How to disclose AI use in manuscripts and grants: Your trainees already use AI. Whether they tell you depends on you.

Table of Contents

Somewhere between 28% and 76% of researchers tell surveys they use AI when they prepare a manuscript.

When BMJ Group made the disclosure question mandatory across 49 journals, 5.7% of 25,114 submissions reported any AI use.

Different samples from different years. But a gap that wide still needs an explanation. And researchers have already given one.

Oxford University Press surveyed 2,606 researchers this year. Of the journal authors, 46% said they had not fully disclosed their AI use in the past year.

  • 44% worried that disclosure would change how others view their work, and
  • 74% were unclear about what they are supposed to disclose.

The stigma of AI use unfortunately runs deep in academia.

A longer policy will not close that gap. The people who can close it run labs, divisions, and institutions.

👋🏽 Quick announcement: Our Black Friday deal for Research Boost (AI academic writing agent) is live. Use code RBBLACKFRI20 to get 20% OFF for 12 months on both monthly and yearly plans: https://researchboost.com/pricing

1. The disclosure gap comes from fear, and the fear has evidence

Two more measurements point the same way.

The fear is partly justified. Schilke and Reimann ran 13 experiments with more than 5,000 participants and found that disclosing AI use lowered trust. But trust fell furthest when the authors did not disclose and a third party exposed the AI use later.

At a recent NIH K Forum, the NIH officials were comfortable with AI as long as the ideas and the science belong to the researcher. However, the researchers I spoke with afterward were far more resistant. They felt “disappointed” and even “disgusted” that their trainees or colleagues used AI.

The written rules are calmer than the hallway. Closing that distance is a leadership job.

2. Lab leaders set the norm before any policy does

Trainees learn what is acceptable by watching you. If you never mention AI, they conclude it is something to hide.

A senior colleague was angry that a mentee had looked up a method in ChatGPT, which invented one. The mentee skipped one step, which was checking the output against a source. In a group where AI use is shameful, nobody teaches that step. You learn someone is using the tool only at the moment it breaks.

Four habits change that:

  1. Go first. Tell your group which AI tools you use, for which tasks, and where you refuse to use them.
  2. Encourage the use, then coach it. Review a trainee’s prompt and output the way you would review a statistical analysis plan.
  3. Ask two questions on every draft. What did you use? How did you check it?
  4. Treat errors as teaching cases. When a fabricated citation shows up, walk through the verification step that was skipped. If you punish the use, you will simply stop hearing about it.

Then write it down in a one-page lab SOP: approved tools, data rules, a shared AI log, and the disclosure templates below.

3. The message from the top decides whether people hide

Data security deserves serious attention. So does research ethics. Patient data, unpublished results, and confidential reviews do not belong in a consumer chatbot.

The tone of the message matters as much as the rule. “AI tools are blocked” and “here is the approved tool, here is the data you can put in it, and here is who to ask” protect the same data. Only the second keeps the use where you can see it. (Also, that internal AI tool needs to be as good or better than the available alternatives for people to actually use it.)

A University of Melbourne study with KPMG >48,000 people, found that 57% of employees conceal their AI use and pass off AI-generated work as their own. Only 40% said their workplace had any policy or guidance on generative AI.

If you are a chair, a dean, or a research office, a useful message has three parts:

  • A named, approved AI tool that people can use today.
  • A data map with four tiers (public, institutional, restricted, critical), each matched to the tools it may enter.
  • A plain sentence saying that AI use within these rules is encouraged, with training to back it up.

A blanket block pushes the use underground, where nobody can teach it or check it.

4. People will not share a workflow they think will replace them

Picture your research coordinator. She has built an AI workflow for chart screening that carries a decade of knowing which records cause trouble. That is tacit knowledge, written down in a form a machine can run.

Now ask what happens to her job if she shows you.

In the 2024 Work Trend Index from Microsoft and LinkedIn, a survey of 31,000 knowledge workers, 53% of AI users feared that leaning on it for important tasks would make them look replaceable.

Ethan Mollick names the same worry in his new book Co-Existence. People suspect they are training their own replacements, so silence is the rational choice. His advice to leaders is to reduce that fear directly: guarantee that nobody is laid off as a result of AI use, and reward the people who share what they built.

Assurance has to be explicit:

  1. “Nobody here loses their position for showing how they use AI.” Put it in writing.
  2. “The time you save goes back into the science.”
  3. “If you build it, you get the credit.” Count the workflow in the annual review and the promotion file.

A PI cannot promise what a hospital or a dean controls. Promise what you do control, and carry the same request up the chain.

Leadership carries real weight while this terrain is still uncertain. Disclosure is the working part. It is how we keep the boundaries visible, how we learn from each other, and how we move the science forward with AI.

