CMSC724: Reading, Research, and Writing With AI Available

Instructor: Amol Deshpande

Important

A note to graduate students. Please read before the class begins.

The purpose of this note is primarily to convince you that you need to be very mindful about how you use AI in the course. I made a decision a while ago to allow effectively unlimited AI use in my classes (senior undergraduate and graduate), both because I see how useful AI tools are, and also because you will be expected to use these tools moving forward.

There is a version of this memo for undergraduates that argues from exam scores: students who lean on AI during practice (for homeworks, programming projects) do worse on closed-book tests (some specific studies below), and larger emphasis on in-class exams exposes that gap.

But that argument doesn’t translate to a graduate course like CMSC724 (where in-class exams don’t make sense, although I have seen some folks reintroduce them), and in some ways, the impact of AI use is likely worse considering the intended learning outcomes for this course.

The gaps in being able to critically read research papers, think independently, write coherently, will not show up for several years, and possibly not at all during your MS or PhD. But in three to five years, your professional value may consist almost entirely of being the person who can tell whether a confident technical claim is actually correct — in a review, in a design meeting, in a system that is about to be trusted with something.

We get better at those things because of the struggles in doing those (what has been called “desirable difficulty”), but AI tools largely remove that struggle, so you need to try to force yourself through it. I have tried to set up the deliverables to achieve that goal, but ultimately its on you to recognize the dangers and use the tools responsibly.

Important note: I have intentionally omitted any discussion of how to use them for “coding” or “architecting”, which is a whole another topic, but you can read this paper (from UMD :) as a starting point: (Im)paired Programming: Coding Agents Improve Productivity but Harm Understanding

The Baseline Evidence

A growing body of experimental and observational evidence suggests that answer-giving AI can improve assisted performance while impairing subsequent unaided performance. The studies are not yet sufficient to establish how general or durable these effects are, especially for graduate research and writing, but the consistency of the emerging pattern warrants caution.

In a randomized trial with ~1,000 students, unrestricted GPT-4 access during practice improved assisted performance 48% and then produced exam scores 17% below students who never had access (Bastani et al., PNAS, 2025). Randomized trials in 2026 (Liu et al., N = 1,222) reproduced the performance reversal after roughly 10-15 minutes of use, and found reduced persistence, especially among the participants who used AI to obtain direct answers. A 30-month panel of 26,811 students found homework scores up 18% while closed-book exam scores fell 20% within six months, with losses concentrated among the ~80% whose time-on-task collapsed and largest among high achievers (Strömberg, Lei & Wu, 2026).

The last result is worth noting, since it highlights that losses were especially large among high-achieving students.

Reading Papers and Relying on AI Summaries

A large portion of the coursework here is reading papers, and we usually ask you to write a summary with positive and negative points, etc. The goal is for you to struggle through the paper despite not knowing all the background or notation (and possibly mistakes in the paper), perhaps go back to portions of it you didn’t understand, think critically about the arguments, etc. If you outsource this to AI, you won’t learn how to do this effectively. Learning science calls this a desirable difficulty: the struggle is not unnecessary friction in the way of learning, it is necessary for you to learn. Another goal is to be able to distinguish a real contribution from a delta, an honest evaluation from a rigged one, a hard problem from one that merely looks hard. That taste is built by making your own judgments and being challenged for it. Consuming someone else’s judgments does not build it, no matter how good they are.

A key goal with reading a bunch of different papers is also to create a mental working map of a field, how different papers connect to each other, how they build upon each other, what the gaps are, etc. An AI-generated deep research document rarely helps build those mental maps, because they lose too much of the detail and you are not the one making those connections.

AI summaries of papers tend to be shallow and not sufficiently critical. Peters and Chin-Yee (2025, Royal Society Open Science) compared 4,900 LLM-generated summaries against their source texts across ten models. Several models produced conclusions broader than the source texts supported in 26–73% of cases, and LLM summaries had nearly five times the odds of overgeneralizing compared with human-written summaries. Explicitly prompting for accuracy increased overgeneralization, and several newer models performed worse than older ones.

