See also Pedagogy, LLM writing styles, LLMs in scientific research.
There seem to be three lines of thinking in education over LLMs—three latent factors that are combined in different ways to form different perspectives:
There are, consequently, many proposals for what we should do in response. My instinct is that LLMs enhance the case for the same pedagogical techniques we should have been doing for years: active learning, peer instruction, realistic projects, and so on. These are what encourage true understanding in students, and can be arranged to limit the student ability to outsource their thinking to LLMs; and once we have students who can master the basics, they can learn to use LLMs to their advantage.
But of course this is an empirical question, addressed by the papers below.
Imundo, M. N., Watanabe, M., Potter, A. H., Gong, J., Arner, T., & McNamara, D. S. (2024). Expert thinking with generative chatbots. Journal of Applied Research in Memory and Cognition, 13(4), 465–484. doi:10.1037/mac0000199
A good overview of how experts think (in cognitive psychology terms), how experts could use LLMs to support them, and the considerations for novice use. Discusses how offloading low-level tasks to a tool can threaten expertise, since (a) experts become experts through deliberate practice of those low-level skills; (b) expertise involves examining the whole problem, including the low-level details; and (c) fields change over time, so engagement with the details is necessary to stay up-to-date. Gives theoretical arguments that automating tasks typically done by junior staff (e.g., new lawyers tediously reviewing contracts) may prevent them from becoming experts.
Yan, L., Greiff, S., Lodge, J. M., & Gašević, D. (2025). Distinguishing performance gains from learning when using generative AI. Nature Reviews Psychology, 4, 435–436. doi:10.1038/s44159-025-00467-5
Opinion piece noting that empirical studies measuring performance – how well students complete some task – do not necessarily measure learning, so research into the effect of AI on education should be careful about what they’re measuring. There’s every reason to think students can perform well even without much learning (e.g., by cramming for an exam and forgetting it all a day later). Points to some of the research below to note that AI may boost performance while diminishing learning by inhibiting metacognition and reducing engagement with tasks.
Thorgeirsson, S., Ewen, T., & Su, Z. (2025). What can computer science educators learn from the failures of top-down pedagogy? In Proceedings of the 56th ACM Technical Symposium on Computer Science Education (Vol. 1, pp. 1127–1133). doi:10.1145/3641554.3701873
Creates a dichotomy of teaching approaches: bottom-up approaches start with the basic core skills and work up to integrate them, while top-down approaches start with the overall goal and work downward. Phonics, in which children learn to read by practicing sounding out individual words and then progressing to whole sentences, is bottom-up; whole-language reading, where students start reading entire sentences early and guess how to read unfamiliar words from context, is top-down. Whole-language reading was trendy until recently, but it turns out that all the empirical evidence suggests it doesn’t work. The authors worry that “approaches that rely heavily on large language models (LLMs) or ‘prompt programming’ run the risk of being the computer science equivalent of whole language, focusing heavily on end results without understanding the underlying mechanics”. (While the paper is about teaching CS, the argument should apply to almost any field.) They review several studies, including some cited below, that top-down LLM-based approaches may leave students with weak metacognitive skills. But, they note, the other part of education is engagement and motivation—whole-language reading is much more fun for teachers, because they get to read stories instead of sounding out words; and LLM-based education might be more fun because of the fancy new technology. And you can’t pick a pedagogy without considering whether it will motivate students and teachers.
Bernstein, S., Rahman, A., Sharifi, N., Terbish, A., & MacNeil, S. (2025). Beyond the benefits: A systematic review of the harms and consequences of generative AI in computing education. In Proceedings of the 25th Koli Calling International Conference on Computing Education Research. ACM. doi:10.1145/3769994.3770036
Systematic review of CS education papers on the harms of AI, covering the obvious categories of harms (cognition and metacognition), but also social harms. equity, integrity, and so on. Good source for papers on each topic. Social harms seem under-studied (students no longer working together or asking each other, or TAs, for help) but quite relevant to motivation, belonging, and soft skills development.
Lovett, N. (2026). Individual differences in generative AI effectiveness: An empirical puzzle for HRD research. New Horizons in Adult Education and Human Resource Development. doi:10.1177/19394225261472792
Reviews the somewhat contradictory literature on AI assistance: some professionals become more productive, some become less, some students gain, some lose. Why should it be so varied? Argues this must depend on individual traits, but also on path dependence: your traits change how you choose to use AI and the role you let it play, changing how its use affects you. Argues for further research to study this, so we can reap the benefits but not suffer the deskilling or lack of growth that AI can cause.