5. What journals require: three principles and a location

Set the stigma aside and read what publishers and standards bodies have written. For manuscripts, most policies reduce to 3 principles:

  1. AI cannot be an author.
  2. Meaningful AI use should be disclosed.
  3. You remain accountable for every claim, citation, and interpretation.

Where the disclosure goes varies by journal. It usually lands in a dedicated AI declaration, the Acknowledgments (for writing help), the Methods (for research tasks), or some combination. ICMJE also wants it in your cover letter.

The stakes are stated plainly. The ICMJE section for authors says nondisclosure “may be construed as misconduct in some circumstances.”

Current journal policies (as of Nov 2026)

SourceWhat they requireWhere to disclose
ICMJE (now a standalone Section V)Journals should require disclosure at submission. AI cannot be an author. AI output cannot be cited as a primary source. You answer for plagiarism in AI-produced text and images.Cover letter and manuscript. Writing help in the Acknowledgments. Research use in the Methods, with the tool, version, and prompts where applicable.
Elsevier (updated June 2026)A separate AI declaration, published with the article, that names the tool, the purpose, and your oversight. Basic grammar, spelling, and punctuation checks are exempt. Substantive changes to sentence structure or organization must be disclosed.A section titled “Declaration of generative AI and AI-assisted technologies in the manuscript preparation process.” Research use in the Methods. Explanatory images in the caption and the declaration.
Springer Nature and Nature Portfolio (2026)Three risk tiers. Green: language editing, readability, formatting, translation. Amber: outlines, drafting sections from your own input, reference management, visuals built from verifiable source data. Red is prohibited: AI authorship, AI “writing results or discussion independently,” AI-generated claims or interpretations, and data or figures presented as empirical. Green and Amber uses both get declared.An AI declaration in the manuscript. Visuals also in the caption or legend.
Cambridge University PressAI use must be declared and clearly explained. AI cannot be an author. Authors are accountable for accuracy and originality.No single fixed location. Follow the journal’s instructions. Data collection or analysis belongs in the Methods.
WileyIf AI helped develop any portion of the manuscript, describe it specifically. Tools for spelling, grammar, and general editing are excluded.Methods, Acknowledgements, or a disclosure statement, depending on the journal.

3 things to highlight:

  • Elsevier draws its line at structure. A grammar check needs no statement. Restructuring sentences or sections does.
  • Springer Nature declares everything in Green and Amber, including language editing. Drafting a section from your own outline and notes is Amber. An AI writing your results or discussion on its own is Red.
  • ICMJE asks for prompts. For AI used in the study, the Methods should carry enough detail to replicate the approach.

AI-generated images

This is a big change. AI-generated images were strictly prohibited until earlier this year.

  • Elsevier: AI may help create explanatory images such as flow charts, timelines, and workflow diagrams, disclosed in the caption and in the AI declaration. Data visualizations qualify only when they are derived from the underlying data by a reproducible method described in the Methods. Primary research images (microscopy, blots, radiology, patient images) are off limits. General-purpose AI image tools are prohibited for graphical abstracts.
  • Springer Nature: AI visuals are allowed when the use is transparent and the output is tied to source data or material a reader can inspect. Disclose in the AI declaration and in the caption or legend. Nature Portfolio lists photorealistic deepfakes as not permitted.
  • ICMJE: You are responsible for plagiarism in AI-produced images too.

For data figures, assume no. For explanatory figures, check the journal first. I expected these rules to loosen as image models improved. It is starting to happen.

Always follow the target journal’s instructions. They override everything here, and they are changing fast.

6. What funders require: originality first

For grants, the weight falls on originality. Funders want to know you did not outsource your ideas, and they hold you fully responsible for what you submit.

Some funders specify where the disclosure goes. Others state the principle and stop there: AI use cannot compromise originality, confidentiality, or accuracy.

Funder policies side by side

FunderWhat they requireWhere and how to disclose
NIH (NOT-OD-25-132)Applications, or sections, “substantially developed by AI” are not treated as your original ideas. AI use detected after award can be referred to the Office of Research Integrity, with enforcement that ranges from disallowed costs to termination. Each PI is capped at six applications per calendar year. In force for receipt dates from September 25, 2025.No required disclosure format. The test is originality.
NSFDisclosure is encouraged and voluntary. You answer for the accuracy and authenticity of the proposal, including anything AI helped with.In the Project Description: whether you used generative AI, how much, and how.
DOE Office of Science (FY 2027 NOFO)Disclose any AI tool used in the application unless it was used solely to edit your own original draft. A Project Narrative that is not the researchers’ own work, including a machine-generated one, may be treated as academic dishonesty or research misconduct.The rule sits in the Project Narrative instructions. No form is named, so confirm placement with your program manager.
NASA Science (SMD)AI is permitted. Unacknowledged content that the team did not create is discouraged. You are responsible for accuracy and authenticity.A citation in the References section with the program name, version, date and time of use, and a statement of how it was used.
NEHAI is permitted in proposal preparation. You must flag any AI-generated text you insert. NEH can declare a non-compliant application ineligible.Footnotes or marginal notes at the point where the text is inserted.

