Undergraduate research with AI

This guide is for an undergraduate starting mentored research: a summer program, an honors or senior thesis, a project in a faculty member’s lab or group, a first conference poster. You already have the practices from Part I of the book: try first, verify, keep the work yours, disclose how AI helped. Research raises the bar on all of them. This guide adapts those practices to the research setting, and points you to the graduate-manuscript standard when your work heads that way.

Before you start: five things that change in research

Research is not coursework. In a course, an unverified answer costs you a grade. In research, an unverified result travels. It goes into your poster, your thesis, your mentor’s report to a funder, and sometimes into the published record. The cost of a result that looks finished but is not rises with every place it travels. That is the reason for everything below.

You are learning to do research, not to produce it faster than you understand it. The point of a research experience is to build research judgment and the ability to check what a tool hands you. That checking ability is what lets a researcher use AI for serious work later. A researcher who can verify what AI produces can put it to real use. One who cannot is not helped by it but exposed by it. Build that ability now, while the stakes are still low and someone is there to teach you.

Your mentor is the relationship to protect. Research is mentored, and the mentoring is the part that makes you a researcher. Ask your mentor and the group what their norms are for AI use, and follow them. Tell them how you are using it. Hiding it corrodes the one relationship the experience is built on, and it costs you the feedback that research training runs on. When a norm is unclear, ask. Asking is not a sign you do not belong. It is how research works.

The one rule, research grade. You must be able to explain and defend every result as your own, without the tool in front of you: to your mentor, at your poster, in your thesis. If you cannot reconstruct a result without the AI, it is not a result yet. It is a draft you have not finished. A useful test comes from the book’s appendix on scientific writing: if someone asks why you did it this way, your answer should come from understanding the work, not from what the tool suggested.

Set up your tool once. Before you ask for anything, tell the AI your project, your field, your level, and what your group is working toward. Ask it to show its work, and to flag the claims you should check and how to check them. The book’s Define Personal Preferences walkthrough and the Study Partner Protocol set this up, and the same setup serves research. Output you can check is worth far more than output that merely sounds right.

The research arc

Research runs as a sequence, so this guide follows one. At each stage the same three questions apply: what is the tool good for here, how do you use it well, and how do you make sure the result is yours.

1. Finding and scoping a question

AI is good for orienting you fast in an unfamiliar area. Ask it what the open questions are, what the standard methods look like, and what terms to search. Use it to get your bearings, then bring candidate directions to your mentor.

The question you work on is yours and your mentor’s to choose. It is shaped by what the group needs, what equipment and data you can actually reach, and what has already been tried. A tool does not know any of that. Do not let it pick your problem. Let it help you understand the landscape so you can have a better conversation with your mentor about where to stand in it.

2. Reading the literature

AI is good for finding related work, summarizing a paper before you read it, and explaining a dense passage in plainer words. Two rules keep this safe.

First, verify every reference. AI invents citations that look completely real: a plausible author list, a plausible title, a plausible journal, and no such paper behind them. Check each reference against the actual paper, or a database your field trusts, before it goes anywhere near your writing. An invented citation in a poster or thesis is the kind of error a careful reader catches first.

Second, read the papers you cite. A summary is a map, not the ground. Your mentor will ask you about a paper on your reference list, and “the AI summarized it” is not an answer you want to be giving. Use the summary to decide what to read, then read it.

3. Methods, code, and data

AI is good for scaffolding code, explaining an unfamiliar method, finding a bug, and planning an analysis. Use it to move faster on the mechanical parts of the work. Keep three lines firm.

Understand the method well enough to defend it. Do not run an analysis you cannot explain, because you will have to explain it, to your mentor and on your poster.

Test any code you did not write by hand. Run it on inputs whose right answers you already know, check the edge cases, and compare against a benchmark or a case you can work out yourself. Code that runs without an error is not the same as code that is correct.

Protect the group’s unpublished work. This barely comes up in coursework, and it matters here: do not paste your lab’s unpublished data, code, or ideas into a public tool without asking. Many tools keep what you give them and may reuse it. Unpublished work is your mentor’s and the group’s to protect. Check the tool’s data-retention terms, and keep anything not already public inside the environment your group approves.

4. Doing the work and verifying at research grade

This is the center of the whole guide. AI produces fluent, confident output that can hide a single weak step: an argument that reads cleanly with one shaky link, a number that looks right and is wrong. In a course, that costs you a grade. In research, it can carry a wrong result into a poster or a paper, where other people build on it.

