Using AI in Your Research and Writing: Guidelines for My Students
Branislav L. Slantchev (slantchev@ucsd.edu)
Department of Political Science, University of California San Diego
October 30, 2026
TL;DR Think of AI as a well-read colleague giving you feedback rather than as a research assistant who can do things you can't. As with a colleague, not all advice will be good and worth taking, but a lot of it will deserve attention. Never rely on AI output you can't verify yourself. For the parts that carry the argument (the model, the methods, the analytical choices) verifying isn't enough: you have to understand them, because that is where the argument lives. And unlike a human collaborator, AI shares none of the responsibility: if it's in your paper, it's yours.
The rest of this note explains what that means in practice: why I care how your drafts are produced, how to get the most out of AI as a colleague, where the line is for writing, models, and data, and what you owe your coauthors and other readers. Near the end is a stage-by-stage workflow with example prompts, and a checklist to run before anything goes out.
Why I care how your drafts are produced
When I read your drafts, I am reading to understand your ideas. Bad writing is useful to me: an awkward paragraph or a confused transition usually points to a problem in your reasoning or analysis, and that is exactly where I can help. If you smooth the draft over with AI before I see it, the problem does not go away. It just goes underground, and it surfaces anyway, usually late in the process, when it is far more costly to fix (a job talk, a referee report, a committee meeting).
So when you send me a draft, send me your draft. I would much rather read something rough that shows me how you think than something polished that hides it. That includes the language: drafts for me do not need polished English. Polishing comes later, for outside readers (see the section on English below). This is part of my job as your advisor, and it is the main reason the rules below are stricter for you than for anyone else whose work I read.
Good uses: AI as a well-read colleague
The right model is a well-read colleague who is always available and never tired of reading your drafts. Write first, then ask it to read. Like any colleague, it will sometimes be wrong, miss the point, or push you toward its own view of the paper, so weigh its advice the way you would weigh a colleague's: take what survives scrutiny and ignore the rest. But a lot of what it says will deserve your attention, and used this way it is very efficient. Some things that work well:
- Finding what will confuse readers. Give it a section you wrote and ask what is unclear, where a reader might get lost, what terms are undefined, and what could help clarify the point. This complements presenting your work to real people; it does not replace it. Talk about your ideas with anyone who will listen. A real person's offhand question can stop you dead in your tracks in a way a chatbot rarely will.
- Stress-testing the argument. Ask it to make the strongest case against your claim, to find gaps in the logic, or to say what a hostile referee would go after first. Then decide for yourself which objections are real.
- Bouncing off ideas. Talk through a puzzle, an alternative mechanism, or a research design before you commit to it.
- Learning techniques. Use AI as a tutor to learn a new estimator, a solution concept, or a proof technique: ask it to explain, to work through examples, to quiz you. This is different from having it do the work for you. The goal is that afterward you can do it yourself.
- Looking for related work, especially outside our field. AI is good at pointing to arguments in other disciplines that bear on what you are doing. I have learned about relevant work this way that I did not know existed. Treat these as leads, not as a literature review. Reviews written by scholars working in the field are preferable, and an AI-generated map is no substitute for reading them. In either case, read the actual sources before you cite or rely on them, because AI can misdescribe an argument or invent a reference.
- Better phrasing and clarity. Ask for alternative ways to say something you have already written, then choose (and adjust) what fits your meaning.
None of this is a substitute for thinking. It is a way to get feedback faster and more often than you could get it from people.
Bad uses: letting AI build the argument
The line that matters is between using AI to get feedback or fix language and using it to build the argument for you. Framing the question, positioning it against the literature, deciding what the contribution is, and structuring the argument are intellectual work. That work is your job, and it is what I am trying to help you learn to do. AI may suggest an idea, mechanism, or approach you had not considered, just as a colleague might. You can pursue it, but you must work out why it makes sense, develop it yourself, and decide whether it survives scrutiny. Adopting a suggestion is not the same as understanding it.
In practice, do not:
- Ask AI to write a section from scratch and then edit it into shape. You will end up endorsing framing choices you never made and may not fully understand.
- Accept suggested rewrites wholesale. Every sentence that goes in should say exactly what you mean, and you should be able to defend it in a seminar without notes.
- Run my comments back through AI and send me whatever comes out. If you do that, I am in a conversation with the model and you are just relaying messages.