Bastani, H., Bastani, O., & Sungu, A. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. doi:10.1073/pnas.2422633122
Randomized experiment of Turkish high school math students. Three treatment arms: control (normal teaching), GPT Base (students get unrestricted GPT-4 access), and GPT Tutor (GPT-4 is prompted to act like a tutor instead of giving answers, and has instructor-written instructions on how to explain things). Found, perhaps unsurprisingly, that students do well on practice problems while they have AI access, but those in the GPT Base condition did worse than the control group when AI access was taken away. The GPT Tutor group didn’t do worse, but they also didn’t do better than the control group. So letting AI write answers for you harms your learning in the ways you’d expect.
Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. doi:10.1038/s41598-025-97652-6
An RCT in Harvard’s intro physics course, with outcomes measured with the Force Concept Inventory. (The FCI is widely used in physics, and Harvard is where Peer Instruction, the most effective early active learning intervention in intro courses, was developed. See Pedagogy.) Crossover study with two treatments: an in-class active learning activity and an at-home AI-tutored lesson, using GPT-4 as the tutor via TeachGPT. The AI tutoring showed double the learning gains compared to the in-class active learning activity (and in fact hit the ceiling that could be measured with the FCI). It would be nice to have the actual lesson materials available, and also to know how the pre-test and post-test were conducted: were students in the at-home AI lesson using ChatGPT to do the post-test for them?
Salinas, A. et al. (2026). Law professors prefer AI over peer answers. SSRN. https://ssrn.com/abstract=6849678
Had 16 law professors who teach contract law write questions that students might ask during office hours, then had the professors and Gemini 2.5 Pro write answers to the questions. In blind matchups, the professors were asked to judge which was a better answer to the question. “Professors rated LLMs far higher than their peers (average win rate = 75.33%),” and the effect was consistent across law professors: they all rated Gemini better than humans, and only one professor’s answers outperformed Gemini. They say they measure “structural organization, reasoning nuance, legal anchors, confidence tone, clarity, pedagogical support, and length” to test if these lexical/syntactic features account for the difference, and claim “the LLM advantage cannot be fully explained by common textual features”, though I don’t see any explanation of how these features were defined or computed.
In a follow-up, they tested if Llama 4 can be used as a judge instead of humans, and claim it was consistent with the human raters; this allowed them to scale up to test a bunch more LLMs, showing that Claude Opus 4.7 is even stronger than Gemini 2.5 Pro was.
One wonders how the law professors were judging the answers, and whether their preferences are aligned with what is most instructionally effective for students asking questions.
LearnLM Team, Google, & Fab AI (2026). Teaching with Gemini: Measuring the impact of Guided Learning on student mathematics progress in Sierra Leone. Google. https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf
Preregistered RCT conducted in Sierra Leone. 7th and 8th grade math classrooms were randomized to either use Gemini’s Guided Learning feature or standard instruction. Collected pre- and post-tests of math learning. Showed a 0.26 SD gain in learning for the students using Guided Learning, relative to the students getting standard teaching.
It’s not clear here what exactly Guided Learning does; Google just says it “breaks down problems step-by-step and adapts explanations to your needs”. Also not clear what the instructional design was. The Appendix just says the teachers came up with learning objectives and an activity, then gave students starting prompts and scaffolding for using Gemini. So there’s something here, but it’s not clear how it compares to, say, training the teachers to a different kind of class activity, or other changes relative to the current baseline of instruction.
LearnLM Team, E., Google (2026). Measuring the impact of AI tutoring, from near transfer to standardised tests: An expanded RCT in UK mathematics classrooms. Google. https://services.google.com/fh/files/misc/learnlm_oct26.pdf
Another preregistered RCT from Google, this time in “10 UK state secondary schools over 10 weeks of classwork”. Students learning math through an online platform were randomized to four conditions: “standard remediation on the Eedi platform, including curated videos and expert-written hints to resolve student misconceptions; expert human tutoring; AI tutoring, guided by a specialised pedagogical prompt; and AI tutoring, guided by the pedagogical prompt and personalised context about the student’s learning history”. AI tutoring was provided by various versions of Gemini (updated throughout the trial). All forms of tutoring were more effective than the standard remediation, at least when measuring student performance on follow-up questions and related questions. However, ultimately the students with AI tutoring performed very similarly on the standardized math exam. They speculate this is because (a) students getting tutoring took longer to progress through the material, and (b) the effects of tutoring seemed to be stronger for girls than boys. They conclude Gemini is just as effective as human tutors, though I would also be interested in long-term measures of student attitudes and motivations (don’t they get bored interacting with an AI instead of a human?).