The direction is that of more transparency.

7. The hard line: confidential peer review

This is where I see people step into real risk by accident.

Grant applications and manuscripts under review contain confidential, proprietary, and unpublished information. The rules here are the strictest of all.

  • NIH prohibits peer reviewers from using generative AI to analyze applications or formulate critiques, and from uploading application content to online AI tools. A separate NIH notice lists the consequences of confidentiality breaches, up to referral to the HHS Office of Inspector General and the Department of Justice.
  • NSF prohibits reviewers from uploading proposal content, review information, or related records to non-approved generative AI tools.
  • Elsevier allows reviewers limited support, such as language help, with private tools and with disclosure in the report. Uploading the manuscript, or any part of it, is ruled out.
  • Nature Portfolio lists delegating peer review to an LLM as not permitted and asks reviewers to disclose any AI use in their report.
  • NEH prohibits reviewers from uploading applications to third-party cloud databases.
  • ICMJE says manuscripts should not go into AI systems where confidentiality cannot be assured, unless the authors have given permission.

Reviewing your own work is different. Using AI to criticize your own study, to point out flaws, stress-test arguments, and find weak spots, is one of the best uses I know.

I would never do it on a free chatbot plan. Free means you pay with your data. Use a paid or institutional subscription and turn off the setting that lets your chats train the model.

The safest option is an open-weight model such as Llama or DeepSeek running locally on your own device. DO NOT use their website or app – that does not count as local.

8. How to write the disclosure

Every AI disclosure should answer 3 important questions:

  1. What tool? Name, version, and date of use (for example, Claude Opus 4.8, June 2026).
  2. What did it do? The specific task, in the specific section.
  3. What did you do? Who reviewed it, what was verified, and who is accountable.

Then put it where it belongs:

  • Writing help (editing, drafting, summarizing) → Acknowledgments or the journal’s AI declaration
  • Research use (screening, extraction, analysis, code) → Methods, with enough detail to replicate
  • AI-assisted figures → figure caption plus the AI declaration
  • All of it → cover letter, for journals that follow ICMJE

Fill-in templates

A. Light editing (Acknowledgments)

“During the preparation of this manuscript, the authors used [TOOL, VERSION (MONTH YEAR)] to improve the grammar and readability of author-written text. The authors reviewed every change and take full responsibility for the content.”

Elsevier requires no statement for basic grammar, spelling, and punctuation checks. For ICMJE and Springer Nature journals, include it.

B. Drafting help (Acknowledgments or AI declaration)

“[TOOL, VERSION (MONTH YEAR)] was used to [restructure / produce a first draft of / summarize the literature for] the [SECTION] from the authors’ outline and notes. The authors rewrote the text, checked every reference against its original source, and take full responsibility for the final manuscript.”

The phrase “from the authors’ outline and notes” carries weight. At Springer Nature it separates Amber from Red.

C. Code (Methods)

“Analysis code in [R / Python, VERSION] was drafted with help from [TOOL, VERSION (MONTH YEAR)]. [INITIALS] reviewed every line, tested it against [manual calculations / a reference dataset], and verified all results. Prompts and final code are available in [SUPPLEMENTARY FILE / REPOSITORY].”

Elsevier published its own declaration template. It follows the same principles: the tool, the reason, and a sentence confirming that the authors reviewed and edited the output and take full responsibility for the article. Copy the exact wording from their page when submitting to an Elsevier journal.

Two worked examples

Systematic review:

“GPT-6 Astra (ChatGPT, September 2026) assisted in refining the search strategy. Covidence and Rayyan supported abstract screening, and [EXTRACTION TOOL, VERSION] prepopulated structured data fields. PK and DP manually reviewed and verified every AI-assisted step. R code drafted with AI aided the statistical analysis, and RP verified all results for accuracy and reproducibility.”

Manuscript writing:

“During the preparation of this manuscript, the authors used Claude Opus 4.8 (June 2026) to refine the structure, improve clarity, and strengthen the coherence of the argument. The same model was used to summarize literature, and the authors critically reviewed, revised, and expanded those summaries. All factual claims and references were independently verified, and the final manuscript reflects the intellectual contribution and judgment of the authors.”

Both statements do 3 things:

  • Transparency. They name the tool and the task.
  • Accountability. AI assisted. People decided.
  • Human oversight. A named person checked each critical step.

A good AI disclosure statement does all three.