- Use AI to paper over a part of the argument you are unsure about. If you cannot write it clearly yourself, that is a signal the thinking is not done yet. Come talk to me instead.
The test I use: if I asked you to explain a paragraph without looking at it, could you? If not, it is not yours yet.
Data, code, models, and methods: verify, understand, own
Many people worry about AI-written prose and think nothing of having AI collect, collate, and analyze data, or write the programs that do it. In my view this is backwards. With English text you can at least read the output and see where it departs from what you meant. With a model, an estimation, or a program, the problems are often invisible unless you already have the skills to see them, and people often turn to AI precisely because they lack those skills. Three rules apply here, and they apply to different kinds of work.
Verify what is mechanical. Some tasks really are mechanical: the syntax of R or C++ once you know what the program should do, data formatting, reference management, language editing. AI can do these for you, but you must be able to tell when it got them wrong. For code, that means testing it on simulated data with known answers, checking special cases by hand, and comparing against an independent implementation. The standard is not whether you could have written it yourself, but whether you can tell when it is wrong.
Understand what carries the argument. A formal model is an argument (Slantchev 2017, 2020b). The math is a form of accounting: it guarantees that the conclusions follow from the premises. The argument itself lives in the choices around the math: which premises go in, how the interaction is structured, what the players know and want, and how the parameters are interpreted. None of these choices is rigorous, and all of them are yours to make and defend. If you do not understand why the result holds, you do not have the argument, and the whole point of the exercise is lost.
Checking the math is not enough. When Fey and Ramsay (2007) argued that mutual optimism cannot cause war, I could follow every step, everything seemed correct, the conclusion followed, and I still did not believe it. It took building alternative models to discover that an innocuous-looking definition was doing all the work (Slantchev and Tarar 2011; the story is in Slantchev 2017). I really do not like solving models, but I do not outsource the analysis, not even to smart graduate students, because when you do not understand how the analysis works, you do not understand what the model is telling you. A simple test: if you cannot explain the behavior of your agents without reference to equations, you have a problem.
Be especially careful with AI when building models. It is very good at producing a model that yields whatever result you ask for, usually through some quirky assumption that buys the conclusion while the rest of the math is ornamentation. You can rationalize almost anything this way, up to bank presidents setting their pants on fire in the street (Postrel 1991; Slantchev 2017). If you do not understand the model, you will not see where the rabbit went into the hat, but your referees will. Do not formalize just to give your hypothesis a formal gloss, and do not ask AI to construct the model for you.
The same holds for statistical methods and the choices made in analyzing the data: specification, identification, measurement, what to control for. These are substantive claims, not mechanics. Check what your variables actually measure, not what they are called. A widely used "immediate balance of forces" variable (Huth and Russett 1984) was treated in later work as information revealed during a crisis (Fearon 1994; Signorino and Tarar 2006), but it is built entirely from indicators observable before the crisis. Researchers trusted the label instead of the coding (Slantchev 2017). AI output is full of plausible labels.
Code is not purely mechanical either. Working out the algorithmic solution to a problem forces you to think it through, which is why writing code often clarifies your thinking. And code embeds choices: how missing data are handled, which observations a merge silently drops, default standard errors, bandwidth selectors, solver tolerances. AI-written code makes these choices for you, often by default and without telling you. Verifying code means knowing what choices it made, not just that it runs.
Own all of it. Scholars routinely collaborate with people who can do things they cannot, and we rely on those colleagues to do the work properly. This mostly works, but it occasionally ends in scandal when someone in the research pipeline turns out to have been less than careful or honest with the data or analysis. Even then, "my coauthor was responsible for the analysis" is barely acceptable, because everyone whose name is on the paper has to stand behind its results. "The AI did it" is not acceptable at all. A human collaborator shares the responsibility; AI shares none of it. If you do not understand what the AI did, nobody in the pipeline does.
I have found math errors in papers more than once that turned out to come from students using AI without catching the problem. That is the kind of mistake that can sink a paper or a reputation. In practice:
- Work through every derivation yourself until you can explain the intuition behind the result without notes and could walk someone through every step of the proof with the paper in front of you. AI can suggest a solution strategy when you are stuck, provided you then reconstruct, verify, and understand the solution independently. But if you are learning formal theory, I strongly discourage this shortcut. Getting stuck, trying different approaches, and figuring out how to proceed are essential parts of the training. Following a solution is much easier than finding one yourself, as I explain in my advice against group work in Analytic Theory II (Slantchev 2020a). Use Mathematica or AI to check your derivations rather than to spare yourself that work.