Bo, J. Y., Kazemitabaar, M., Deng, M., Inzlicht, M., & Anderson, A. (2026). Invisible saboteurs: Sycophantic LLMs mislead novices in problem-solving tasks. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3772318.3791365
Clever randomized crossover study with novice ML undergrads. Each completed two tasks with LLM assistance, but one LLM was tuned to be “High Sycophancy” and the other was “Low Sycophancy”. (This wasn’t just a prompt; they used a separate LLM to identify misconceptions in the student’s answer, then passed these to the answering LLM. The high sycophancy answering LLM receives these misconceptions and is told to be more sycophantic.) Students used the LLMs while doing basic ML prediction tasks; those with the Low Sycophancy LLM did much better than those with the High Sycophancy LLM. They tried to explore why, and concluded that “the HighSycophancy chatbot induced overwhelming inappropriate over-reliance through validating users’ misconceptions”. But the “users were largely unobservant of the effect of sycophancy on the LLM’s response characteristics”; some attributed the difference between tasks to their own prompts, and did not realize one LLM was hindering them. Some liked the High Sycophancy LLM because it was “encouraging and friendly”. Seems to fit with prior misconception research: students don’t realize they have misconceptions, even when they’re struggling because of them, and having those misconceptions confronted is uncomfortable.
Yu, S., Cheng, M., Jabbar, A., Sucholutsky, I., Collins, K. M., Jurafsky, D., & Hawkins, R. D. (2026). The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks. arXiv. https://arxiv.org/abs/2605.22687
Three studies using online participants to do toy tasks, randomized to use or not use LLMs to support. Finds that users are not good judges of whether AI saves them time, and adopt it even when it doesn’t:
We identify systematic miscalibration at two levels: (1) a self-estimate miscalibration where people on average believe that they are using AI less than they actually are, and (2) efficiency-gain illusions where people overestimate how much time and effort savings AI use affords. We also identify a session-level carryover effect where a participant’s prior AI use leads to further AI adoption and entrenches their miscalibration about time savings.
While I’d like to see this tested in a more realistic setting (real tasks done by people with a stake in their result), I can certainly believe the message that people are poor judges of time use and mistake easy for fast.
Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026). AI assistance reduces persistence and hurts independent performance. arXiv. https://arxiv.org/abs/2604.04721
Another study of online participants doing toy tasks, this time fraction problems or reading comprehension tests. Participants are randomly assigned to do the tasks with AI help (GPT-5) or without; after doing tasks for a while, the AI was removed, and their performance on the subsequent tasks was measured. In each experiment, “AI assistance improved performance during the learning phase, but solve rates dropped and skip rates increased once the AI was removed”. That is, people with AI help performed worse than the control group when it was removed, and also were more likely to press the button to skip the task, suggesting the AI “reduced persistence” by making them more likely to give up on the task without its help. They suggest this is evidence for “gradual disempowerment”, where reliance on AI erodes human skills and persistence, “displacing human participation over time”. They speculate on the metacognitive mechanisms for this displacement. While the paper is not explicitly about teaching and learning, I’d expect the mechanisms to apply just as well to students using AI to help them do schoolwork.
Autor, D., Rodchenko, T., Martin, J., Iscenko, Z., Strand, S., Pearl, D., & Ferere, M. (2026). Does AI assistance enhance or erode expertise? Evidence from a three-month field experiment in patent drafting. NBER Working Paper. doi:10.3386/w35720
Randomized trial of patent attorneys using “a custom AI drafting assistant” (made by Google, which helped run the study). Lawyers using AI performed better on patent drafting tasks. But after three months of AI use, when they were forced to do a task without AI, the results changed. Scores were higher on average among AI users,
But these learning gains were concentrated entirely among senior lawyers; junior lawyers, the group that captured the largest contemporaneous gains in drafting tasks with AI access, showed no comparable carryover. As in AI-assisted drafting, the higher redlining performance of treated senior lawyers stemmed from fewer mediocre and poor scores rather than from any gain at the top of the distribution. Conversely, treated juniors showed no average gain in performance and substantially greater dispersion: sharply fewer mediocre (fourth-quintile) scores, significantly more poor (bottom-quintile) scores, and an offsetting rise in good (second-quintile) scores. This dispersion is consistent with AI accelerating skill acquisition among some juniors and substituting for it among others.
