AI-Assisted Research Ethics: A Practical Guide

AI tools are now part of many research workflows, and the published guidance on them has converged on a small number of requirements. This guide covers what those requirements are, where they are written down, and what they mean in practice.

Where the Rules Actually Live

There is no single global rule. Three layers apply, in this order:

1. Your institution's policy. Disclosure requirements are set locally, and they are the ones that govern your thesis or coursework. Read yours before you rely on any tool.
2. The journal's or publisher's policy. These are stated in the author guidelines, and they vary.
3. The cross-publisher recommendations, of which the most explicit are the ICMJE recommendations on the use of AI by authors, written for medical journals but widely cited beyond them.

What the ICMJE Recommendations Say

The ICMJE section on AI use by authors sets out four things (quoted here; read the section in full at the link above):

• Disclose at submission. "The journal should require authors to disclose at submission whether they used AI-assisted technologies (such as LLMs, chatbots, or image creators) in the production of submitted work." Authors who use such technology "should describe, in both the cover letter and the submitted work in the appropriate section if applicable, how they used it."
• A tool is not an author. "Chatbots (such as ChatGPT) and other AI-assisted tools should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work, and these responsibilities are required for authorship." It follows that "humans are responsible for any submitted material that included the use of AI-assisted technologies."
• Review and edit the output. "Authors should carefully review and edit the AI-generated content as the output can be incorrect, incomplete, or biased." Authors "should be able to assert that there is no plagiarism in their paper, including in text and images produced by the AI", and the ICMJE puts the next duty on people rather than on authors specifically: "Humans must ensure there is appropriate attribution of all quoted material, including full citations."
• Do not cite the tool as the source of a fact. "Referencing AI-generated material as the primary source is not acceptable."

And on the consequences: "Nondisclosure of AI use may require corrective action and may be construed as misconduct in some circumstances."

Core Principles

Transparency and Disclosure

State which tools you used, how, and for what purpose. This is what lets a reviewer evaluate your methodology rather than guess at it. The wording is set by your institution; the habit that makes it easy is keeping a log as you work.

Accuracy and Verification

Language models are probabilistic: they predict likely text, and the output can be wrong in ways that read as confident. Treat an AI-generated summary or reference as a suggestion to check, not as a finding.

For a reference, that means resolving the DOI or finding the record in a database, then reading the paper. For a statistic, it means finding the number in the source and confirming the population and year it describes.

Fair Attribution

Using a tool to tighten your prose is generally accepted; letting it produce your argument is not. If a tool shaped a hypothesis or structured a section, disclose that according to your institution's rules. The analysis has to be yours, because you are the one answerable for it.

Best Practices

Document Your AI Usage

Keep a log, for reproducibility and for your own protection:

• Prompts: what exactly you asked
• Tools and versions: which model, which product, which release
• Date of access: models change, and results change with them
• Extent of editing: how much of the output survived

Mitigate Bias

Models trained on web-scale text carry the biases of that text, and a discovery tool's ranking is not neutral either. If you rely on one tool for literature discovery, you can systematically miss non-English work and research from under-represented regions. Run your search in more than one place, and check whether whole literatures are absent.

Protect Participant Data

Do not put identifiable participant data into a public model. Whether a given service trains on submitted content depends on the product and the plan, and it changes — read the current terms for the specific tool, and default to anonymising before you paste anything.

Common Mistakes

Over-Reliance Without Validation

The recurring failures are accepting wrong dates, names or figures; citing papers that do not exist; and using stale information on a fast-moving topic. If you cannot find the primary source for a claim a tool made, do not use the claim.

Not Disclosing

Failing to disclose AI use can be treated as misconduct — the ICMJE says so directly for journal submissions, and institutional codes generally follow. Disclosure costs a sentence.

Treating a Summary as a Reading

A summary is a triage tool. A citation is a claim that the source says what you say it says, and only reading the source supports that claim.

How to Structure an Ethical Workflow

The goal is augmentation, not automation. Use tools for the mechanical parts — an initial sweep of the literature, formatting references — so your time goes on synthesis and judgement.

Treat a tool as an assistant whose work you check, not as a co-author:

• You make the decisions
• You verify every claim
• You are answerable for the output

Conclusion

Ethical AI use in research is mostly procedural: disclose, verify, attribute, protect data. None of it is onerous, and all of it is written down somewhere you can cite.

Sources

• International Committee of Medical Journal Editors. Use of AI by authors (Recommendations, Section V.A). https://www.icmje.org/recommendations/browse/artificial-intelligence/ai-use-by-authors.html (read 22 September 2026)