- Write down the intuition for each result in plain prose as you derive it, as if explaining it to someone unfamiliar with the model (Slantchev 2020b). This is your test of understanding, and it is the one thing you should never delegate.
- Justify every premise. Before the model specification, write a paragraph listing the features the model must have to represent your puzzle and why. Anything you cannot justify there should come out, and this applies doubly to assumptions AI suggested.
- Use AI for numerical exploration, then prove. Having AI write code for numerical examples, simulations, and plots of best responses or comparative statics is a good way to get a feel for a model before solving it. The patterns you find are conjectures until you prove them analytically.
- Make every methodological choice deliberately and be able to defend it. If AI suggested it, find out why it is standard and what the alternatives are.
- Understand and test every consequential part of AI-written code, including the analytical choices and defaults it relies on.
- Use AI to check your work, not just to produce it. It is good at finding bugs and gaps in work you did yourself, which is a much safer direction than the reverse.
- If AI is part of your data pipeline, treat it as a method. Document what it does, validate it against human coding, report error rates, and describe it in the paper as you would any other measurement instrument.
If English is not your first language
We are not in the English Department, and research is evaluated on its ideas and on how well it was done. But a weak grasp of idiomatic English can obscure good ideas, and readers do penalize that, fairly or not. Using AI to improve your language is legitimate, in the same way that hiring a copyeditor always has been. I see no principled difference between the two.
A few cautions, though:
- Drafts for me do not need polishing. Use AI on the language when the work is going to outside readers, not before I see it.
- Write the draft in your own words first, then use AI on the language. Ask it to fix grammar and idiom, not to rewrite the argument.
- Check every change against your meaning. Fluent text can quietly shift a claim, add a hedge, or overstate a result. You may be the person least able to notice, so read slowly and compare.
- Keep the register consistent. A paper where some sections are AI-smooth and others are rough reads oddly and invites suspicion. Better to bring the whole thing to the same level.
- Proofread everything yourself, and make the abstract and introduction flawless. The whole paper is your responsibility, so read all of it carefully before it goes out. Be 100% sure the abstract and introduction are correct and free of grammatical mistakes: they are read first and by everyone, and many readers never get further. AI edits followed by your own changes are a common way for new errors to slip in.
This is advice for native speakers as well. Keep working on your own writing. Read widely and emulate writers you admire. Sword (2012) and Thomas and Turner (2011) are excellent on prose in general, and McCloskey (2019) and Thomson (2001) are indispensable for academic work in economics and formal theory. AI can help here if you give it examples of writing you want to emulate and discuss with it why you like that style, or give it the principles from the books above. Together you can build a list of principles of your own to apply consistently and revise as your writing improves.
Disclosure and coauthorship
When someone asks you to write something (a referee report, a memo, a section of a joint paper), disclose substantive use of AI to generate text, code, or analysis; asking for feedback on your own draft is different. There is no good reason to hide it, and finding out after the fact damages trust far more than the AI use itself would have. For confidential work, including referee reports, first check whether AI use is permitted at all by the journal or institution.
My rules for papers we write together (the first three are about AI; the last is how I run every joint paper, AI or not):
- Clear it with me before using AI to generate text, analysis, or code for our joint work. Asking AI for feedback on a draft, as described above, does not need clearance. This applies to all my coauthors, not only students.
- Tell me whenever AI touched the analysis or the math, so we know what needs to be checked independently.
- Follow journal and funder policies. Many now require a statement on AI use. Read them and comply; they change often.
- Do not edit the text of a joint paper directly. Send me comments, suggestions, and proposed wording, and I will decide what goes in and how. This keeps one consistent voice and one person accountable for every sentence.
What not to share with AI
Anything you paste into an AI tool leaves your control, and some material is not yours to share:
- Manuscripts you are reviewing, and your reports on them. Peer review is confidential, and many journals prohibit uploading submissions to AI tools. Check the journal's policy; when in doubt, do not.
- Coauthors' unpublished work, without their permission.
- Data covered by IRB protocols or data-use agreements, including interview transcripts and anything that could identify a person. Check your protocol before any AI use; your approval may not cover it.
- Confidential material: letters of recommendation, evaluations, applications, personnel files.
