tools' capabilities in programming, code generation, and debugging tasks.
Content in Development
VIP Evaluating Tools subteam is currently curating resources on AI coding capabilities. We welcome your contributions and findings.
Creativity
Assessing AI's ability to generate creative content and support creative thinking processes.
Content in Development
VIP Evaluating Tools subteam is currently curating resources on AI creativity and creative content generation. We welcome your contributions and findings.
Error / Hallucination
Understanding AI errors, hallucinations, and accuracy limitations.
Özer, M. (2024). Is Artificial Intelligence Hallucinating? National Library of Medicine. Retrieved September 11, 2025.
This article explores the causes and types of hallucinations that AI experiences. The article can be described as a literature review, although not super in depth. But I still think reading it is still extremely helpful to anyone who wants to understand the basics of AI hallucinations. Hallucinations are classified into intrinsic and extrinsic. Özer writes, "intrinsic hallucinations result in outputs that contradict the source content and the conversation history, while extrinsic hallucinations correspond to outputs whose accuracy cannot be verified based on the source content or conversation history." According to the author, conflicting information within the training data that is used on the AI is what can cause hallucinations. I feel like this point is quite obvious, since AI can be simplified to a mapping of relative probabilities of words. The author also suggests using a word other than "hallucination" to describe this phenomenon. They suggest "confabulation," since AI models don't have consciousness. In the conclusion, the author states that hallucinations can have varying degrees of impact on the user, based on what the user is using the AI for. The author brought up the example of using AI in the healthcare sector, but I found that confusing because, to my knowledge, generative AI isn't currently being in the diagnosis or treatment process.
Yufei Huang 11/9/2025
Translation
Evaluating AI translation capabilities and multilingual support quality.
Content in Development
VIP Evaluating Tools subteam is currently curating resources on AI translation capabilities. We welcome your contributions and findings.
Tutoring / Instruction
Evaluating AI tutoring systems, instructional capabilities, and educational feedback quality.
Owan, V. J., Abang, K. B., Idika, D. O., Etta, E. O., Bassey, B. A. "Exploring the Potential of Artificial Intelligence Tools in Educational Measurement and Assessment." EURASIA Journal of Mathematics, Science and Technology Education. 2023, em2307.
This study evaluates how artificial intelligence tools are being integrated into educational measurement and assessment which revolutionizes the traditional practices. AI supports personalized learning, intelligent tutoring and automated grading which improves the speed and feedback quality. The authors noted down that the predictive analytics and natural language processing tools further gives individualized support and expand the assessment possibilities. Importantly, while AI streamlines many tasks, they emphasize the need for ethical safeguards, transparency and ongoing teacher involvement to counter the risks around privacy, fairness and overdependency on technology. The authors recommend Teacher training and a blend of human and AI analysis are recommended for best practices.
Unmesh Achar 10/15/2025
Steiss, J.; Tate, T.; Graham, S.; Cruz, J.; Hebert, M.; Wang, J.; Moon Y.; Tseng, W.; Warschauer, M.; Olson, C. B. "Comparing the quality of human and ChatGPT feedback of students' writing." Learning and Instruction. 2024. 101894.
This study investigates the question of whether AI can serve as a viable tool for providing formative feedback on students' essays by comparing ChatGPT-3.5 feedback with teacher feedback. They evaluated the feedback based on five key criteria: being criteria-based, offering clear directions, accuracy, prioritization of essential features, and using a supportive tone. The key findings are that teacher feedback outperformed ChatGPT in four of the five categories. The only area where AI was stronger was "criteria-based," meaning that it was good at explicitly referencing writing rubrics and standards. Importantly, the study also found that the quality of ChatGPT's feedback varied depending on the quality of the essay; for example, ChatGPT struggled with accuracy on high-quality essays and a supportive tone on low-quality ones.
Meiling Zhang 9/24/2025
The study looked at how ChatGPT3.5's feedback on student essays compared to teacher feedback. They judged the feedback in five areas: whether it followed the criteria, gave clear directions, was accurate, prioritized what mattered most, and had a supportive tone. Teachers were stronger in almost every category, except for being "criteria-based," where ChatGPT did better because it directly tied back and connected to the feedback to rubrics and standards. The researchers also found that ChatGPT's performance depended on the quality of the essay because it sometimes gave less accurate feedback on strong essays and wasn't very supportive when commenting on weaker ones. In the reflection, Zhang mentions that AI is definitely useful and can save a lot of time, but it shouldn't replace the writing process we know (writing drafts, getting feedback, etc.). According to her, writing should be messy and all over the place, which is where its true beauty lies.
I think Zhang makes valid points about not relying on AI to completely change the way we write. The convenience is definitely tempting especially when you're stuck and don't know where to start. However, I think that true skill when it comes to writing comes from struggling to get something down, constant revisions, and hearing what other real humans have to say about it. I think this is where our creativity can really come out. I think that in the end, while ChatGPT and AI are useful tools that can definitely be integrated in some ways and forms, it is best for us humans to continue to learn how to develop our writing ourselves in the traditional way that we know of.
I think Zhang makes valid points about not relying on AI to completely change the way we write. The convenience is definitely tempting especially when you're stuck and don't know where to start. However, I think that true skill when it comes to writing comes from struggling to get something down, constant revisions, and hearing what other real humans have to say about it. I think this is where our creativity can really come out. I think that in the end, while ChatGPT and AI are useful tools that can definitely be integrated in some ways and forms, it is best for us humans to continue to learn how to develop our writing ourselves in the traditional way that we know of.
Ethan Lu 10/01/2025
Sarkar, T. "The evolution of mastery learning: Challenges, technologies, and the AI journey leading to the intelligent tutoring system DARTS." Issues in Information Systems. 2025, 26(4), 45–55.
These articles explore the intersection of technology and education in Mastery Learning. Working off Bloom's work and his findings that one-on-one tutoring can produce learning outcomes two standard deviations above conventional instruction, the articles explore how AI-driven Intelligent Tutoring Systems (ITS) can serve to make this a more accessible possibility for all students. The first paper explores the motivations for the author's system: Dynamic Academic Response and Tutoring System (DARTS), which he initially proposed in his PHD thesis in 2022. DARTS expands on the framework of other ITS systems by having the system interface with students through SMS on their phones. It includes the inner workings of other ITS systems, which are the domain model, which holds subject knowledge and expert strategies; the student model, which tracks learner progress and misconceptions to build a personalized profile; and the tutoring (pedagogical) model, which selects instructional strategies and feedback. While other ITS models require internet or a specific app, DARTS allows any student with a mobile device to interact with the system and integrate it into the classroom.
Jesse Noppe-Brandon 10/29/2025
Sarkar, T. "Designing DARTS: A unified AI system for tutoring, classroom interaction, assessment, and students' success." Issues in Information Systems. 2025, 26(4), 92–101.
The second article delves more into the design of DARTS. DARTS aims to serve two primary functions: supporting real-time classroom interaction and enabling personalized out-of-class tutoring. Real-time classroom interaction comes in many forms. One is taking attendance, where students are given a code in class and are prompted to send that code within a specified time limit; the system will log and store excused absences for that student, and excused or not DARTS automatically retrieves a tutoring plan generated by its AI engine and sends the student a personalized learning message. This allows the student to learn alongside what they are missing. Another feature is class quizzes, where students respond using the SMS system. The AI engine analyzes correctness and frequency of participation. Based on this, personalized feedback or review content is automatically pushed to the student's phone. Finally, there is brainstorming, where a prompt is put on the board by the teacher, and students send in responses.
These responses, or the lack of a response, are saved and analyzed in order to recommend follow-up tutoring or enrichment tasks based on inferred strengths and weaknesses. All of these factors go into creating a personalized system for each individual student, which the student can go home and interact with like a one-on-one tutor. Feedback and interaction are based on student history and knowledge gaps, and students can engage in two-way conversations through SMS with their tutor. Teachers also have dashboards, which allow them to guide and initiate tutoring interventions.
DARTS uses secure cloud-based databases with encrypted data transmission and storage, and all data is handled in compliance with FERPA. Using cloud-native design, the system is scalable to classrooms of any size, and through SMS, the system can be used even without an internet connection.
These papers are a three-part series, and the third part will look at actual classroom integration. I think this is a fascinating set of articles, and I am very interested in technology like this. I think there are many things that need to be extensively researched and tested before something like this could be rolled out, but I feel very optimistic about the possibilities that this gives to allow for Mastery Learning in a school system that doesn't accommodate those who can't master the content at the set pace. The article, like my previous two articles, stresses the importance of teacher training for integrating technology like this and introducing students to the technology at the beginning of the school year.
