How do you disclose that you used AI in your work?

A good AI disclosure answers three questions in one or two sentences: which tool, for what, and who is responsible for the result. "AI was used" answers none of them. "I used ChatGPT to suggest a structure for this report; I wrote, checked and edited all of the text" answers all three, and takes ten seconds to read.

Disclosure is not a confession. It is the same courtesy as crediting a photographer or a co-author: it tells the reader how the work was made, so they can judge it. It also protects you. If a reader later spots the hallmark of an AI tool, a clear note written up front reads very differently from a discovery.

The three parts of a useful disclosure

Be specific about the role. "Help with writing" can mean anything from fixing commas to producing the whole draft, and readers will assume the worst version.

Worked disclosure lines

Here are sample lines for common situations. Adjust the tool names and tasks to match what you actually did.

A blog post or newsletter, at the end of the article:

How this was made: I used Claude to suggest an outline and to tighten
two paragraphs. The research, the final wording and any errors are mine.

A research paper, in the methods or acknowledgements section. This follows the rule Nature and all Springer Nature journals set in January 2023: no AI tool can be listed as an author, because authorship carries accountability that a tool cannot take, and any use of a language model must be documented in the methods or acknowledgements, or in another suitable section if those do not exist.

We used ChatGPT to improve the grammar and clarity of the draft
manuscript. The authors reviewed and edited all text and take full
responsibility for the content.

An image, in the caption or alt text area, where the viewer sees it next to the picture:

Illustration generated with Midjourney from the author's prompt,
then cropped and colour-corrected.

Code, in a comment above the relevant block and in the commit message, so the next developer knows where to look carefully:

# Initial version of this parser drafted with GitHub Copilot;
# reviewed, refactored and covered by tests in test_parser.py.

A video, through the platform's own tool when it has one, plus a line in the description if you want to explain more.

Where to put it

Put the disclosure where the audience meets the work, not on a separate page they will never open. A caption beats an "about" page for images; a methods section beats a footnote for research. Keep the same wording across similar pieces so readers learn to recognize it, and record the details for yourself, including which tool and version you used, in case anyone asks later.

When is disclosure required?

Requirements depend on where the work goes. Three real examples show how different the rules are.

Scholarly publishing. Nature's January 2023 editorial set two principles for all Springer Nature journals: language models cannot be credited as authors, and researchers must document their use. The editorial expected other publishers to take a similar stance, and rules differ in detail, so check the author guidelines of the specific journal, and your institution's policy, before you submit.

Video platforms. Since March 2024, YouTube has required creators to flag realistic content that was altered or synthetically made, generative AI included. "Realistic" here means a viewer might take it for a real person, place, event or scene: a face swapped onto someone else, a synthetic version of a real person's voice, altered footage of a real event, or a realistic scene of a fictional disaster. Creators do not have to disclose clearly unrealistic content such as animation, ordinary filters and effects, or AI used for productivity, such as drafting a script or generating captions. YouTube shows a label in the expanded description, and a more prominent one on the video itself for sensitive topics like health, news, elections and finance. It has said it may add a label itself when content could mislead.

Copyright registration in the United States. The US Copyright Office's guidance, in effect since March 16, 2023, says copyright protects only human authorship. When you register a work that contains more than a trivial amount of AI-generated material, you must say so in the application and exclude that material from your claim, while claiming the parts you wrote, selected or arranged yourself. Editing your own work with ordinary tools such as Photoshop does not take away your authorship.

Outside rules like these, disclosure is usually a matter of policy rather than law: your employer's, your client's, your school's or your publisher's. When no rule applies, a simple test helps. Would the person relying on your work feel misled if they learned how it was made? If the answer is yes, or even maybe, disclose.