5 mistakes to avoid

  1. “AI was used to prepare this manuscript.” No tool, no task, no check.
  2. Disclosing in the cover letter only. ICMJE wants it in the manuscript too.
  3. No version or date. Models change every few months.
  4. Claiming responsibility without saying what you verified.
  5. Citing AI output as a primary source. ICMJE says that is unacceptable.

Pro tip: Log the tool, the version, and the task on the day you use it. You will not reconstruct it six months later at submission. Barely a third of researchers in the OUP survey keep that record.

Start with one sentence in your own lab

If you lead a group, tell them this week which AI tools you use and how you check the output. Then ask them the same.

If you are a trainee or a staff member, start your AI log today and bring one workflow to your next lab meeting.

I would love to know how this looks where you work. Does your lab talk about AI use openly, or does everyone find out when something breaks? I read every message (no bots involved :).

Top Papers on AI in research this week:

  1. Neurology LLM Scoping Review – Beth Israel Deaconess and Mount Sinai researchers mapped 32 studies of LLMs in clinical neurology. Models did well on narrow tasks like information extraction (F1 0.89 to 0.90). Open-ended case reasoning dropped to 42% to 54% accuracy. Almost none of the studies (31 of 32) had external validation.
  2. Have LLMs Changed Science Yet? – Bekkerman and colleagues analyzed 2.5 million full-text papers from five disciplines, spanning 2015 to 2025. Most pre-2023 trends simply continued, with limited detectable breaks. Some word frequencies shifted. The authors position the work as a baseline for tracking science as LLM adoption deepens.
  3. Novel Research Sounds Old – Frontier LLMs tend to make open research questions sound partly studied. Asked to check full texts separately, they catch most of these overstatements. In one test, a model judged 98% of 40 questions unstudied but said so plainly in just 8% of ordinary answers. A fixed-answer pipeline with separate citation checks lifted that figure to 95%.
  4. Hallucinated Authorship – An EMNLP 2026 paper tested three LLMs on authorship questions for 1.92 million papers across eight disciplines. Hallucination rates topped 98% for rarely cited papers. They fell below 37% for highly cited ones. The authors link most failures to retrieval limits rather than missing knowledge.
  5. AI Prompts in Journal Papers – A Taylor & Francis biomechanics paper published last month contained text telling AI reviewers to write a positive review. Scholars called it a prompt injection attempt to game peer review. The publisher has opened an investigation.
  6. EASD Rules for AI Evidence Synthesis – The European Association for the Study of Diabetes published a framework for using AI in guideline evidence syntheses. It builds on the RAISE guidance. Requirements cover transparency, preplanning, credibility, accountability, and evaluation. Human supervision stays central.

Top Papers on AI in education this week:

  1. AI Patients vs Human Patients – In a randomized crossover trial at Chiba University, 48 medical students interviewed a GPT-4o patient and a human standardized patient. Human patients produced higher Mini-CEX scores (23.9 vs 22.7). Satisfaction and rapport were similar. Students reported more active engagement with the AI patient.
  2. AI Appraisal Workshop for Residents – Mayo Clinic ran a 30-minute workshop teaching internal medicine residents to critique LLM output. Confidence in appraising AI rose from 60.4% to 79.5%. Concern about inaccuracy and bias rose too. The surveys were anonymous and unpaired, so the comparison is descriptive only.
  3. LLMs Grade Simulation Debriefings – A Montreal team had ChatGPT assess 30 anesthesia resident debriefings using the DASH framework. It produced a complete structured evaluation every time. “Structures the debriefing” scored lowest, with a median of 4 out of 7. No human expert comparison was done, so validity remains untested.
  4. Scoring Surgical Feedback with LLMs – University of Florida researchers fine-tuned models to rate 1,509 narrative comments from surgical resident assessments. Encoder models reached 0.93 accuracy on evidence scoring, while generative models reached 0.77. A simplified rubric cut rater disagreement from 27.8% to 8.5%.
  5. Mining Residency Advice with an LLM – LSU researchers used Claude Opus 5 to extract 8,069 recommendations from 74 plastic surgery application advice sessions. About 70.5% were actionable. Advice assuming access to mentors or money outnumbered alternatives 6.8 to 1. Human coders found lower actionability (48.6% to 63.4%), so treat the model figure with care.
  6. Why AI Tutors Underdeliver – The Syllabus covers a randomized trial of Khanmigo in 18 Tennessee middle schools. Gains were about 1.3 percentile ranks per term, similar to Khan Academy practice alone. Nearly all students messaged the tutor, yet only about 15% of messages contained real math reasoning. The same issue cites a Stanford review finding that students who let AI do the work tend to do worse once it is removed.

Leave a Comment

Your email address will not be published. Required fields are marked *

Related Posts

Join the ONLY NEWSLETTER You Need to Publish High-Impact Clinical Research Papers & Elevate Your Academic Career

I share proven systems for publishing high-impact clinical research using AI and open-access tools every Friday.