These responses, or the lack of a response, are saved and analyzed in order to recommend follow-up tutoring or enrichment tasks based on inferred strengths and weaknesses. All of these factors go into creating a personalized system for each individual student, which the student can go home and interact with like a one-on-one tutor. Feedback and interaction are based on student history and knowledge gaps, and students can engage in two-way conversations through SMS with their tutor. Teachers also have dashboards, which allow them to guide and initiate tutoring interventions.
DARTS uses secure cloud-based databases with encrypted data transmission and storage, and all data is handled in compliance with FERPA. Using cloud-native design, the system is scalable to classrooms of any size, and through SMS, the system can be used even without an internet connection.
These papers are a three-part series, and the third part will look at actual classroom integration. I think this is a fascinating set of articles, and I am very interested in technology like this. I think there are many things that need to be extensively researched and tested before something like this could be rolled out, but I feel very optimistic about the possibilities that this gives to allow for Mastery Learning in a school system that doesn't accommodate those who can't master the content at the set pace. The article, like my previous two articles, stresses the importance of teacher training for integrating technology like this and introducing students to the technology at the beginning of the school year.
Jesse Noppe-Brandon 10/29/2025
Assessing Impacts
These resources help us to answer questions like: What are the effects of AI tools on student learning outcomes? How do these tools impact academic integrity, assessment validity, and student wellbeing? What are the equity implications of AI adoption in education?
Learning Outcomes 0
Measuring how AI tools affect student learning, comprehension, and skill development.
Academic Integrity 0
Assessment of AI's effect on academic honesty and institutional integrity measures.
Assessment Validity 0
How AI tools impact the reliability and validity of educational assessments and grading practices.
Student Experience + Wellbeing 0
Impact of AI on student stress, motivation, confidence, and overall educational experience.
Equity + Access 0
How AI tools affect educational equity and access across different student populations.
Learning Outcomes
Measuring how AI tools affect student learning, comprehension, and skill development.
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2024). Generative AI Can Harm Learning. The Wharton School Research Paper. https://doi.org/10.2139/ssrn.4895486
A large field experiment in U.S. high-school math classes deployed two GPT-4 tutors for ~1,000 students: a general "GPT Base" chat interface and a guardrailed "GPT Tutor" designed with prompts to encourage learning behaviors. Results show a tradeoff: unguided assistance can boost practice performance yet undermine actual learning/retention, while structured guardrails can mitigate harms and better support learning objectives—evidence that design choices (not just model quality) matter in education.
Yanze Wu 10/15/2025
Ahn, J., Verma, R., Lou, R., Liu, D., Zhang, R., & Yin, W. (2024). Large Language Models for Mathematical Reasoning: Progresses and Challenges. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop (pp. 1-8). Association for Computational Linguistics.
This article not only examines the progress of LLM's in mathematical reasoning but also its implications in regards to education. Specifically looking at Math Word Problems (MWP). These problems may include tables and/or geometric figures. The article examines how tokenization, pre-training, prompting techniques, interpolation and extrapolation, scaling laws, Chain of Thought (COT), and In-Context Learning (ICL) impact the reasoning of LLM's. These factors play a direct role in how well a model does with MWP's. The article then evaluates the advantages and disadvantages of LLM's in math education. One disadvantage is that LLM's may not suit a student's learning style, especially if they learn from hands-on activities or visual aids. Though, LLM's also introduce a 'conversational style' in problem-solving which is important in mathematics and can offer further insight into subjects like Calculus. This research is unique in its assessment of LLM's in math education while also offering insight into LLM performance with MWP's. Something I'm curious about after reading this article is how well various models perform on more theoretical math courses like Discrete Math or Analysis. The MWP's assessed have concrete solutions involving equations or numbers, but in more advanced math courses answers can just be words and symbols.
Karen Maza Delgado 11/11/2025
Bettayeb, A. "Exploring the impact of ChatGPT: conversational AI in education." Frontiers in Education. 2024.
This journal provides a literary review of a research study conducted between 2022-2023 on the specific impacts of ChatGPT on educational environments. The findings showed that the adaptable nature of ChatGPT had a positive impact on student learning, allowing for tailored learning experiences and increased accessibility for students due to the conversational mode of information retrieval. The study overall showed positive outcomes on student learning, while citing a number of persisting concerns largely in regards to the ethics of models like ChatGPT. The main ethical and practical concerns the study raises are those of bias and authorship and misinformation.
Elan Hebert 10/9/2025
Boubker, O.; Ben-Saghroune, H.; Bourassi, J.; Abdessadek, M.; Sabbahi, R.; Abdellah, M. "Examining the Impact of OpenAI's ChatGPT on PhD Student Achievement." International Journal of Information and Education Technology, 2024, 14, 443–451.
This article presents a study that explores the influence of ChatGPT on PhD students' performance and examines factors that motivate their usage as well as their level of satisfaction with the tool. They found that there was a direct positive effect of ChatGPT's perceived usefulness on ChatGPT use and PhD students' satisfaction. They found a significant positive relationship between students who use ChatGPT and satisfaction with the tool. Finally, they found that students who use ChatGPT and are satisfied with ChatGPT reported net benefits on their studies, like increasing productivity, enhancing performance, simplifying research tasks, and improving research quality.
Jesse Noppe-Brandon 10/15/2025
Habib, S.; Vogel, T.; Anli, X.; Thorne, E. "How does generative artificial intelligence impact student creativity?" Journal of Creativity, 2024.
This paper presents a mixed-methods study that explores how ChatGPT-3 influences students' creative thinking abilities and confidence in a college-level creativity course. The authors found that using ChatGPT-3 significantly improved scores in all four creativity dimensions (fluency, flexibility, elaboration, originality). However, qualitative findings revealed that while AI boosted idea generation speed and provided inspiration, it sometimes weakened originality and creative confidence. The study concluded that AI can enhance divergent thinking but may also lead to overreliance and cognitive fixation, where students depend too much on AI-generated ideas.
Dan Ahimbisibwe 10/16/2025
Liu, S., Guo, X., Hu, X., & Zhao, X. (2024). Advancing Generative Intelligent Tutoring Systems with GPT-4: Design, Evaluation, and a Modular Framework for Future Learning Platforms. Electronics, 13(24), 4876.
This article presents a framework for designing an AI-assisted ITS using ChatGPT. It begins with a literature review of classical ITS, ChatGPT, and Intelligent ITS. It then breaks down the framework of how to design the system using ChatGPT. Four components are created: a content retriever which pulls information based on the topic the user is interested in; a data analyzer which monitors the user and generates a dynamic user profile; an instructional advisor, which adapts the difficulty and type of content based on the user's performance; a feedback assessor which generates reports based on the data analyzer and the instructional advisor. In order for the system to work, all data is generated as JSON files, where the prompts given to the models specify exactly what needs to be in the files, and they are iterated upon until they pass validation. This allows the different models to communicate with each other through lightweight JSON information.
The authors create an example ITS called the Socratic Playground for Learning (SPL), which implements this system. Through a user interface, students are able to interact with these models by defining, through a menu or a text box, what they are learning about or practicing. The student then engages in back-and-forth Socratic-style conversation. The questions presented to them dynamically scale in difficulty and complexity depending on their answers and responses.
To test this system, the authors involved 30 first-year students from the Faculty of Education at Central China Normal University, all majoring in Early Childhood Education. In order to measure the system's effectiveness quantitatively, all of these students were studying beginner English, and they were provided with a pre- and post-test before and after a week of using SPL for 30-40 minutes each day. The students were also provided with a survey asking questions grouped into effectiveness, engagement, adaptivity, satisfaction, and recommendation activity. By the end of the week, there was a significant improvement on the post-test. Effectiveness, engagement, and adaptivity were rated very high on the qualitative side, with recommendation accuracy just slightly below, and satisfaction received good but lower scores. The study showed the efficacy of the system, though there is no control group so it is hard to fully call it a causal relationship.
As the authors point out, it's interesting that the satisfaction scores were lower, and that highlights the importance of determining how to create systems that students actually want to use. It's unsurprising that they provided results, but figuring out how to design these systems in ways that can not just become tools for efficiency, but student-desired tools for better experiences in their education is paramount. My last review was about another ITS system that used text as an interface, and while one of the major issues of SPL is that the use of ChatGPT like this makes the system not applicable in lower resource environments, it's likely that a text-based system would score even lower on satisfaction. I will be interested in continuing to explore this idea of how to make these systems integrate into classrooms in ways that give students useful personalized attention, but don't detract from or change for the worse pedagogical practices that exist now.