Perhaps there is a difference in how these attorneys used the AI – replacing their judgment or supplementing it – but they don’t have any data on usage to tell.
Handa, K. et al. (2025). Anthropic Education Report: How university students use Claude. https://www.anthropic.com/news/anthropic-education-report-how-university-students-use-claude
Reviews half a million Claude conversations from accounts with university emails. CS and STEM topics were disproportionately overrepresented, as you might expect if those students are early adopters. Finds that “nearly half (~47%) of student-AI conversations were Direct—that is, seeking answers or content with minimal engagement”. This could be for legitimate use (checking your work on a practice exam, say) or just to get answers to homework problems. In Bloom’s taxonomy, Claude was often asked to complete the higher-level functions, not the lower ones, raising “legitimate worries that AI systems may provide a crutch for students, stifling the development of foundational skills needed to support higher-order thinking.” Of course, there were also significant proportions of conversations that are less concerning, like asking for explanations of concepts or for study materials.
Bent, D., Handa, K., Durmus, E., Tamkin, A., McCain, M., Ritchie, S., Donegan, R., Martinez, J., & Jones, J. (2025). Anthropic Education Report: How educators use Claude. https://www.anthropic.com/news/anthropic-education-report-how-educators-use-claude
Reviews 74,000 anonymized Claude conversations from “higher education professionals”, as well as surveying Northeastern University faculty on their use. Finds that faculty tend to use AI to augment their work, not replace it—except for boring administrative tasks, which they happily outsource. Some educators use AI to build educational games, assessment tools, demos, and other cool educational products. They did observe some faculty using AI directly for grading, though their survey respondents were very uneasy with this.
Specifically, how LLMs affect student learning of writing. For effect on writing style, see LLM writing styles.
Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. doi:10.1111/bjet.13544
An experimental study of ESL students completing a writing task in several experimental conditions, such as with ChatGPT support, with a human tutor, or with a writing checklist. The task focused on revision, so the students worked to revise their writing to meet a rubric. Their actions (edits, clicks, chat interactions, etc.) were collected. The AI-assisted students produced better writing, but possibly because they were copying its generated text to cater to the scoring rubric. The students using AI spent much less time evaluating their writing than those using a checklist (which forced them to evaluate each item), making the authors worry that using ChatGPT might “optimise performance at the expense of developing genuine human skills”. They call this “metacognitive laziness”: the students outsource the metacognitive skills (of evaluating their work and determining how to improve it) to a tool, and show no sign of learning those skills.
Mahapatra, S. (2024). Impact of ChatGPT on ESL students’ academic writing skills: A mixed methods intervention study. Smart Learning Environments, 11. doi:10.1186/s40561-024-00295-9
An experimental study using ChatGPT to help ESL students learn writing. Students were trained to use it to get feedback on the content, organization, and style. Results showed higher scores (on a writing rubric) for the ChatGPT group in post-tests. Interviews with students showed they found the feedback helpful. However, the writing tasks were paragraphs of 150 words and the paper is vague about whether the post-tests were conducted with or without ChatGPT access, so it’s unclear if this tests if the students learned how to write without ChatGPT’s aid or if the graders just prefer ChatGPT-mediated output; and whether this generalizes to longer writing forms.
Oppenheimer, D. M., Cash, T. N., & Connell Pensky, A. E. (2025). You’ve got AI friend in me: LLMs as collaborative learning partners. doi:10.31219/osf.io/8q67u_v2
A similar experimental intervention in a large (154 students) intro course. Students used ChatGPT to help improve their argumentative essays, turned in their drafts and chat transcripts, and provided self-reflections on the revision process. They wrote five essays throughout the course, and performance on the final essay was evaluated before they got LLM feedback, to test if they had improved. Found significant improvements in writing. But there was no control group, so it’s unclear how ChatGPT compares to their progress without AI assistance.