If the university provides an AI tool with data protections, use it for anything sensitive. Those protections do not override journal or IRB rules.
Readers outside our group
Norms around AI are unsettled, and colleagues hold very different views. When I evaluate other people's work, I judge the ideas and the research. Some readers, however, will refuse to take seriously anything they suspect was AI-generated, and some will assume it from polished or uneven prose alone. You cannot reliably control what readers suspect, especially as AI-generated prose improves. What you can control is the quality of your work and your ability to explain it:
- Before you send work to anyone outside our group (letter writers, discussants, editors, fellowship committees), make sure every sentence is one you would stand behind as your own.
- Follow the advice on language above: a consistent register and a clean abstract and introduction make the paper clearer and more professional, regardless of what readers assume about AI.
- If you used AI substantively and the context calls for it, say so briefly and plainly.
- If someone raises the question, explain how you used AI and what is yours. If it goes further than that, let me know.
A workflow, start to finish
Here is what this looks like over the life of a paper. At every stage the pattern is the same: you do the thinking and make the choices, AI reads, questions, checks, and handles the mechanical parts, and you verify what comes back. The prompts are examples, not scripts.
- Puzzle and literature. You: find the puzzle. You have to know enough to understand that you do not understand something. AI can: suggest leads in the literature and point you to work in other fields. Start with reviews written by scholars in the field whenever possible; an AI-generated map is not equivalent. Try "What literatures address why autocrats tolerate some dissidents and jail others? List the main arguments and key papers" or "Is there work in economics or sociology on the same problem?" Check: read every paper before you rely on it. Confirm that each reference exists and says what the AI claims it says.
- The argument. You: write a page stating your argument in plain words, including what you assume and why. AI can: attack it. Try "Here is my argument. What are the three strongest objections? What would a skeptic offer as an alternative explanation? What am I assuming without saying so?" Do not ask: "Give me an argument for X."
- The model (if you are doing theory). You: write the paragraph listing the features the model must have, then build and solve it. AI can: point you to similar existing models ("Which published models have a setup like this?"), find what might be driving a result ("Which of these assumptions is doing the work in Proposition 2?"), write code for numerical exploration ("Write R code that computes the cutoff types over a grid of costs and plots them"), check your algebra ("Here is my derivation of the cutoff. Check each step for errors and tell me where, without fixing it"), and teach you a technique on a simpler example ("Walk me through finding a PBE in a two-type signaling game"). Do not ask: "Write me a model in which X leads to Y." Check: prove analytically whatever the simulations suggest, and do the derivations by hand.
- Data and analysis (if you are doing empirics). You: choose the design, the measures, and the specification, and be able to say why. AI can: write cleaning and estimation code to your specification ("Merge these two files on country-year and report every row that fails to match"), explain defaults ("What does this function do with missing values and standard errors by default?"), and suggest robustness checks ("What are the standard robustness checks for this design, and what does each one guard against?"). Check: test the code on simulated data with a known answer, count the observations at every step, and read the documentation for anything the AI set by default.
- Writing. You: write the draft. AI can: read it as a skeptical referee ("Where is this section unclear? What does a reader need to know before the third paragraph that I have not told them?"), offer alternatives for sentences you wrote ("Give me three other ways to say this"), and, if English is not your first language, fix the language ("Correct grammar and idiom only. Do not change the content. List every change you make"). Check: compare every change against what you meant.
- Sending it out. To me: send the rough draft as is. To anyone else: go through the checklist below and proofread the whole paper yourself, with extra care on the abstract and introduction ("List any errors in this section; do not rewrite it" is a fine second pass, not a substitute).
A few habits make the interaction much more useful:
- Ask for critique, not content. "What is wrong with this?" teaches you something; "Write this for me" does not.
- Ask it to list changes, not rewrite. A list of suggested changes lets you decide each one. A rewrite slips choices past you.
- Give it context. Paste your actual text, say who the audience is, and say what you are trying to do. Vague questions get generic answers.
- Push back. Ask it to justify a claim, to give a source, or to say what would change its mind. If it simply caves, its first answer was not worth much either.
- Watch out for agreement. These tools tend to agree with whoever they are talking to, especially in long conversations. Ask explicitly for the opposing case, and start a fresh conversation when you want a fresh read.