The authors create an example ITS called the Socratic Playground for Learning (SPL), which implements this system. Through a user interface, students are able to interact with these models by defining, through a menu or a text box, what they are learning about or practicing. The student then engages in back-and-forth Socratic-style conversation. The questions presented to them dynamically scale in difficulty and complexity depending on their answers and responses.
To test this system, the authors involved 30 first-year students from the Faculty of Education at Central China Normal University, all majoring in Early Childhood Education. In order to measure the system's effectiveness quantitatively, all of these students were studying beginner English, and they were provided with a pre- and post-test before and after a week of using SPL for 30-40 minutes each day. The students were also provided with a survey asking questions grouped into effectiveness, engagement, adaptivity, satisfaction, and recommendation activity. By the end of the week, there was a significant improvement on the post-test. Effectiveness, engagement, and adaptivity were rated very high on the qualitative side, with recommendation accuracy just slightly below, and satisfaction received good but lower scores. The study showed the efficacy of the system, though there is no control group so it is hard to fully call it a causal relationship.
As the authors point out, it's interesting that the satisfaction scores were lower, and that highlights the importance of determining how to create systems that students actually want to use. It's unsurprising that they provided results, but figuring out how to design these systems in ways that can not just become tools for efficiency, but student-desired tools for better experiences in their education is paramount. My last review was about another ITS system that used text as an interface, and while one of the major issues of SPL is that the use of ChatGPT like this makes the system not applicable in lower resource environments, it's likely that a text-based system would score even lower on satisfaction. I will be interested in continuing to explore this idea of how to make these systems integrate into classrooms in ways that give students useful personalized attention, but don't detract from or change for the worse pedagogical practices that exist now.
Jesse Noppe-Brandon 11/10/2025
Yannier, N., Hudson, S. E., & Koedinger, K. R. (2020). "Active Learning is About More Than Hands-On: A Mixed-Reality AI System to Support STEM Education." International Journal of Artificial Intelligence in Education, 30(1), 74-96.
The study's most effective approach, the Combined condition, interleaves hands-on construction with guided discovery to let students "learn the theory" and then "apply the theory". The AI acts as an intelligent tutor, orchestrating a structured predict-explain-observe-explain cycle to provide real-time, personalized feedback. This approach, enabled by AI, ensures that children don't just tinker; they actively engage with scientific principles. It led to a more than 10-fold improvement in tower-building skills compared to the minimally-guided hands-on condition alone, proving that adding interactive guidance to hands-on experiences makes them more effective for learning. Guidance for hands-on activities is often inconsistent in learning environments like museums or even in schools. Not all parents, museum staff, or teachers have the necessary scientific background to provide knowledgeable and consistent support to students. This variability in support creates an unequal learning experience for children.
Muhammad Farhan 09/25/2025
Qurat-ul-Ann Mirza, N. Sanjay Rebello. "Help or Hype? Students' Engagement and Perception of Using AI to Solve Physics Problems." Physics Education Research Conference (PERC). 2025.
This study explores how students use ChatGPT during a physics problem-solving task embedded in a formal assessment. The findings indicated that students who engaged with ChatGPT generally performed better than those who did not. In particular, students who provided more complete and contextual prompts experienced greater benefits. There was also the presence of incorrect AI-generated responses, which calls for evaluating AI output. These results suggest that while AI can be a valuable aid in problem solving, its effectiveness depends significantly on how students use it, reinforcing the need to incorporate structured AI literacy into STEM education.
Dan Ahimbisibwe 10/08/2025
Vieriu, A. M., & Petrea, G. (2025). The impact of artificial intelligence (AI) on students' academic development. Education Sciences, 15(3), 343.
This study was conducted on 85 students at the National University of Science and Technology POLITEHNICA Bucharest, Romania, focusing on second-year Aerospace and Medical Engineering students with direct experience in AI-integrated learning environments. The study uses purposive sampling and a structured questionnaire with both closed and open-ended questions to evaluate AI usage, perceptions, and its impact on learning. Data was collected through a structured questionnaire comprising 11 items: seven closed-ended questions assessing perceptions, usage, and the effectiveness of AI tools; and four open-ended questions exploring experiences, expectations, and concerns.
It finds that 95.6% of students use AI in academic activities through virtual assistants (88.2%) and educational platforms (42.4%). Most of the participants use these tools weekly or daily, making AI an integral part of the educational process for tasks like homework, projects, and knowledge enhancement. While 80% of respondents agree that AI enhances their educational experience and 82.4% believe it improves academic performance, some express uncertainty or skepticism about its ultimate effects, including concerns about reducing critical thinking and accuracy issues. When asked for suggestions on improvement, most students suggest proper integration of AI in education, focusing on guided use that doesn't limit critical thinking. Overall, the majority of participants highlight both the positive influence on collaboration and learning efficiency, as well as challenges such as the need for reliable information and ethical use.
This paper gives valuable insight into the real-time experiences and perceptions of students who have direct exposure to AI-integrated learning environments, rather than relying solely on expert or educator perspectives. By using mixed-methods analysis across both closed and open-ended survey questions, it provides a nuanced account of how students engage with artificial intelligence for academic activities, covering not just usage statistics but also the emotional, ethical, and cognitive challenges and benefits they encounter. The study emphasizes student-driven recommendations for balanced and effective AI integration, including calls for enhanced training, reliability, and critical engagement. This makes it a key resource for educators and policymakers aiming to understand and responsibly advance AI adoption in education, learning from the people who would be most affected by these policies.
It finds that 95.6% of students use AI in academic activities through virtual assistants (88.2%) and educational platforms (42.4%). Most of the participants use these tools weekly or daily, making AI an integral part of the educational process for tasks like homework, projects, and knowledge enhancement. While 80% of respondents agree that AI enhances their educational experience and 82.4% believe it improves academic performance, some express uncertainty or skepticism about its ultimate effects, including concerns about reducing critical thinking and accuracy issues. When asked for suggestions on improvement, most students suggest proper integration of AI in education, focusing on guided use that doesn't limit critical thinking. Overall, the majority of participants highlight both the positive influence on collaboration and learning efficiency, as well as challenges such as the need for reliable information and ethical use.
This paper gives valuable insight into the real-time experiences and perceptions of students who have direct exposure to AI-integrated learning environments, rather than relying solely on expert or educator perspectives. By using mixed-methods analysis across both closed and open-ended survey questions, it provides a nuanced account of how students engage with artificial intelligence for academic activities, covering not just usage statistics but also the emotional, ethical, and cognitive challenges and benefits they encounter. The study emphasizes student-driven recommendations for balanced and effective AI integration, including calls for enhanced training, reliability, and critical engagement. This makes it a key resource for educators and policymakers aiming to understand and responsibly advance AI adoption in education, learning from the people who would be most affected by these policies.
Nasrin Alanna Aidi 11/13/2025
Nguyen, A. "Editorial: Enhancing Student Engagement Through Artificial Intelligence (AI): Understanding the Basics, Opportunities, and Challenges." Journal of University Teaching and Learning Practice. 2024, Vol 21 No. 6.
This paper examines how AI can be effectively integrated into classrooms (and even outside classrooms) to promote meaningful and engaging learning. The potential can be seen in Interactive Teaching Aids or Intelligent Tutoring Systems that provide 24/7 learning, even outside traditional classrooms. This system mirrors one-to-one tutoring, where students can submit their work and the system adjusts the feedback based on their current level of understanding. But one important aspect that this paper includes is how crucial technology accessibility is for the widespread adoption and maintenance of this system in classrooms. On one hand, this system offers positive benefits to students who are able to use it, but on the other, this "digital divide" could actually worsen the education gap in more under-resourced educational environments.
What I thought was important to note in this paper was how it cited the need to include emotional and behavioral cues in AI, adapting to the learning environment based on the individual student's needs; this particular emphasis is because merely adjusting the difficulty of tasks based on student performance is insufficient to become a personalized tutor. Since we as humans need more than just efficient learning tools, but also crave more connection, it is important to recognize that GenAI has a long way to go before even threatening to replace teachers in the classroom.
Overall, the paper highlights the need for equitable AI use across diverse educational contexts: rural and urban schools, as well as among students with disabilities. Ensuring transparency and accountability in AI decision-making is crucial to building trust and promoting inclusive adoption of these technologies in education.