Melumad and Yun (2025, PNAS Nexus) ran seven experiments with 10,462 participants comparing learning from LLM syntheses against learning from web sources. LLM users developed measurably shallower knowledge and then produced advice that was sparser, less original, and less likely to be adopted by readers. The effect held even when the AI summaries were augmented with real-time links to sources: the results suggest that the gap was not about which facts were available, but about who did the work of finding and assembling them.

There is also a randomized crossover study (Etkin et al., 2025, Frontiers in Education, n = 195) with a result that should worry you specifically. Reading an AI summary instead of the passage hurt comprehension most among the strongest readers — 76.5% down to 67.6% — while a Socratic discussion chatbot helped weaker readers. The summary tool was the one that hurt; the tool that asked questions was the one that helped. That distinction runs through a number of studies in this memo.

Again the suggestion here is not to use AI tools for this, but to be aware of these issues and to use them to complement your analysis of the paper rather than substitute for it. For example, I would highly recommend forming your own impressions of a paper by at least skimming it, before asking AI for its take, or better yet, asking it to challenge your take.

Doing Research

A large part of research is to identify gaps and contradictions in prior work, so we know what are some interesting open problems to work on. LLMs tend to smooth over those gaps in their high-level summaries unless you are careful and can ask the right questions (which itself requires a deep understanding of the problem space).

Sycophancy is another problem. If you ask an AI tool whether your idea is good, you will get a strong case for it. Cheng et al. (2026, Science) tested 11 frontier models and found they affirmed users’ actions 49% more often than human respondents (in the context of interpersonal advice). In three preregistered experiments (N = 2,405), a single interaction with a sycophantic model increased participants’ conviction that they were right — and participants rated the sycophantic responses as more trustworthy and said they would return to them. Discarding bad ideas is an important skill to learn, and a tool that may make every idea sound viable may not allow you to focus your energies. This is especially problematic for junior researchers like yourself.

Liu et al. (2026) paper talks about another really important trait of a researcher, namely persistence. They measured reduced willingness to stay with an unsolved problem after about ten minutes of assistance. A dissertation is three to five years of staying with unsolved problems, and persistence is not just a nice-to-have in your training, but critical. This of course applies to any long-term and challenging projects, not just dissertations.

Doshi and Hauser (2024, Science Advances) found that AI-assisted writers produced work rated more creative individually, while the resulting body of work was measurably more similar across authors. They describe it as a social dilemma: each person is better off, and the collective output narrows. Generalizing that to a research field — if everyone entering a subfield asks similar models what the open problems are, the field converges on the same problems — not ideal in the long term.

Writing Technical Prose

It is well-established that writing is critical to learning, and is the process by which thought gets completed. Composition researchers distinguish knowledge-telling from knowledge-transforming (Bereiter & Scardamalia, 1987): in real technical writing, the attempt to state a claim changes the claim. When you try to write “our approach is better because X,” you may discover that X doesn’t quite follow, that a condition is missing, that the two halves of the argument don’t meet in the middle. Sentence-level resistance, i.e., trouble connecting sentences logically, is a correctness checker for ideas.

LLM prose is exceptionally good at the surface features of logical connection without the connection itself, through fragments like “Furthermore”, “This naturally leads to”, “Building on this insight” — exactly what an argument should look like. However, often times, if you think critically about those statements, you might find that they don’t actually follow, but only seem to follow.

LLM prose is particularly hard to challenge if you don’t know the topic well (cf. Gell-Mann Amnesia effect), and for junior researchers working on cutting-edge topics, that is often likely to be the case.

“But I only draft with AI and then edit heavily”: Many people don’t see issues with this, including established researchers. I do it all the time (e.g., for writing this memo), but I am also actively trying to avoid it. Editing plausible text and generating precise text are quite different, with the second much harder. That needs practice, and drafting-then-editing removes it almost entirely.

