AI Tools in Research

How we use language models and related tools, and the one principle that decides the hard cases.

The Principle

These tools are for enhancing your judgment, not replacing it.

You are here to become a researcher. That means developing the ability to read a hard paper, find your own bug, notice when a result is too clean, and know what you think before someone tells you. Every one of those is a skill, and skills are built by doing the thing — including the parts that are slow and uncomfortable.

A tool that does the slow part for you has not saved you time. It has moved a cost from this week to the point, later, where you need the skill and do not have it.

Why We Care About This Specifically

The concern is not that these tools are inaccurate. It is that they are useful, which is what makes the trade easy to take without noticing.

The pattern has a name and predates the technology. Bainbridge’s “Ironies of Automation”¹ observed that the more capable an automated system becomes, the more demanding the human role gets rather than less: operators end up expected to monitor and take over systems they no longer have the skills to run, because the automation removed exactly the routine practice that built those skills.

¹ Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8. Pre-dates and generalizes the AI-specific findings.

² Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., et al. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. https://arxiv.org/abs/2604.04721. Randomized trials, N = 1,222. Preprint.

³ Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. https://arxiv.org/abs/2506.08872. Preprint.

More recent work finds the same shape in tasks closer to ours. A randomized study of AI assistance on reasoning and comprehension tasks found that assistance improved performance while it was available, and that participants then performed worse without it and gave up sooner². An EEG study of essay writing found weaker neural connectivity in LLM-assisted writers than in unassisted ones, and degraded engagement in participants moved from assisted to unassisted work³.

Treat these as suggestive rather than settled — they are recent, some are preprints, and the effect sizes are contested. But the direction is consistent, and it matches what the group cares about: we are trying to produce researchers, not just papers.

The Test

For any use you are unsure about, ask: would a competent version of me be able to check this?

If yes, the tool is an accelerator and you are still the one responsible. If no — if you would be accepting the output because you cannot evaluate it — you have outsourced the judgment, and that is the thing we are trying not to do.

This is a test about you, not about the task. Having a model draft boilerplate in a language you know well is fine. Having it write an analysis in a method you do not understand is not, even though the code may run and the numbers may look reasonable.

Where This Lands in Practice

Reasonable, with you still checking:

Explaining an unfamiliar error, a library, or an equation, as a starting point you then verify
Boilerplate you could have written: plotting scaffolds, argument parsing, file I/O
Rubber-ducking a design decision before you commit to it
Tightening prose you have already written
Adversarial review of your own draft before you send it to anyone

Not reasonable:

Anything going to a third party as your own judgment — a referee report above all. Several journals and funders prohibit it outright, and the Peer Review section below covers why.
Generating citations. Models produce plausible, correctly formatted, non-existent references, and a single invented citation discredits everything else you wrote. If you did not read it, do not cite it.
Writing the parts that are supposed to be your contribution: the hypothesis, the framing, the interpretation, the decision about what the result means.
Analysis in a method you cannot evaluate.

Never Cite What You Have Not Read

This is the failure mode most likely to reach print. A model will give you an author, a year, a journal, and a DOI that look exactly right and describe a paper that does not exist — or a real paper that does not say what you were told.

Verify every reference against the actual source. Not a search snippet, not an abstract when the claim is about what the paper found. This applies to citations a model suggested, and equally to ones you half-remember.

The Group’s Skills

The group maintains claude-research-skills, a set of Claude Agent Skills built around our workflow: adversarial reading, paper planning, literature review, figure design, code review, and revision response.

These are deliberately built as critique tools, not authorship tools, and the design notes in the repository state the stance better than a summary would. From the section on starting a study:

What comes out of this conversation is not your contract. It is a draft, with placeholders, the assistant’s compressions of what you said, and suggestions marked as suggestions. The hypothesis, the scope decision, and the framing have to be yours. […] A contract you did not write is a contract you will not maintain, and its entire value is that it records what you predicted before you knew the answer.

The same logic runs through the rest: referee simulation is scoped to your own papers, the revision-response skill evaluates a reply you have written rather than originating its substance, author lists are never inferred, and the literature skill refuses to name a paper it has not seen.

Installation instructions are in the repository README.

Reproducibility

Our reproducibility commitments do not have an exception for code a model helped write.