What I thought was important to note in this paper was how it cited the need to include emotional and behavioral cues in AI, adapting to the learning environment based on the individual student's needs; this particular emphasis is because merely adjusting the difficulty of tasks based on student performance is insufficient to become a personalized tutor. Since we as humans need more than just efficient learning tools, but also crave more connection, it is important to recognize that GenAI has a long way to go before even threatening to replace teachers in the classroom.
Overall, the paper highlights the need for equitable AI use across diverse educational contexts: rural and urban schools, as well as among students with disabilities. Ensuring transparency and accountability in AI decision-making is crucial to building trust and promoting inclusive adoption of these technologies in education.
Nasrin Alanna Aidi 10/23/2025
Luckin, R. "Nurturing human intelligence in the age of AI: Rethinking education for the future." Development and Learning in Organizations: An International Journal. 2025, 39(1), 1-4.
Artificial intelligence (AI) is rapidly transforming education by personalizing learning, supporting teachers, and automating administrative tasks. While AI tools can adapt lessons to fit individual student needs and give immediate feedback, they also free up teachers to focus more on creative teaching and student guidance, helping to build deeper relationships and individual support.
However, the paper stresses that education shouldn't just be about memorizing facts or passing standardized tests. Instead, schools need to cultivate "Learning Mastery"—helping students understand how they learn best, think critically, set goals, and reflect on their progress. Alongside this, "Knowledge Mastery" means achieving a deep, conceptual understanding of subjects, not merely recalling information, so students can transfer their knowledge to real-world challenges.
The article emphasizes the importance of ethical AI deployment, warning against issues like bias, privacy risks, and inequality. It calls for collaboration between educators, policymakers, and technologists to ensure AI benefits all students. Teachers need professional development focused on both using AI and nurturing the unique human qualities—creativity, empathy, independent thinking—that future societies and workplaces will need. In conclusion, schools must combine the best aspects of AI and human intelligence to prepare students for lifelong learning and success in an AI-driven world.
However, the paper stresses that education shouldn't just be about memorizing facts or passing standardized tests. Instead, schools need to cultivate "Learning Mastery"—helping students understand how they learn best, think critically, set goals, and reflect on their progress. Alongside this, "Knowledge Mastery" means achieving a deep, conceptual understanding of subjects, not merely recalling information, so students can transfer their knowledge to real-world challenges.
The article emphasizes the importance of ethical AI deployment, warning against issues like bias, privacy risks, and inequality. It calls for collaboration between educators, policymakers, and technologists to ensure AI benefits all students. Teachers need professional development focused on both using AI and nurturing the unique human qualities—creativity, empathy, independent thinking—that future societies and workplaces will need. In conclusion, schools must combine the best aspects of AI and human intelligence to prepare students for lifelong learning and success in an AI-driven world.
Debika Dharma Lingam 11/13/2025
Innovating EdTech
This section catalogs existing AI-powered educational tools to help innovators understand the current landscape and identify market gaps. Unlike other sections focused on research evaluation, these pages describe tools and their capabilities to inform product development and innovation strategy.
AI Tutoring & Learning Support 0
Tools that act as instructors, mentors, or study aids including tutorbots, homework helpers, and adaptive learning systems.
Writing & Content Generation Tools 0
Tools that help students create or refine text and media, including writing assistants and translation tools.
Detection Tools 0
Tools that monitor or verify the authenticity of work, including AI-writing detectors and plagiarism checkers.
Organizational & Productivity Tools 0
Tools that structure learning and manage large projects, such as note-taking and research assistants.
Feedback & Assessment Tools 0
AI that evaluates or comments on student work, including automated grading and feedback generators.
Accessibility & Inclusion Tools 0
AI that expands educational access through speech-to-text, captioning, and assistive technologies.
Classroom & Learning Analytics 0
AI systems that track and interpret educational activity, including analytics dashboards and early warning systems.
Scheduling 0
Tools that facilitate appointment and meeting scheduling for educational contexts.
Prototype Development
Creating and iterating on educational technology prototypes that leverage AI.
Kumar, S., Thompson, A., & Lee, J. (2024). "Human-centered design principles for AI-enhanced educational technologies." Proceedings of the ACM Conference on Learning @ Scale, 234-245.
This paper outlines key design principles for creating educational technologies that effectively integrate AI while maintaining pedagogical integrity. The authors emphasize transparency, user agency, and educational effectiveness over technological novelty. Essential reading for the Innovating EdTech subteam's prototype development work.
Alexander Landfair, January 2025
AI Tutoring & Learning Support
Tools that act as instructors, mentors, or study aids including tutorbots, homework helpers, and adaptive learning systems.
Context - Generative Learning Planner
Visit Site →
Pain Point
Teachers struggle to find time to create personalized learning plans for each student; students receive generic instruction that doesn't match their learning preferences.
What it does
A generative learning planner for higher education that creates AI-taught courses based on student preferences. Teachers upload material and teaching structure/vision to create learning plans; students receive personalized plans tailored to their needs.
Target Users
Teachers and students in higher education
Platform
Web application
Pricing
Freemium
Business Model
B2B2C (institutions → students)
Competitors
Coursera, Khan Academy, traditional LMS platforms
Key Differentiators
Specifically designed for higher education with dual-sided personalization—teachers define structure/vision while students receive individually tailored plans. Founded by NYU ITP post-graduate.
Cate Hackett 11/2025
Feedback & Assessment Tools
AI that evaluates or comments on student work, including automated grading and feedback generators.
Brisk AI - Multi-Function Browser Extension
Visit Site →
Pain Point
Teachers spend excessive time on tedious administrative tasks (grading, feedback, recommendation letters, IEP goals) that could be better spent on instruction.
What it does
A web/browser extension with functions including writing inspection, feedback & grading, presentation making, IEP goal generator, and recommendation letter generator.
Target Users
Teachers and students (K-12 focus)
Platform
Browser extension (Chrome), integrates with Google Docs, Slides, Classroom
Pricing
Free for teachers
Business Model
B2C / B2B (free individual, school plans)
Competitors
Magic School AI, Diffit, Eduaide.ai
Key Differentiators
Browser-based approach that works within existing workflows (Google Suite). Exhaustively presents AI possibilities through extensive function menu—a "quick option" effect that shows teachers AI uses they wouldn't think of themselves.
Cate Hackett 11/2025
Organizational & Productivity Tools
Tools that structure learning and manage large projects, such as note-taking and research assistants.
Almanack - Course Asset Generator
Visit Site →
Pain Point
Teachers spend hours creating quality worksheets, activities, and slide-decks instead of focusing on direct instruction.
What it does
Generates course assets including worksheets, activities, assessments, and slide-decks.
Target Users
Teachers (K-12 focus)
Platform
Web application
Pricing
Freemium (limited free tier, paid plans)
Business Model
B2C / B2B (freemium individual, school plans)
Competitors
Magic School AI, Canva for Education, Google Slides + AI tools
Key Differentiators
Focused specifically on asset generation (worksheets, slides, assessments) rather than broader AI assistant functions. Higher education applicability unclear—primarily K-12 oriented.
Cate Hackett 11/2025
Magic School AI - Lesson Planning & Assessment Suite
Visit Site →
Pain Point
Teachers are overwhelmed by lesson planning and IEP writing; school administrators need AI tools they can trust with student data.
What it does
Lesson planning, assessments, and IEP writing capabilities with 60+ AI tools for educators.
Target Users
Teachers (K-12 focus)
Platform
Web application, Chrome extension
Pricing
Free tier, paid school/district plans
Business Model
B2B / B2C (school/district focus, free individual tier)
Competitors
Brisk AI, Almanack, Eduaide.ai, Diffit
Key Differentiators
Heavy emphasis on "safeguards," evidence basis, and data security—strong trust-building messaging. Comprehensive suite approach with 60+ specialized tools. Higher education applicability unclear—primarily K-12 oriented.
Cate Hackett 11/2025
Writing & Content Generation Tools
Tools that help students create or refine text and media, including writing assistants and translation tools.
QuillBot - AI Writing Assistant
Visit Site →
Pain Point
Writers struggle with paraphrasing, grammar, and refining their work; students need writing assistance that helps them improve rather than writes for them.
What it does
AI writing-specific tool that provides pop-up style recommendations as you write, including paraphrasing, grammar checking, and summarization.
Target Users
Teachers, students, and general users
Platform
Web app, browser extension, MS Word add-in
Pricing
Freemium (limited free, premium subscription)
Business Model
B2C (individual subscriptions)
Competitors
Grammarly, Wordtune, ProWritingAid, Hemingway Editor
Key Differentiators
Markets itself as a collaborator rather than "writing for you"—emphasizes responsible AI use. Inline recommendations appear as you write. Strong paraphrasing focus. Inclusive but not specifically tuned for formal academic settings.