Taylor, M. R. (2026). The obligation to restrict AI in student writing. Journal of Philosophy of Education. doi:10.1093/jopedu/qhaf097
Exactly what it says on the tin. If the purpose of writing in education is to expose the student’s thinking and growth to the instructor for feedback and assessment, allowing AI use is morally equivalent to allowing plagiarism. Both impede the assessment role and render the feedback useless, because it does not help develop the student. “But they’ll be able to use AI on the job”, you might object; but a rich student hiring someone to do their coursework could just as well hire someone to do their job, and you still wouldn’t let them plagiarize, would you? Besides, “What matters, in evaluating such restrictions, is whether the technology prevents the internalizing of foundational knowledge, not whether students will one day have access in the workforce. Appeals to workforce productivity are therefore insufficient for marking a moral distinction between AI appropriation and plagiarism.”
There is much more to the essay; it covers all the usual counterarguments and shows that to accept AI in coursework is to imply that “productive results are all that matter” and “the education of the person behind those results is irrelevant.” That is, if you believe AI renders personal growth and skills morally valueless, then let students use AI all you want; but if you still value human development, you have an obligation not to let them use AI to write.
Savelka, J., Agarwal, A., An, M., Bogart, C., & Sakr, M. (2023). Thrilled by your progress! Large language models (GPT-4) no longer struggle to pass assessments in higher education programming courses. In Proceedings of the 2023 ACM Conference on International Computing Education Research (Vol. 1, pp. 78–92). doi:10.1145/3568813.3600142
Evaluates GPT-4 (not 4o or any of the later, more coding-focused models) on undergrad Python course exercises. (These are online courses for certifications, not advanced CS courses.) Finds GPT-4 can likely earn a passing grade. I suspect this would be true of newer LLMs and our own undergraduate statistical computing course.
McDanel, B., & Novak, E. (2025). Designing LLM-resistant programming assignments: Insights and strategies for CS educators. In Proceedings of the 56th ACM Technical Symposium on Computer Science Education (Vol. 1, pp. 756–762). doi:10.1145/3641554.3701872
Tests GPT-4o and Claude 3.5 Sonnet on SIGCSE’s Nifty Problems, which are basically papers presenting clever CS homework assignments and projects. Claude does particularly well, succeeding at most assignments, though a few assignments were consistently hard for the LLMs. Figure 4 summarizes what LLMs were bad at: anything involving visual output (despite being multimodal, LLMs are not very good at interpreting images), tasks requiring sequences of many steps, and, oddly, very detailed and specific assignment prompts: “Assignments that limit the solution space by giving detailed, clear, and explicit instructions are challenging for LLMs. Open-ended and greenfield projects are easier.”
Prather, J., Reeves, B. N., Denny, P., Becker, B. A., Leinonen, J., Luxton-Reilly, A., Powell, G., Finnie-Ansley, J., & Santos, E. A. (2023). “It’s weird that it knows what I want”: Usability and interactions with Copilot for novice programmers. ACM Transactions on Computer-Human Interaction, 31(1), 1–31. doi:10.1145/3617367
Prather, J., Reeves, B. N., Leinonen, J., MacNeil, S., Randrianasolo, A. S., Becker, B. A., Kimmel, B., Wright, J., & Briggs, B. (2024). The widening gap: The benefits and harms of generative AI for novice programmers. In Proceedings of the 2024 ACM Conference on International Computing Education Research (Vol. 1, pp. 469–486). doi:10.1145/3632620.3671116
Fascinating pair of think-aloud interview studies with novice programmers using GitHub Copilot and ChatGPT to do a simple programming task. In the first, they observe students “shepherding” Copilot toward solutions, but sometimes “drifting” when Copilot generated “large blocks of unhelpful code” that they had to read and ultimately reject (or accept and then debug).
In the second, half the subjects, who knew what they wanted and had a rough idea of strategy, did quite well. Half, however, struggled greatly: they were derailed by constant interruptions from Copilot offering suggestions, sidetracked by suggestions that actually solved a different problem, got stuck in misunderstandings of the task and couldn’t get un-stuck, and so on. Most ultimately finished the task, but slowly. They conclude: “From the evidence presented above, it appears that most of these ten who struggled thought they understood more than they actually did. The patterns of behavior above describe how participants were often led along by GenAI such that each step was able to be rationalized as understanding, making it even more difficult for participants to assess their own learning.” They caution that studies saying novice programmers self-report finding AI useful may be misleading, because their own interviews show students being derailed by it and then claiming it was helpful, perhaps because the students don’t have the metacognitive skills to recognize the result was wrong.