You are responsible for code you did not write yourself. If it is in your analysis, you own its correctness. Generated code that runs but is subtly wrong is a serious failure mode, because it removes the syntax errors that usually tell you something is off while leaving the logic errors that do not.
Test it the way you would test any code. Positive and negative controls are the relevant discipline: does it recover known values, and does it correctly find nothing in noise?
This matters most where the numbers are least checkable — uncertainty propagation, tail estimates, anything where a plausible-looking wrong answer will not announce itself.
Internal code review before submission is already required, and the reviewer should be told which parts were generated.

Journal and Funder Disclosure

No major publisher permits AI authorship. The stated reason is consistent: authorship requires accountability, and a tool cannot be accountable. Our own authorship criteria rule it out already, since they require contribution and participation in drafting and revision.

Beyond that, the common shape of current policy is:

Disclosure of substantive use, usually in Methods or an equivalent section. Basic grammar and spelling checks generally do not need declaring.
Full author responsibility for everything in the manuscript, including generated parts.
Peer review is different and stricter — see below.
Image rules vary the most. Some publishers ban AI-generated figures outright; others permit them with disclosure. Check before you make one.

Check Your Target Journal, Every Time

These policies change faster than this page will. The table below was checked in September 2026 and is a starting point, not an authority. Confirm against the journal’s current author guidelines when you submit, and add the disclosure statement to the paper checklist for that submission.

Venue	Policy
AGU	Our primary venue. No AI authors. Disclose tool and use in Methods. LLMs may not be used to review a manuscript.
Elsevier	No AI authors. Required declaration section before the references. Tiered image rules.
Copernicus / EGU	Assistive language use need not be declared; generative use must be. Notably permits AI-generated illustrations if cited.
Springer Nature	No AI authors. Document in Methods. AI-generated images not permitted.
PNAS	No AI authors. Disclose in Methods or Acknowledgments.
AMS	No AI authors. Disclose tool and use. Reviewers prohibited from using AI tools.

NSF

Two separate things, easy to conflate:

Proposal preparation. Disclosure is encouraged, not required — NSF asks proposers to indicate in the project description whether and how generative AI was used (December 2023 notice).
Research misconduct. PAPPG 24-1 Supplement 1, effective December 2025, extends the definition of misconduct to fabrication, falsification, or plagiarism committed “through the use or assistance of other persons, entities, or tools, including artificial intelligence (AI)-based tools.”

The second does not create a disclosure requirement. It does mean a fabricated citation in a proposal is misconduct whether or not a model produced it.

Peer Review and Confidentiality

Do Not Put a Manuscript Under Review Into a Language Model

A manuscript sent to you for review is confidential. Uploading it to an external service is a breach of that confidence regardless of what the service does with it, and essentially every publisher now says so explicitly. NSF likewise prohibits reviewers from uploading proposal content to non-approved tools.

Beyond confidentiality: a review is a judgment you were asked for personally. AMS puts the reason well — AI systems cannot explain the reasoning behind their evaluations, which makes the scholarly exchange a review is supposed to start impossible.

If you are asked to review and want help, ask Vivek. Reviewing is a skill worth building, and doing it badly with a tool is how you avoid building it.

Applicants

Separate guidance applies to people applying to join the group — see the graduate, postdoc, and undergraduate pages.

The short version is the same principle from the other side. An application is the only evidence we have that there is a specific person with specific interests behind it, so a generated one defeats its own purpose. The guided paper evaluation is more permissive precisely because it gets discussed afterwards, which makes outsourcing self-defeating rather than prohibited.

This Page Will Age

The tools, the policies, and the evidence are all moving. If something here is out of date, fix it — and if you find a use that is genuinely valuable and not covered here, raise it at lab meeting. How people actually use these tools well is something the group is still working out, and nobody has this settled.