Cate Hackett 11/2025
Detection Tools
Tools that monitor or verify the authenticity of work, including AI-writing detectors and plagiarism checkers.
Content in Development
VIP Innovating EdTech subteam is currently curating resources on detection tools. We welcome your contributions and findings.
Accessibility & Inclusion Tools
AI that expands educational access through speech-to-text, captioning, and assistive technologies.
Content in Development
VIP Innovating EdTech subteam is currently curating resources on accessibility and inclusion tools. We welcome your contributions and findings.
Classroom & Learning Analytics
AI systems that track and interpret educational activity, including analytics dashboards and early warning systems.
SchoolAI - Integrated School Platform
Visit Site →
Pain Point
School administrators lack visibility into AI usage across their institution; teachers need AI tools while administrators need oversight and analytics on student/teacher performance.
What it does
Comprehensive platform with administrative oversight capabilities including rule setting and insight to student/teacher performance. Interface includes quick options like rubric generator, multiple choice quiz builder, and text translator.
Target Users
Teachers, school leaders, students, and administrators
Platform
Web application
Pricing
Free for teachers, paid district/school plans
Business Model
B2B (school/district contracts)
Competitors
Magic School AI, Khanmigo, traditional LMS platforms with AI add-ons
Key Differentiators
All-in-one platform where students, teachers, and administrators coexist. Has higher education tab. Quick options give teachers AI ideas. However, raises concerns about surveillance and data collection when building infrastructure with administrative oversight of all users.
Cate Hackett 11/2025
Scheduling
Tools that facilitate appointment and meeting scheduling for educational contexts.
Content in Development
VIP Innovating EdTech subteam is currently curating resources on scheduling tools. We welcome your contributions and findings.
Faculty Perspectives
Faculty attitudes toward AI integration and their teaching practice adaptations.
Content in Development
VIP Understanding Users subteam is currently curating resources for this research area. We welcome your contributions and findings.
Administrator Views
Administrative perspectives on AI policy, implementation, and institutional change.
Content in Development
VIP Understanding Users subteam is currently curating resources for this research area. We welcome your contributions and findings.
AI Competencies & Literacy
Assessment of AI literacy levels and competency development across user groups.
Long, D., & Magerko, B. (2020). "What is AI literacy? Competencies and design considerations." Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, 1-13.
This foundational paper defines AI literacy as a set of competencies enabling individuals to critically evaluate AI technologies, communicate effectively with AI, and use AI as a tool online, at home, and in the workplace. The framework includes understanding AI capabilities, limitations, and ethical implications.
Alexander Landfair, January 2025
GenAI Tool Capabilities
Assessment of what current generative AI tools can and cannot do for education.
Bommasani, R., Hudson, D. A., Adeli, E., et al. (2021). "On the opportunities and risks of foundation models." arXiv preprint arXiv:2108.07258.
This comprehensive analysis examines foundation models (like GPT) across multiple dimensions including capabilities, limitations, and societal impacts. Essential reading for understanding what these models can and cannot do in educational contexts, including issues of bias, reliability, and appropriate use cases.
Alexander Landfair, January 2025
Tool Limitations & Biases
Understanding the constraints, biases, and failure modes of AI educational tools.
Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., & Galstyan, A. (2021). "A survey on bias and fairness in machine learning." ACM Computing Surveys, 54(6), 1-35.
This survey provides a comprehensive overview of different types of bias in machine learning systems, sources of bias, and mitigation strategies. Particularly relevant for educational AI tools where bias can significantly impact student outcomes and educational equity.
Alexander Landfair, January 2025
Comparative Tool Analysis
Side-by-side comparisons of different AI tools for specific educational tasks.
Content in Development
VIP Evaluating Tools subteam is currently curating comparative studies. We welcome your tool evaluation research and findings.
Academic Integrity
Assessment of AI's effect on academic honesty and institutional integrity measures.
Eaton, S. E. (2023). "Postplagiarism: Transdisciplinary ethics and integrity in the age of artificial intelligence and neurotechnology." International Journal for Educational Integrity, 19(1), 1-10.
Eaton proposes a "postplagiarism" framework for academic integrity in the AI era, arguing that traditional notions of plagiarism are insufficient for addressing AI-assisted work. The paper calls for new ethical frameworks that focus on learning outcomes rather than detection of AI use.
Alexander Landfair, January 2025
Seco, D.; Grösser, S.; Pedrosa, A.M. "Use of Generative Artificial Intelligence tools in higher education environments." Multidisciplinary Journal for Education, Social and Technological Sciences. 2025, pages 156-175.
This paper examines the integration of Gen AI into a university setting by analyzing its current applications and collecting perspectives from stakeholders like students and professors. The study finds that while AI is becoming increasingly common, its use remains somewhat limited. Students view AI as "valuable resource for personalized learning", while educators are focused on adapting their teaching methods to maintain quality in this new technological landscape. The research mentions a central tension between the significant potential of AI and its balance against major ethical concerns and risks. Key challenges include a rise in academic dishonesty, as AI provides new ways to cheat and plagiarize, and the risk of over-reliance on these tools, which can lead to negative cognitive effects like procrastination, memory decline, and decreased academic performance. The authors conclude that for AI to be integrated successfully, educators must evolve their assessment strategies to address these challenges, ensuring a student-centered approach that balances the technology's benefits and risks.
The concerns raised by Seco et al. parallel the arguments from the Veritasium video and the issues of bias detailed by Baker & Hawn (2022). While the former two focus on how AI threatens the process of learning by eliminating cognitive effort and undermining academic integrity, Baker & Hawn reveal how AI can corrupt the equity of learning outcomes. Together, they expose a potential double bind for students, especially those from marginalized groups. A student might use an AI tool that allows them to bypass the "desirable difficulties" essential for deep system 2 thinking, and lead to academic dishonesty. Then, the same AI system could evaluate their work through a biased lens, systematically disadvantaging them based on their race, gender, or socioeconomic background. This creates a scenario where AI not only fails to support genuine learning but also actively perpetuates systemic inequality. Thus, the solution should be two-pronged, where we need the pedagogical shifts that Seco et al. and the Veritasium video imply, such as redesigning assessments to focus on process, but we also desperately need the technical and policy-level interventions that Baker & Hawn advocate for to ensure the tools themselves are built on principles of fairness and equity.
Josephine Li 10/01/2025
The paper's insights align with urging institutions and professors to balance rapid adaptation of generative AI. Not only it offers personalized learning but also deep understanding of concepts that's crucial for employability. Seco et al also recommends student centered approaches for conducting assessments to foster deep reflective thinking. The authors emphasize a central tension: AI enhances flexibility and accessibility but also introduces serious concerns, including the risk of academic dishonesty, overreliance, and negative cognitive impacts. The paper concludes that universities must strategically balance innovation with integrity, guiding the sector toward inclusive, process-oriented, and human-centered education in the AI era.
Debika Dharma Lingam 10/9/2025
Lu, N., Liu, S., He, R., Ong, Y.-S., Wang, Q., & Tang, K. (2023). Large language models can be guided to evade AI-generated text detection [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2305.10847
The paper introduces a concept named SICO (Substitution-based In-Context example Optimization). It is a novel and cost-efficient method for constructing prompts that guide large language models (LLMs) to evade AI-generated text detectors. The method automatically optimizes a few human-written examples to be less "AI-like", which are then used as in-context demonstrations for the LLM.
The research's strength is in demonstrating a significant vulnerability in existing detection systems. The study shows that prompts generated by SICO enabled GPT-3.5 to successfully bypass six different detectors, significantly outperforming baseline paraphrasing methods. A comprehensive human evaluation further confirmed that the SICO-generated text maintained human-level readability and task completion rates.
This work is valuable because it highlights that evasion can be achieved through simple prompt-guiding rather than external tools, and it provides a new, effective benchmark (SICO) for evaluating the robustness of future detectors.
It also connects well with several past papers I've found that looks into prompt engineering and using that to bypass AI-detection tools (through targeting perplexity and burstiness).
The research's strength is in demonstrating a significant vulnerability in existing detection systems. The study shows that prompts generated by SICO enabled GPT-3.5 to successfully bypass six different detectors, significantly outperforming baseline paraphrasing methods. A comprehensive human evaluation further confirmed that the SICO-generated text maintained human-level readability and task completion rates.