Anecdotally, I clearly remember many instances in the past where I was stuck on just the opening line of the abstract or introduction, in some cases for 10–15 minutes or longer. For me, that opening line is crucial and getting it right was not just about that sentence but about setting tone for the entire section. I actively try to avoid using AI unless I have that opening now.

You also inherit the model’s framing and structure — its choice of what to highlight, what to concede, what order to make the argument in — regardless of how thoroughly you rewrite the words on top of it. The structure survives your revisions, and the structure itself is a major part of the argument.

On a more practical note, readers/reviewers tend to form judgments very quickly about whether the authors know what they are doing, and that judgment runs almost entirely on the precision of the framing. Generic prose now reads as an active signal of not knowing, in a way it did not three years ago. The cost of writing like everyone else has gone up sharply, and it will keep going up.

Summarizing for the Course

Ultimately you need to decide how you use AI tools in this course. I can make some recommendations, but I am not planning to police this. Following are what I think are some good ways to use AI, based on the research as well as my personal experience.

  • Asking the model to argue against a position or a plan you have already formed and/or written down.
  • Finding relevant references to a topic, perhaps with a one-sentence description for each reference.
  • Using it to understand specific background from a paper you are reading, or summarize a background paper you are not planning to read at all (be careful in guiding the model with specific questions — it may simply work towards confirming your take).
  • Any mechanical tasks (e.g., pointing at the course webpage and asking it download all papers, creating bibtexs, etc.).
  • Line-level copyediting (refining) after text is written (I usually tell it to “make as few changes as possible”).
  • Brainstorming if you are careful in not letting it guide your thinking.

Following are things to avoid.

  • Asking the model to summarize a paper you have been asked to read in the class.
  • Generating the positive/negative points for a paper.
  • Generating related work sections — the models are very good at this so it is hard to resist the temptation.
  • Drafting full text from outlines.

Sources

  • Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning. PNAS, 122(26). doi:10.1073/pnas.2422633122
  • Strömberg, D., Lei, V., & Wu, Y. (2026). The generative AI learning penalty: Evidence from Chinese secondary education. CEPR Discussion Paper 21577 / SSRN 6868618. (working paper, not yet peer reviewed)
  • Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026). AI assistance reduces persistence and hurts independent performance. arXiv:2604.04721 (preprint)
  • Peters, U., & Chin-Yee, B. (2025). Generalization bias in large language model summarization of scientific research. Royal Society Open Science, 12(4), 241776. doi:10.1098/rsos.241776
  • Melumad, S., & Yun, J. H. (2025). Experimental evidence of the effects of large language models versus web search on depth of learning. PNAS Nexus, 4(10), pgaf316. doi:10.1093/pnasnexus/pgaf316
  • Etkin, H. K., Etkin, K. J., Carter, R. J., & Rolle, C. E. (2025). Differential effects of GPT-based tools on comprehension of standardized passages. Frontiers in Education, 10. doi:10.3389/feduc.2025.1506752
  • Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792), eaec8352. doi:10.1126/science.aec8352
  • Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. doi:10.1126/sciadv.adn5290
  • Padmakumar, V., & He, H. (2024). Does writing with language models reduce content diversity? ICLR 2024. arXiv:2309.05196
  • Lee, H.-P., et al. (2025). The impact of generative AI on critical thinking. CHI ’25. doi:10.1145/3706598.3713778
  • Rismanchian, S., Uzun, H., Matayoshi, J., Cosyn, E., & Kurd-Misto, E. (2026). Faster completion, less learning. arXiv:2605.21629 (preprint)
  • Bereiter, C., & Scardamalia, M. (1987). The Psychology of Written Composition. Erlbaum. (knowledge-telling vs. knowledge-transforming)
  • Bjork, R. A., & Bjork, E. L. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning.
  • Belapur N. et al. (2026). (Im)Paired Programming: Coding Agents Improve Productivity but Harm Understanding. arXiv:2607.26375