--- title: AI Tools in Research --- How we use language models and related tools, and the one principle that decides the hard cases. ::: {#the-principle .callout-important} ## The Principle **These tools are for enhancing your judgment, not replacing it.** You are here to become a researcher. That means developing the ability to read a hard paper, find your own bug, notice when a result is too clean, and know what you think before someone tells you. Every one of those is a skill, and skills are built by doing the thing — including the parts that are slow and uncomfortable. A tool that does the slow part for you has not saved you time. It has moved a cost from this week to the point, later, where you need the skill and do not have it. ::: ## Why We Care About This Specifically The concern is not that these tools are inaccurate. It is that they are *useful*, which is what makes the trade easy to take without noticing. The pattern has a name and predates the technology. Bainbridge's "Ironies of Automation"[^bainbridge] observed that the more capable an automated system becomes, the more demanding the human role gets rather than less: operators end up expected to monitor and take over systems they no longer have the skills to run, because the automation removed exactly the routine practice that built those skills. More recent work finds the same shape in tasks closer to ours. A randomized study of AI assistance on reasoning and comprehension tasks found that assistance improved performance while it was available, and that participants then performed *worse* without it and gave up sooner[^liu]. An EEG study of essay writing found weaker neural connectivity in LLM-assisted writers than in unassisted ones, and degraded engagement in participants moved from assisted to unassisted work[^kosmyna]. Treat these as suggestive rather than settled — they are recent, some are preprints, and the effect sizes are contested. But the direction is consistent, and it matches what the group cares about: we are trying to produce researchers, not just papers. [^bainbridge]: Bainbridge, L. (1983). Ironies of automation. *Automatica*, 19(6), 775–779. <https://doi.org/10.1016/0005-1098(83)90046-8>. Pre-dates and generalizes the AI-specific findings. [^liu]: Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., et al. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. <https://arxiv.org/abs/2604.04721>. Randomized trials, N = 1,222. Preprint. [^kosmyna]: Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. <https://arxiv.org/abs/2506.08872>. Preprint. ## The Test For any use you are unsure about, ask: **would a competent version of me be able to check this?** If yes, the tool is an accelerator and you are still the one responsible. If no — if you would be accepting the output because you cannot evaluate it — you have outsourced the judgment, and that is the thing we are trying not to do. This is a test about *you*, not about the task. Having a model draft boilerplate in a language you know well is fine. Having it write an analysis in a method you do not understand is not, even though the code may run and the numbers may look reasonable. ## Where This Lands in Practice **Reasonable, with you still checking:** - Explaining an unfamiliar error, a library, or an equation, *as a starting point you then verify* - Boilerplate you could have written: plotting scaffolds, argument parsing, file I/O - Rubber-ducking a design decision before you commit to it - Tightening prose you have already written - Adversarial review of your own draft before you send it to anyone **Not reasonable:** - Anything going to a third party as your own judgment — a referee report above all. Several journals and funders prohibit it outright, and the [Peer Review](#peer-review-and-confidentiality) section below covers why. - Generating citations. Models produce plausible, correctly formatted, non-existent references, and a single invented citation discredits everything else you wrote. If you did not read it, do not cite it. - Writing the parts that are supposed to be your contribution: the hypothesis, the framing, the interpretation, the decision about what the result means. - Analysis in a method you cannot evaluate. ::: {.callout-warning} ## Never Cite What You Have Not Read This is the failure mode most likely to reach print. A model will give you an author, a year, a journal, and a DOI that look exactly right and describe a paper that does not exist — or a real paper that does not say what you were told. Verify every reference against the actual source. Not a search snippet, not an abstract when the claim is about what the paper found. This applies to citations a model suggested, and equally to ones you half-remember. ::: ## The Group's Skills The group maintains [`claude-research-skills`](https://github.com/srikrishnan-lab/claude-research-skills), a set of Claude Agent Skills built around our workflow: adversarial reading, paper planning, literature review, figure design, code review, and revision response. These are deliberately built as **critique tools, not authorship tools**, and the design notes in the repository state the stance better than a summary would. From the section on starting a study: > What comes out of this conversation is not your contract. It is a draft, with placeholders, the assistant's compressions of what you said, and suggestions marked as suggestions. The hypothesis, the scope decision, and the framing have to be yours. [...] A contract you did not write is a contract you will not maintain, and its entire value is that it records what *you* predicted before you knew the answer. The same logic runs through the rest: referee simulation is scoped to your own papers, the revision-response skill evaluates a reply you have written rather than originating its substance, author lists are never inferred, and the literature skill refuses to name a paper it has not seen. Installation instructions are in the repository README. ## Reproducibility Our [reproducibility commitments](../expectations/open.qmd) do not have an exception for code a model helped write. - **You are responsible for code you did not write yourself.