This work is valuable because it highlights that evasion can be achieved through simple prompt-guiding rather than external tools, and it provides a new, effective benchmark (SICO) for evaluating the robustness of future detectors.
It also connects well with several past papers I've found that looks into prompt engineering and using that to bypass AI-detection tools (through targeting perplexity and burstiness).
Josephine Li 11/13/2025
Ardito, C. G. (2025). Generative AI detection in higher education assessments. New Directions for Teaching and Learning, 2025, 11–28.
This paper explores the pitfalls of AI Detection software and offers remedies to the problem. It explains that AI Detection is both fragile and misaligned with how education is today. It states that detection with these softwares is unreliable, as students are able to avoid it by paraphrasing AI generated work, while non-fluent English writers are being flagged as their writing uses common English sentences. These inconsistencies make it almost impossible to prove whether AI is used or not, and often lead to false positives, and leave students anxious about their writing and left to deal with proving themselves to their professors that AI was not used. The author also disproves watermarking, a technique experts have suggested to help prove AI was used in writing by having GenAI tools leave a hidden signature in their writing. An example of a signature could be by first dividing a dictionary of words into 2 sets, and then alternating between the two when generating text. AI detection tools would then search for this signature and give a score based on that. However, watermarking isn't immune to paraphrasing, and additionally, there are already models released without watermarking, and implementing it would just lead to a new market of non-watermarked models, essentially making it hard to regulate.
This paper recommends shifting school work to in-person centered assessments, presentations, and having on-going check-ins for writing. This could limit the use of AI and lead to students writing on their own. Another approach to this issue, the paper suggests having transparent disclosures on the proper uses of AI. Having these set rules will help properly integrate AI into learning. As the paper mentions, AI mirrors the invention of the calculator, as like AI, it was at first looked down upon and considered cheating and banned in schools. However, a gradual shift in perception occurred where calculators were later accepted in schools and they began regulating its use and now see it as a tool to make mathematics more efficient. I think once we are past this period of AI slander and begin to see its potentials, creating clear and transparent guidelines for its use will bring down the uses for cheating and bring down our reliance on unreliable AI Detection tools.
This paper recommends shifting school work to in-person centered assessments, presentations, and having on-going check-ins for writing. This could limit the use of AI and lead to students writing on their own. Another approach to this issue, the paper suggests having transparent disclosures on the proper uses of AI. Having these set rules will help properly integrate AI into learning. As the paper mentions, AI mirrors the invention of the calculator, as like AI, it was at first looked down upon and considered cheating and banned in schools. However, a gradual shift in perception occurred where calculators were later accepted in schools and they began regulating its use and now see it as a tool to make mathematics more efficient. I think once we are past this period of AI slander and begin to see its potentials, creating clear and transparent guidelines for its use will bring down the uses for cheating and bring down our reliance on unreliable AI Detection tools.
Rohan Kapur 11/13/2025
Assessment Validity
How AI tools impact the reliability and validity of educational assessments and grading practices.
Owan, V. J., Abang, K. B., Idika, D. O., Etta, E. O., Bassey, B. A. "Exploring the Potential of Artificial Intelligence Tools in Educational Measurement and Assessment." EURASIA Journal of Mathematics, Science and Technology Education. 2023, em2307.
This paper explores the potential of AI tools in educational measurement and assessment, examining how artificial intelligence can enhance the reliability, validity, and efficiency of assessment practices in educational settings.
VIP Assessing Impacts Team
Student Experience + Wellbeing
Impact of AI on student stress, motivation, confidence, and overall educational experience.
Gerlich, M. "AI Tools in Society; Impacts on Cognitive Offloading and the Future of Critical Thinking." Societies 2025, 15, 6.
This study looks at the implications of the overreliance on AI tools to users' critical thinking abilities. The study sought to explore two central claims: That high use of AI is correlated to decreased critical thinking abilities, and that cognitive offloading mediates the relationship between AI tool usage and critical thinking. The results of the study found that individuals who self-identified as having high reliance on AI tools consistently scored lower on the critical thinking tests than those who did not declare such a reliance. The study urged that as a result of the tested linkage between AI usage and decreased critical thinking, that precautionary protocols should be implemented to avoid overall degradation of individuals cognitive functions.
Elan Hebert 10/16/2025
Equity & Access Issues
How AI tools affect educational equity and access across different student populations.
Rahimi, S., & Chen, L. (2024). "The AI divide: How socioeconomic factors influence access to and benefits from educational AI tools." Computers & Education, 205, 105112.
This research examines how socioeconomic factors create disparities in AI tool access and effectiveness. The study finds that while AI tools promise to democratize education, they may actually exacerbate existing inequalities without intentional equity-focused implementation strategies.
Alexander Landfair, January 2025
Baker, R.S.; Hawn, A. "Algorithmic Bias in Education." Int J Artif Intell Educ 2022, 32, 1052–1092.
It is clear that algorithmic bias can be harmful; from racial bias in sentencing decisions to gender bias in assigning credit limits. Bias can arise in the data collection, data preparation, model development, model evaluation, model post-processing, and model deployment stages. Having under-sampled groups in training data can result in representational bias. This follows in education, requiring modifying algorithms to be fair and reducing bias when predicting GPAs or dropout rates. Additionally, omitting data like socioeconomic background can result in fairer models. A current challenge in machine learning is determining how many members of an underrepresented group must be included in a combined model so that it is fair and reliable. Additionally, some underrepresented groups have yet to be studied, like children of immigrants or religious minorities. Ultimately, Groups of learners should not be oversimplified in educational datasets where feasible and can improve labeling. Although we cannot optimize for all definitions of fairness, there is still opportunity to improve significantly, as many algorithms do not even consider fairness. Unfortunately, there is "disincentive for the developers of learning systems to reveal the algorithmic biases in their systems" due to privacy and critique. Some solutions going forward are to include the use of reference data and avoiding bias in the labeling process by allowing the group being studied to conduct the data labeling.
Karen Maza Delgado 9/24/2025
This article emphasizes higher education institutions' responsibilities in promoting ethical AI usage in classrooms. One key point the authors made was that penalizing students for using AI is less effective than raising awareness around the implications of overreliance on AI to prevent academic dishonesty from happening in the first place. On the flip side, previous research has shown that AI can assist educators as well, in tasks such as lesson planning and trend-mapping. The article suggests four main ways that AI can be used in the classroom as a tool, including creating role-play scenarios to prepare students for stressful workplace environments. Although this was not mentioned in the article, I would also argue that AI can be especially helpful in language learning classes, as it would reduce stress on the professor to simulate conversational exams, since an AI tool can be used to do so. A pattern I am seeing in both the real world and the article is that AI has the biggest comparative advantage when it comes to personalizing a student's learning plan. Whether it be conversations in a different language or degree mapping, AI is a powerful tool that is at every student's hands.
Yufei Huang 10/03/2025
Saltman, K.J. "Artificial intelligence and the technological turn of public education privatization: In defence of democratic education." London Review of Education, 2020, 18 (2): 196–208.
This article highlights what the EdTech sphere does wrong. The author claims adaptive technology doesn't have as strong positive standardized tests results that we expect it to. He mentions that young students are not very interested in 2d digital learning, and also makes claims about how biometric pedagogies (using webcams to see if students are paying attention/responding well to lessons) are problematic. The root of what he is saying is that the increasing investment towards private EdTech is only to the detriment of true learning. His current qualms with the current integration of AI in Education sheds light on ways we can work around the detrimental, for-profit, neoliberal frameworks regarding tech in the education system.
Sabria Islam 10/16/2025
User Testing & Feedback
Methods for testing educational prototypes with real users and gathering feedback.
Nielsen, J., & Gilutz, S. (2024). "Usability testing methods for educational AI applications." Proceedings of the CHI Conference on Human Factors in Computing Systems, 456-467.
This paper presents adapted usability testing methods specifically for educational AI applications. The authors emphasize the importance of testing with actual students and educators rather than general users, and provide frameworks for evaluating both usability and educational effectiveness.
Alexander Landfair, January 2025
Integration Strategies
Approaches for integrating new AI technologies into existing educational systems.
Content in Development
VIP Innovating EdTech subteam is currently curating integration methodologies. We welcome your implementation research and case studies.
Scalability Assessment
Evaluating the potential for scaling educational technology innovations.
Clarke, J., & Dede, C. (2024). "Factors influencing the scalability of educational technology innovations." Computers & Education, 210, 105234.