O’Brien, G. (2025). How scientists use large language models to program. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3706598.3713668
Survey and interview study of researchers at the University of Michigan who use LLMs (either chatbots or Copilot-style editor plugins) to write code for their research. Finds that a common use case is working with code or libraries the researcher doesn’t understand, such as when they work in a different language they’re unfamiliar with or work with code written by someone else. Most of the researchers don’t have a formal way to check if the generated code is correct, mainly just running it and seeing if the output looks right; during one interview, as the subject showed a chat log of their apparently successful use of ChatGPT, the interviewer noticed a bug they had missed through lack of testing and verification. The paper gives several other examples of missed errors in LLM code suggestions.
Several of the researchers also expressed misconceptions about ChatGPT, such as believing that it Googles the answer to their questions (at a time when ChatGPT did not have web access) or thinking it performs mathematical and logical calculations.
The paper doesn’t make the connection to critical thinking and metacognition, but the results suggest even expert users are not critical enough of AI output, and do not have effective strategies for testing and verifying their code.
Lucchetti, F., Wu, Z., Guha, A., Feldman, M. Q., & Anderson, C. J. (2025). Substance beats style: Why beginning students fail to code with LLMs. In L. Chiruzzo, A. Ritter, & L. Wang (Eds.), Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Vol. 1, pp. 8541–8610). Association for Computational Linguistics. doi:10.18653/v1/2025.naacl-long.433
Reviews a dataset of student interactions with LLMs to solve programming problems (“write a function to…”). Goal is to distinguish between two explanations for students failing to get good results: (a) students don’t have the technical vocabulary to state what they want, or (b) they don’t provide the LLM sufficient information to solve the problem. Finds support for (b) through various clever analyses of the prompt logs, such as by swapping out the nonstandard vocabulary students used for the technically correct terms. That didn’t help much, but they found that unsuccessful prompts were typically missing some of the clues necessary to solve the problem. In other words, students didn’t realize what knowledge is needed to do the task, so teaching them prompt engineering that focuses on wording will not be sufficient alone.
Thorgeirsson, S., Weidmann, T. B., & Su, Z. (2026). Computer science achievement and writing skills predict vibe coding proficiency. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3772318.3791666
Recruits 100 CS students to complete programming tasks with vibe-coding. Finds that “that both writing skill and CS achievement are significant predictors of vibe-coding performance, and that CS achievement remains a significant predictor after controlling for domain-general cognitive skills”. That is, being able to write well helps you vibe-code, and having domain knowledge does too (perhaps for the reasons noted by Lucchetti et al above, though they do not test this). They add that “prior LLM usage is negatively and significantly correlated with written communication and vibe coding performance. We speculate that this relationship is caused either by (i) LLMs having a negative effect on students’ ability to express themselves or (ii) that students who are less proficient at writing are more likely to use LLMs, or (iii) a combination of both.”
Ma, Q., Peng, W., Yang, C., Shen, H., Koedinger, K. R., & Wu, T. (2025). What should we engineer in prompts? Training humans in requirement-driven LLM use. ACM Transactions on Computer-Human Interaction, 32(4). doi:10.1145/3731756
Explores prompt engineering for building applications with AI. Finds that conventional prompt engineering does not meaningfully help humans succeed in building AI-based tools, but “Requirement-Oriented Prompt Engineering”, which teaches humans to define a clear set of requirements and encode those in the prompts, performs much better. This seems to fit well with the above studies showing knowledge and writing skills are necessary to use AI, since prior knowledge is necessary to know what requirements to state; perhaps, then, prompting is a teachable skill conditional on having the necessary background knowledge.
Ma, Q., Koedinger, K. R., & Wu, T. (2026). Not everyone wins with LLMs: Behavioral patterns and pedagogical implications for AI literacy in programmatic data science. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3772318.3791283
Follows students in a spring 2025 course as they use Google Colab and its built-in Gemini to do data science programming tasks. A smallish course (36 students), so not super definitive, and this is observational rather than an experiment. Found that students with more prior technical experience did better – the AI didn’t eliminate the gap – and students often had poor mental models of the AI. That is, they didn’t know what it was capable of and what information it had access to. At the time, Colab’s Gemini didn’t have access to the environment and loaded data, for example, so it often used the wrong column names in code; students didn’t realize this (despite an in-class demo explaining how to use AI!) and thus kept getting bad results. Also, “Asking AI to explain code, bugs, or critique outputs remained limited, despite its potential benefit to improve understanding. We especially observe that students may not persist in reading convoluted output, reflecting reluctance or difficulty in engaging with evaluative or complex information.” That is, students don’t want to press the “Explain error” button because they don’t understand (or don’t want to read) a big long explanation. Students may need clearer instruction and practice using AI to use it effectively, and novices may not have the skills to benefit from it.