So verification is part of the work, not a step you save for the end. Reproduce your own key results from scratch, without the tool, before you report them. And because these tools are stochastic, meaning the same prompt can return a different answer the next time you run it, anything you plan to report should still hold when you run it again. A result that rests on one lucky run is not yet a result.

The habit underneath all of this is one the book calls self-explanation: say, in your own words, why a result is true and how you checked it. If you can do that cleanly, the result is yours. If you cannot, you have just found the next thing to work on.

5. Writing it up: poster, report, or thesis

AI is good for turning your notes into a first draft, tightening loose prose, and helping with clear, idiomatic English if it is not your first language. That last use has a side benefit: it leaves a record of what you wrote and what the tool suggested, and that record protects you if your writing is ever questioned.

Watch one specific trap when a tool edits your prose. It quietly strengthens claims. A sentence that said a result “suggests” something comes back saying it “shows” or “proves” it. Read every edited paragraph and put the hedging back where your evidence actually sits. Every claim in your poster or thesis has to be one you can stand behind.

When your work starts heading toward a journal, a real manuscript submitted for peer review, you have crossed into a stricter standard. The book’s appendix on ethical AI-assisted scientific writing is written for exactly that step, and it is where this guide hands off.

6. Disclosure and credit

Be honest about how AI helped, both in your writeup and with your mentor. Match the detail to the use: a line for light editing help, a clearer note when a tool drafted text or assisted an analysis.

A short research disclosure block can read like this:

AI assistance: I used [tool] to draft parts of this [poster/report], to help debug the analysis code, and to find related work. I verified every result independently, read each paper I cite, and tested the code against known cases. The research questions, methods, analysis, and conclusions are my own, developed with my mentor.

Change it to match what you actually did. The book’s disclosure templates and its scientific-writing appendix give fuller versions for longer or more formal work.

One rule holds in every field: AI is not an author. Authorship means taking responsibility for the work, and a tool cannot take responsibility. When you are unsure how to credit AI’s help, or how to credit your group’s, ask your mentor. Credit is a place where guessing quietly is worse than asking plainly.

Tools that help

The stages above do not depend on any particular tool. The list below maps current tools to those stages, to save you the search. It is a snapshot from 2026, and this kind of list ages fast, so treat each name as an example of a category and check what your field and your group actually use. Most of these have free versions worth trying before you pay for anything.

General assistants. ChatGPT, Claude, and Gemini are the everyday workhorses for explaining a method, drafting prose, and finding a bug. They are strong once the papers and the problem are in front of you. They are also the tools most likely to invent a citation, so keep them away from your reference list unless you check every entry.

Finding and mapping the literature. Paper-grounded search tools tie their answers to real papers, which makes them safer than a general assistant for this job. Semantic Scholar is a free discovery engine across a large index. Consensus answers focused evidence questions from published studies. For seeing how a field connects, Connected Papers and Research Rabbit map citation networks. Elicit pulls structured details out of many papers at once. Even with these, cross-check every citation against the actual paper before you use it.

Reading and synthesizing. NotebookLM works across sources you upload and keeps its answers tied to them, which helps when you are summarizing a stack of papers and do not want the tool drifting into invention. SciSpace helps with dense PDFs and heavy notation.

Checking a finding. Scite shows whether later papers supported, contradicted, or merely mentioned a result, so you do not build on a finding that has since been challenged.

Managing references. Zotero is a free, open-source reference manager that stores your sources and formats citations. It is also a good place to confirm a reference is real, since you add it from the actual record rather than from AI text.

Code and analysis. GitHub Copilot is free for students and works inside common editors; Cursor has a free tier and can reason across a whole project. For understanding an unfamiliar method or a stubborn bug, the general assistants do well. Whatever you use, test the code it gives you, since running without an error is not the same as being correct.

Why this matters now

A first research experience is where you build the judgment that lets you use AI as a real tool instead of being fooled by it. The habits in this guide are the same ones that graduate research and published science run on: verify before you report, read what you cite, defend what you claim, and tell the truth about how the work was made. Building them now, with a mentor beside you, is the point of the whole exercise. When your work reaches manuscript stage, the book’s scientific-writing appendix picks up where this guide leaves off.

How this connects

Part of the companion repository for Learning with AI: A Framework for Students, Instructors, and Universities. These materials are free to use and adapt with attribution.