This research identifies key factors that determine whether educational technology innovations can be successfully scaled. The authors present a framework for assessing technical, pedagogical, organizational, and financial scalability dimensions, with specific attention to AI-enhanced tools.
Alexander Landfair, January 2025
Research Methods & Resources
These sources will help you identify a research gap, design appropriate research methodologies, interpret your findings, and find publication in a peer-reviewed journal.
Literature Review Methods 0
Systematic approaches to finding, evaluating, and synthesizing existing research in AI education.
Collaborative Research 0
Best practices for team-based research, project management, and interdisciplinary collaboration.
Experimental Design 0
Methods for designing controlled studies to test AI tool effectiveness and educational interventions.
Survey Design & Analysis 0
Best practices for creating effective surveys and analyzing quantitative data in educational research.
Interview & Focus Group Methods 0
Techniques for conducting qualitative interviews and focus groups with students, faculty, and administrators.
Data Analysis Techniques 0
Statistical and qualitative analysis methods for educational research data and findings interpretation.
Research Ethics & IRB 0
Ethical considerations and IRB approval processes for educational research involving human subjects.
Academic Writing & Publication 0
Guidelines for writing research papers, conference presentations, and academic publications.
Literature Review Methods
Systematic approaches to finding, evaluating, and synthesizing existing research in AI education.
Content in Development
VIP Research Methods subteam is currently curating resources on literature review methodologies. We welcome your contributions and findings.
Collaborative Research
Best practices for team-based research, project management, and interdisciplinary collaboration.
Content in Development
VIP Research Methods subteam is currently curating resources on collaborative research practices. We welcome your contributions and findings.
Experimental Design
Methods for designing controlled studies to test AI tool effectiveness and educational interventions.
Content in Development
VIP Research Methods subteam is currently curating resources on experimental design. We welcome your contributions and findings.
Survey Design & Analysis
Best practices for creating effective surveys and analyzing quantitative data in educational research.
Content in Development
VIP Research Methods subteam is currently curating resources on survey design and analysis. We welcome your contributions and findings.
Interview & Focus Group Methods
Techniques for conducting qualitative interviews and focus groups with students, faculty, and administrators.
Content in Development
VIP Research Methods subteam is currently curating resources on interview and focus group methods. We welcome your contributions and findings.
Data Analysis Techniques
Statistical and qualitative analysis methods for educational research data and findings interpretation.
Content in Development
VIP Research Methods subteam is currently curating resources on data analysis techniques. We welcome your contributions and findings.
Research Ethics & IRB
Ethical considerations and IRB approval processes for educational research involving human subjects.
Content in Development
VIP Research Methods subteam is currently curating resources on research ethics and IRB processes. We welcome your contributions and findings.
Academic Writing & Publication
Guidelines for writing research papers, conference presentations, and academic publications.
Content in Development
VIP Research Methods subteam is currently curating resources on academic writing and publication. We welcome your contributions and findings.
Learning Science
Background readings in the science of teaching and learning.
Writing & Composition 0
Cognitive and developmental theories of writing processes and instruction.
Cognitive Science 0
Research on cognitive load, memory, creativity, and learning processes.
Reading & Literacy 0
Deep reading, textual interpretation, and the ethics of literary engagement.
Educational Philosophy 0
Philosophical foundations of education, purpose, measurement, and teacher agency.
AI/ML Foundations 0
Technical foundations of AI and machine learning relevant to education.
Writing & Composition
Cognitive and developmental theories of writing processes and instruction.
Flower, L. (1979). "Writer-based prose: A cognitive basis for problems in writing." College English, 41(1), 19-37.
In this seminal paper, Flower describes the core challenge of writing from a cognitive psychological perspective. In doing so, she challenges the dominant theories of writing development practiced by expressionists like Peter Elbow. It begins with a deceptively simple question, "If writing is simply the act of 'expressing what you think' or 'saying what you mean, why is writing often such a difficult thing to do?" Along with Janet Emig's work, Flower's early articles marked the beginning of composition's cognitive turn in the late 70s.
Alexander Landfair July 27, 2025
Flower explains that the difficulty in writing isn't a lack of ideas, but in "orchestrating" a cognitive transformation. Writing operates through networks, personal narratives, and words that are full of stories and meanings only we know. Many students, including myself, would input a raw, writer-based draft, and the AI can generate a more reader-oriented revision. However, this is the dangerous part. The real writing happens within the struggle of transforming internal monologue to words on paper.
Meiling Zhang 9/24/2025
Kellogg, R. T., & Raulerson III, B. A. (2007). "Improving the writing skills of college students." Psychonomic Bulletin & Review, 14(2), 237-242.
Advanced writing skills are an important aspect of academic performance as well as of subsequent work-related performance. In order to achieve higher levels of writing performance, the working memory demands of writing processes should be reduced so that executive attention is free to coordinate interactions among them. This can in theory be achieved through deliberate practice that trains writers to develop executive control through repeated opportunities to write and through timely and relevant feedback. Automated essay scoring software may offer a way to alleviate the intensive grading demands placed on instructors and, thereby, substantially increase the amount of writing practice that students receive.
Author abstract
McCutchen, D. (1988). "Functional Automaticity in Children's Writing: A Problem of Metacognitive Control." Written Communication. Vol. 5. No. 3. July 1988. P. 306-324.
McCutchen argues that the challenges inherent to writing are importantly different from most other cognitive tasks. In most domains, mastery happens when subtasks become automated, able to be performed subconsciously. The key point here is that the subtasks of writing are not really eligible for automaticity, because writing subtasks are not routine or generalizable in the same ways. She proposes, therefore, that writing mastery stems from fluency rather than automaticity. The difference is that fluency involves maintaining metacognitive control.
Alexander Landfair, August 5, 2025
McCutchen, D. (2000). "Knowledge, Processing, and Working Memory: Implications for a Theory of Writing." Educational Psychologist, 35:1, 13-23.
This article surveys writing research and attempts to sketch a principled account of how multiple sources of knowledge, stored in long-term memory, are coordinated during writing within the constraints of working memory. The concept of long-term working memory is applied to the development of writing expertise. Based on research reviewed, it is speculated that lack of fluent language generation processes constrains novice writers within short-term working memory capacity, whereas fluent encoding and extensive knowledge allow skilled writers to take advantage of long-term memory resources via long-term working memory.
Author abstract
Sommers, N., & Saltz, L. (2004). "The Novice as Expert: Writing the Freshman Year." College Composition and Communication, 56(1), 124-149.
In this longitudinal study, Sommers and Saltz identify the key characteristics of the most successful college writing students. "We argue that students who make the greatest gains as writers throughout college (1) initially accept their status as novices and (2) see in writing a larger purpose than fulfilling an assignment." The paper explains how first-year writing classes foster these mindsets that promote the growth of writing skills.
Alexander Landfair, Jul 27, 2025
Cognitive Science
Research on cognitive load, memory, creativity, and learning processes.
Sweller, J. (2020). "Cognitive load theory and educational technology." Educational Technology Research and Development, 68(1), 1-16.
Here, Sweller argues for cognitive load theory (which describes the limitations of our working memory) as a central concern in classroom design. "Technology-based instruction used without reference to the instructional design principles that flow from human cognition," he says, "is likely to be random in its effectiveness." What a mic drop!
Alexander Landfair, Jul 27, 2025
Marrone, R., Taddeo, V., & Hill, G. (2022). "Creativity and Artificial Intelligence—A Student Perspective." Journal of Intelligence, 10(3), 65.
This article explores how AI is involved with education by looking at focus groups of secondary school children. Their findings are similar to our current class discussions about AI, but it is especially insightful because these focus on younger students. The results point to 4 buckets: technological, social, affective, and learning. Students believed that AI could harm social skills, making people less capable of real-world communication, while others saw potential for AI as connection if used collaboratively. Their affective responses varied, students who understood AI felt comfortable and trusting toward it, whereas those unfamiliar with it felt anxious or fearful. They often equated AI with robots and futuristic machines, showing a limited understanding of its everyday uses. Overall, they viewed AI as technically advanced but not truly creative, though it could inspire or "spark" human creativity. The researchers also proposed a 4AI model that mirrors the 4C models of learning: Mini-AI (personal use), Little AI (creative collaboration with AI), Big-AI (Advanced AI in professional use), and Legendary-AI (transformative), which students often appreciated the applications of Big/Legendary (describing them as "futuristic robots"), but undervalued everyday AI. The researchers point to educators teaching with AI in mind.