Shen, J. H., & Tamkin, A. (2026). How AI impacts skill formation. In Post-AGI Science and Society Workshop, ICLR. https://arxiv.org/abs/2601.20245
Study by Anthropic (!) in which junior software engineers complete tasks using an unfamiliar Python library. Participants were randomized to either write the code solo or to have access to a chat-based AI assistant (GPT-4o). After completing the tasks, participants took a quiz “with debugging, code reading, and conceptual questions” about core concepts used in the Python library. Participants who used AI did 17% worse on the quiz and took a similar amount of time to complete the programming tasks. Reviewing screen recordings for the participants, they found several patterns. Participants with low quiz scores tended to delegate everything to the AI or relied heavily on it to debug and verify their code. Participants with high scores asked conceptual questions, asked for explanations of code, or otherwise followed up to understand how the code worked. AI users weren’t any faster because they “spent significant time (up to 11 minutes) interacting with the AI Assistant”. They conclude that “our results suggest that the aggressive incorporation of AI into the workplace can have negative impacts on the professional development workers if they do not remain cognitatively engaged.”
Al Labadi, L., & Ly, A. (2025). Enhancing statistics education through Project‐Based Learning (PBL) and the emergence of ChatGPT. Teaching Statistics, 47(3), 200–218. doi:10.1111/test.12405
Discussion of two project-based statistics courses at the University of Toronto Mississauga. In the first course, they noticed several groups using AI “to help write the introduction, discussion, limitations, and conclusion”, so they switched to oral exams in the second course, interrogating each group about their project and their decision-making. 43% of students admitted to using AI in an anonymous survey, despite it being officially forbidden. There’s a lot of other discussion here about student impressions of the projects and the logistical details of project-based learning, but the generative AI parts are interesting; ultimately they say they should stop forbidding AI and adopt strategies like requiring draft submissions and revisions to ensure students are supervising AI and learning from it instead of just copying its output. They also intend to introduce more coding instruction, because students “tend to rely on generative AI codes, which occasionally provide incorrect suggestions or provide content that is inconsistent with the course material.”
Lu, Y. et al. (2025). StatEval: A comprehensive benchmark for large language models in statistics. arXiv. https://arxiv.org/abs/2510.09517
Extracts 14,000 textbook stats problems, from undergrad to doctoral level, and 2,400 “proof-based questions” from peer-reviewed journals (using an LLM-based automated pipeline). Then scores LLMs on solving the problems. On textbook problems, GPT-5 hits 83%, Gemini 2.5 Pro gets 81%, Wen3-235B gets 77%, and other models do worse. (Claude was not tested.) On the research problems, all models score around 50%. “Our systematic evaluation reveals that even the strongest closed-source models struggle with research-level tasks, particularly in advanced machine learning theory.” However, I’m skeptical of LLM-compiled benchmark sets with only light human review; I wonder how good these problems are.
Their website offers a leaderboard with more recent models; for example, GPT-5.4 is up to 60% on the research tasks: https://statai-lab.github.io/StatEval.github.io/
Glickman, M., Yan, J., Eric, N., & Çetinkaya-Rundel, M. (2026). A tutor, a tool, and a sounding board: How statistics students use generative AI. CHANCE, 39(2), 39–44. doi:10.1080/09332480.2026.2698398
Summary of a 2025 American Statistical Association survey of statistics students on their use of AI. Got only 53 responses, across bachelor’s, master’s, and PhD programs. “Roughly 70 percent of respondents reported using AI tools in their coursework”, mainly ChatGPT, and they describe some of the use cases; students report faculty have widely varied AI policies and recommendations. Students also report using AI for lots of outside-class things, like improving cover letters, generating recipes, supporting hobbies, etc. They conclude that “Our findings suggest that students were engaging with generative AI in nuanced, deliberate, and evolving ways rather than having adopted it uncritically”, though I don’t see the evidence for this in the responses: the students are selective in their use cases, but that doesn’t mean they’re thinking particularly critically about the use of AI for particular use cases.