** If it is in your analysis, you own its correctness. Generated code that runs but is subtly wrong is a serious failure mode, because it removes the syntax errors that usually tell you something is off while leaving the logic errors that do not. - **Test it the way you would test any code.** [Positive and negative controls](best-practices.qmd) are the relevant discipline: does it recover known values, and does it correctly find nothing in noise? - **This matters most where the numbers are least checkable** — uncertainty propagation, tail estimates, anything where a plausible-looking wrong answer will not announce itself. - **Internal code review before submission** is [already required](papers.qmd), and the reviewer should be told which parts were generated. ## Journal and Funder Disclosure **No major publisher permits AI authorship.** The stated reason is consistent: authorship requires accountability, and a tool cannot be accountable. Our own [authorship criteria](authorship.qmd) rule it out already, since they require contribution *and* participation in drafting and revision. Beyond that, the common shape of current policy is: - **Disclosure** of substantive use, usually in Methods or an equivalent section. Basic grammar and spelling checks generally do not need declaring. - **Full author responsibility** for everything in the manuscript, including generated parts. - **Peer review is different and stricter** — see below. - **Image rules vary the most.** Some publishers ban AI-generated figures outright; others permit them with disclosure. Check before you make one. ::: {.callout-warning} ## Check Your Target Journal, Every Time These policies change faster than this page will. The table below was checked in **September 2026** and is a starting point, not an authority. Confirm against the journal's current author guidelines when you submit, and add the disclosure statement to the [paper checklist](paper-checklist.qmd) for that submission. ::: | Venue | Policy | |---|---| | [AGU](https://www.agu.org/publications/authors/policies) | Our primary venue. No AI authors. Disclose tool and use in Methods. LLMs may not be used to review a manuscript. | | [Elsevier](https://www.elsevier.com/about/policies-and-standards/generative-ai-policies-for-journals) | No AI authors. Required declaration section before the references. Tiered image rules. | | [Copernicus / EGU](https://publications.copernicus.org/for_authors/ai_policy.html) | Assistive language use need not be declared; generative use must be. Notably permits AI-generated illustrations if cited. | | [Springer Nature](https://group.springernature.com/gp/group/ai/ai-guidance-for-our-researchers-and-communities) | No AI authors. Document in Methods. AI-generated images not permitted. | | [PNAS](https://www.pnas.org/author-center/editorial-and-journal-policies) | No AI authors. Disclose in Methods or Acknowledgments. | | [AMS](https://www.ametsoc.org/ams/publications/ethical-guidelines-and-ams-policies/author-disclosure-and-obligations/) | No AI authors. Disclose tool and use. Reviewers prohibited from using AI tools. | ### NSF Two separate things, easy to conflate: - **Proposal preparation.** Disclosure is *encouraged, not required* — NSF asks proposers to indicate in the project description whether and how generative AI was used ([December 2023 notice](https://www.nsf.gov/policies/ai/merit-review)). - **Research misconduct.** [PAPPG 24-1 Supplement 1](https://www.nsf.gov/policies/document/pappg24-1-supplement-1), effective December 2025, extends the definition of misconduct to fabrication, falsification, or plagiarism committed "through the use or assistance of other persons, entities, or tools, including artificial intelligence (AI)-based tools." The second does not create a disclosure requirement. It does mean a fabricated citation in a proposal is misconduct whether or not a model produced it. ## Peer Review and Confidentiality {#peer-review-and-confidentiality} ::: {.callout-important} ## Do Not Put a Manuscript Under Review Into a Language Model A manuscript sent to you for review is confidential. Uploading it to an external service is a breach of that confidence regardless of what the service does with it, and essentially every publisher now says so explicitly. NSF likewise prohibits reviewers from uploading proposal content to non-approved tools. Beyond confidentiality: a review is a judgment you were asked for personally. AMS puts the reason well — AI systems cannot explain the reasoning behind their evaluations, which makes the scholarly exchange a review is supposed to start impossible. If you are asked to review and want help, ask Vivek. Reviewing is a skill worth building, and [doing it badly with a tool is how you avoid building it](#the-principle). ::: ## Applicants Separate guidance applies to people applying to join the group — see the [graduate](../procedures/graduate.qmd), [postdoc](../procedures/postdoc.qmd), and [undergraduate](../procedures/undergrad.qmd) pages. The short version is the same principle from the other side. An application is the only evidence we have that there is a specific person with specific interests behind it, so a generated one defeats its own purpose. The guided paper evaluation is more permissive precisely because it gets discussed afterwards, which makes outsourcing self-defeating rather than prohibited. ## This Page Will Age The tools, the policies, and the evidence are all moving. If something here is out of date, [fix it](../contributing/index.qmd) — and if you find a use that is genuinely valuable and not covered here, raise it at [lab meeting](meetings.qmd). How people actually use these tools well is something the group is still working out, and nobody has this settled.