Sabria Islam 10/9/2025
This qualitative grounded theory study explored secondary students' perceptions of artificial intelligence (AI) and creativity following an eight-week hands-on curriculum involving both domains. The analysis, based on 12 focus groups and 8 interviews, classified into four central dimensions of student thinking: social, affective, technological, and learning factors. Most students considered AI a useful augmentative tool, not a replacement for human creativity—yet their conception of AI was often limited to "robotic" or extraordinary systems, with little awareness of everyday AI.
To frame the findings, the researchers mapped student conceptions of AI to a new "4AI Model," closely paralleling Kaufman and Beghetto's established 4C model of creativity (mini-c, little-c, Pro-c, Big-C). The 4C model describes creativity along a developmental spectrum, from personal insight (mini-c) to world-changing innovation (Big-C); similarly, the 4AI approach recognizes that students engage with AI at multiple levels, ranging from everyday feedback tools (Mini-AI) to legendary, society-transforming systems (Legendary-AI). Interpreting the results through this model shows that while students readily appreciate the marvels of advanced AI (e.g., AlphaGo or GPT), they undervalue—or simply overlook—the creative potential in routine, embedded AI applications that scaffold daily learning. Thus, the study provides empirical support for AI literacy curricula that not only demystify AI technology, but also teach students to recognize and leverage the full spectrum of AI–creativity interplay across classroom and professional contexts.
Debika Dharma Lingam 10/15/2025
Reading & Literacy
Deep reading, textual interpretation, and the ethics of literary engagement.
Hendrick, C. (2025). "The Humility of the Page: The Lost Ethics of Deep Reading." The Learning Dispatch.
In this anecdotal article learning scientist Carl Hendrick laments the lost art of deep reading. He argues that reading is a moral and ethical duty in that it is the practice of paying attention to the general other while it also transforms the participator into a more conscientious citizen. He argues that the nature of 'soulless' social media algorithms are at least one huge contributor to the problem of making us more like machines - shallower, and less compassionate - and therefore less human. The article connects to Mikhail Bakhtin's theories of Dialogism, Heteroglossia, and Polyphony, and to Pierre Hadot's philosophy: 'Only he who is capable of a genuine encounter with the other is capable of an authentic encounter with himself.'
Paul Marinos, Sep 26, 2025
Zhernokleyev, D. (2021). "Dostoevskii, the Feuilleton and the Confession." Slavonic and East European Review, pp 71-97.
Zhernokleyev analyzes communication, formalism, and epistemology in the work of Fyodor Dostoevsky. He argues that, from an epistemological level, communication itself is inherently violent, and that 'reality strives towards fragmentation'. He argues that there is an 'epistemic fallenness' inherent in our language and imagination, and he cites our givenness to fascination - deliberately taken advantage of with the genre of the Feuilleton - as being the primary reason for distraction, miscommunication, and fragmentation. This paper, like Dostoevsky, is a deep, and often dark, read, and if you argue that the depth is out of scope for our research into AI in Education and Learning Science, then perhaps you'd be right, but I would argue that such psychological depth is not only relevant but also necessary for an adequate exploration of these topics.
Paul Marinos, Sep 26, 2025
Educational Philosophy
Philosophical foundations of education, purpose, measurement, and teacher agency.
Biesta, G. (2015). "Beautiful risk of Education."
This is a book about what many teachers know but are increasingly being prevented from talking about: that real education always involves a risk. The risk is there because, as WB Yeats has put it, education is not about filling a bucket but about lighting a fire. It is there because students are not to be seen as objects to be moulded and disciplined, but as subjects of action and responsibility. The Beautiful Risk of Education is organised around a critical discussion of seven key educational concepts: creativity, communication, teaching, learning, emancipation, democracy, and virtuosity. By opposing the risk aversion that characterises many contemporary educational policies and practices, Gert JJ Biesta makes a strong argument for giving risk a central place in our educational endeavours and brings risk taking to the forefront of a critical pedagogical practice.
Paul Marinos 10/18/2025
Biesta, G. (2015). "Good education in an age of measurement: Ethics, politics, democracy."
The widespread use of the measurement of educational outcomes in order to compare the performance of education within and across countries seems to express a real concern for the quality of education. This book argues that the focus on the measurement of educational outcomes has actually displaced questions about educational purpose. Biesta explores why the question as to what constitutes good education has become so much more difficult to ask and shows why this has been detrimental for the quality of education and for the level of democratic control over education. He provides concrete suggestions for engaging with the question of purpose in education in a new, more precise and more encompassing way, with explicit attention to the ethical, political and democratic dimensions of education.
Paul Marinos 10/18/2025
Biesta, G., Priestley, M., & Robinson, S. (2015). "The role of beliefs in teacher agency."
There is an ongoing tension within educational policy worldwide between countries that seek to reduce the opportunities for teachers to exert judgement and control over their own work, and those who seek to promote it. Some see teacher agency as a weakness within the operation of schools and seek to replace it with evidence-based and data-driven approaches, whereas others argue that because of the complexities of situated educational practices, teacher agency is an indispensable element of good and meaningful education. In this paper, the authors draw from a two-year study into teacher agency against the backdrop of large-scale educational reform – the implementation of Scotland's Curriculum for Excellence.
Paul Marinos 10/18/2025
Watanabe, A. (2024). "Have Courage to Use your Own Mind, with or without AI: The Relevance of Kant's Enlightenment to Higher Education in the Age of Artificial Intelligence." Journal of Critical Education Policy Studies. Vol. 22 No. 2 (2024): Special Issue: Artificial Intelligence (AI) in Education Part 1. 2024, 46-58.
This article explores the relevance of Kant's concept of Enlightenment to higher education in the age of AI, arguing that students must have the courage to use their own minds critically, with or without AI assistance.
Elan Hebert
AI/ML Foundations
Technical foundations of AI and machine learning relevant to education.
Radford, A., Narasimhan, K., Salimans, T., & Sutskever, I. (2018). "Improving language understanding by generative pre-training." OpenAI.
The paper addresses key challenges in natural language understanding. The study demonstrates that significant improvements can be achieved by first pre-training a language model on a large, diverse corpus of unlabeled text, then fine-tuning it discriminatively for each specific task. The Transformer architecture is chosen for its ability to handle long-term dependencies more effectively than recurrent networks. For unsupervised pre-training, the model uses the BooksCorpus dataset, consisting of 700 unpublished books across genres such as adventure, romance, and fantasy (about 1 billion words in total), enabling the model to capture long-range context. The study highlights that generative pre-training combined with discriminative fine-tuning is a powerful method for advancing natural language understanding, setting the foundation for modern pre-trained language models.
Dan Ahimbisibwe 10/09/2025
Schiff, D. (2025). "Education for AI, not AI for Education: The Role of Education and Ethics in National AI Policy Strategies."
The paper analyzes 24 national AI policy strategies to explore the role education plays in policymaking, stating that the strategies focus more on "education for AI" rather than "AI for education" (AIED). This means that the strategies mainly emphasize on preparing a workforce skilled in AI technologies through training AI experts, reskilling workers, and promoting public AI literacy, but lack discussions about the ethical and social implications of AI applied within educational settings. The author argues for bridging the current policy gap by encouraging policymakers to integrate ethical considerations explicitly into the design, deployment, and governance of AI in education. This article pushes for a dual approach of investing in human capital for AI and harnessing AI responsively to improve education itself.
Nasrin Alanna Aidi 10/23/2025
Ongoing Projects
Current research initiatives by the AI in Education VIP team.
Domain-Specific AI Detection at NYU
We are building a domain-specific AI detector trained on Congressional Research Service reports.
Most existing detectors are "all-purpose" and fail under adversarial conditions. By focusing on a narrow domain (student essays, starting with Congressional Research Service reports), we aim to develop a detector that is more accurate and interpretable than general tools like GPTZero/Turnitin.
And perhaps the most exciting part: we won't just train on better data but we'll also explore architectural improvements to the transformer itself, pushing beyond standard models to capture structural differences between human and AI writing that current detectors miss.
NYU Writing Center Platform
This project continues the development of an online writing platform that bridges first-year and upper-level writing by embedding expert strategies from EXPOS courses into an accessible platform.
Designed as a "reinforcement point," it helps students carry forward effective writing habits beyond first-year writing. Unlike traditional word processors, it both (a) promotes the kind of meta-reflection that motivates radical revision and (b) provides tools for enacting those revisions.
In this way, the program demystifies the habits of expert writers, supports lasting skill development, and reduces the attractiveness of AI shortcuts.
Our Team
Our VIP team is composed of dedicated students and faculty contributing to AI in Education research.