Know what its critiques are worth:
- It is uneven. It is sometimes overly cautious, hedging claims that are fine, and sometimes generalizes boldly and incorrectly. Its criticism can be nitpicky in the worst possible way: a dozen trivial complaints that bury the one that matters, all delivered with the same confidence. Above all, it is bad at prioritizing problems. Some of what it lists will be trivialities or confusions caused by sloppy writing; others may be fundamental and would break the argument unless fixed. Do not rely on its ranking, because it will often mistake the forest for the trees. You have to decide which critiques to take seriously and which to discard.
- It is good at catching inconsistencies. A term defined one way and used another, notation that drifts between sections, a claim in the introduction that the results do not deliver, numbers that do not match across tables and text. This is a big part of why asking it to critique your draft pays off.
- It is often bad at detecting flaws in the logical structure of an argument. It can miss an assumption that does all the work, a conclusion that is built into the premises, or a step that does not follow, especially when the prose reads smoothly. Do not rely on it for this.
The upshot: when AI says something is fine, that is no proof that it is. You will have to stand behind the claims on your own. When AI says something is wrong, that might be a signal of a real problem, or it might not. You will have to decide whether it is, and what to do about it.
Before you send anything out
- I wrote the draft myself; AI was used for feedback, checking, tutoring, or language, not to build the argument.
- I can explain every paragraph and the intuition behind every result, and I could walk someone through every consequential step in the code and every step of every derivation.
- I can explain what drives each result without reference to equations, and I can justify every premise of the model.
- Any code or math AI touched has been tested or re-derived independently, and I know what defaults the code relies on.
- Every source AI suggested has been read and checked by me.
- Every AI language edit still says exactly what I mean.
- I proofread the whole paper myself, the prose is consistent throughout, and the abstract and introduction are free of errors.
- I have not uploaded anything confidential: manuscripts under review, coauthors' unpublished work, or protected data.
- I have disclosed AI use where it was asked for or expected, including to coauthors.
Disclosure: how this document was written
I wrote this policy with Claude, following the approach it describes. The original argument came from a reply I wrote to a colleague about AI and student writing, and from my work on using formal models. Claude turned that into a set of guidelines, which I edited iteratively using its feedback until I was satisfied that the result correctly conveys my ideas. I then circulated the draft among colleagues and incorporated their feedback. I stand behind every word in this document.
References
Fearon, James D. 1994. "Signaling versus the Balance of Power and Interests: An Empirical Test of a Crisis Bargaining Model." Journal of Conflict Resolution 38: 236–269.
Fey, Mark, and Kristopher W. Ramsay. 2007. "Mutual Optimism and War." American Journal of Political Science 51: 738–754.
Huth, Paul, and Bruce Russett. 1984. "What Makes Deterrence Work? Cases from 1900 to 1980." World Politics 36: 496–526.
McCloskey, Deirdre N. 2019. Economical Writing: Thirty-Five Rules for Clear and Persuasive Prose. 3rd ed. Chicago: University of Chicago Press.
Postrel, Steven. 1991. "Burning Your Britches Behind You: Can Policy Scholars Bank on Game Theory?" Strategic Management Journal 12: 153–155.
Signorino, Curtis S., and Ahmer Tarar. 2006. "A Unified Theory and Test of Extended Immediate Deterrence." American Journal of Political Science 50: 586–605.
Slantchev, Branislav L. 2017. "On the Proper Use of Game-Theoretic Models in Conflict Studies." Peace Economics, Peace Science and Public Policy 23(4): 1–14.
Slantchev, Branislav L. 2020a. Analytic Theory II (POLI 203B), course syllabus.
Slantchev, Branislav L. 2020b. "The Simple, the Trivial, and the Insightful: Field Dispatches from a Formal Theorist." In The SAGE Handbook of Research Methods in Political Science and International Relations, ed. Luigi Curini and Robert Franzese. London: SAGE.
Slantchev, Branislav L., and Ahmer Tarar. 2011. "Mutual Optimism as a Rationalist Explanation of War." American Journal of Political Science 55: 135–148.
Sword, Helen. 2012. Stylish Academic Writing. Cambridge, MA: Harvard University Press.
Thomas, Francis-Noël, and Mark Turner. 2011. Clear and Simple as the Truth: Writing Classic Prose. 2nd ed. Princeton: Princeton University Press.
Thomson, William. 2001. A Guide for the Young Economist: Writing and Speaking Effectively about Economics. Cambridge, MA: MIT Press.