Besides how students think and learn using AI, it’s important to consider broader questions about how LLMs and AI will affect society and work more generally. Also, automation augmenting human skills is not a new problem, so there are useful historical parallels.
[To read] Leslie Paul Thiele (2025), Human Agency, Artificial Intelligence, and the Attention Economy: The Case for Digital Distancing. Springer. doi:10.1007/978-3-031-82086-1
I’ve only read the chapter on deskilling, which argues that historical parallels (e.g., the shift from artisan work to factory labor) suggest that AI will not simply augment our skills, making us more powerful; it will replace our skills, making us dependent much like the people in Wall-E.
Simkute, A., Tankelevitch, L., Kewenig, V., Scott, A. E., Sellen, A., & Rintel, S. (2025). Ironies of generative AI: Understanding and mitigating productivity loss in human-AI interaction. International Journal of Human–Computer Interaction, 41(5), 2898–2919. doi:10.1080/10447318.2024.2405782
A paper by human-computer interaction researchers who note that some of the problems in the studies above (that using Copilot to help write code actually creates new problems, like dealing with its interruptions and incorrect suggestions) are problems that have been well-known for years in human-computer interaction. Autopilots, industrial automation, and many other systems have run into similar problems. Reviews a number of common themes from past automation research:
They give examples of how each of these has happened in past automation and in generative AI research, then discuss how the field of HCI offers approaches to better design generative AI tools to avoid these issues. Basically, “AI designers should remember that our field exists instead of repeating every mistake from scratch.”
Social interactions
Students who have AI to help them can avoid asking their peers or instructors for help. Perhaps the AI gives excellent answers; perhaps it does not, or helps them too much. But I am also interested in the social effects. Office hours are how students learn from instructors how to think like an expert; interacting with peers and instructors is how students grow into the culture of the discipline and develop a sense of belonging. Are those social benefits going away?
Hou, I., Man, O., Hamilton, K., Muthusekaran, S., Johnykutty, J., Zadeh, L., & MacNeil, S. (2025). "All roads lead to ChatGPT": How generative AI is eroding social interactions and student learning communities. In Proceedings of the 2025 Innovation and Technology in Computer Science Education Conference (Vol. 1, pp. 79–85). doi:10.1145/3724363.3729024. https://arxiv.org/abs/2504.09779
Through interviews with CS undergrads, notes a tendency for students to seek help from ChatGPT rather than peers, which some interviewees suggest reduces their social interactions, makes them feel more isolated, and reduces opportunities for peers to be mentors or share experiences, ultimately reducing their sense of belonging in the field.
Ko, S.-H., & Stephens-Martinez, K. (2026). Connecting computing students’ external help resource preferences and internal help resource usage: 2021-2025. In Proceedings of the 57th ACM Technical Symposium on Computer Science Education (Vol. 1, pp. 596–602). doi:10.1145/3770762.3772519
Uses data from Duke CS courses that use a queuing system for office hours, so they have data from 2021–2025 on office hour use. They also have data on the course discussion forums and, since 2023, student surveys on help-seeking and resource use. Office hour use has declined since 2023, in some courses by about half, though trends for the course discussion forums are less clear. The survey results are not consistent between courses, and do not show a huge jump in preference for AI help. But I would tend to believe the revealed preference (in behavior, in the decline in office hour use) over the stated one, and the decline in office hour use is concerning.
Gao, Z., Zahn, M., Heckman, S., & Lynch, C. (2026). From Covid to GenAI: Tracing long-term changes of students’ help-seeking behaviors in a CS2 course. In Proceedings of the 2nd ACM Virtual Global Computing Education Conference (Vol. 1, pp. 136–142). doi:10.1145/3795867.3830998
Similar study in a CS course at NC State, tracking office hour use from 2020 to 2025. Students used an online system to request office hour slots. Use declined from about 10 requests per student pre-AI to about 4 requests per student in Fall 2025; activity in the course online help forum (Piazza) also dropped from 2020 (when the course was online). They give several hypotheses for these trends, including increasing AI use, the COVID disruption in social behaviors, and the rise of short-form media making students want instant gratification; they “believe the current situation is the result of all these factors together and cannot be explained by any single one”.