Telling Readers That AI Was Involved
9 min read · updated August 4, 2026
Disclosure is not an apology and it is not a legal shield. It exists so a reader can calibrate how much to trust a piece and know who is answerable for it. That purpose gives you a threshold: disclose when the tool affected what the piece says, not when it affected how quickly it was typed.
What disclosure is actually for
A reader assessing an article is asking two questions: is this likely to be right, and who is responsible if it is not. Everything a disclosure note should do serves one of those.
That rules out two common approaches immediately. A blanket footer on every page saying the site may use AI tools answers neither question, because it does not vary with the piece and therefore carries no information. And a disclosure used to lower the standard — publishing something unverified because a label says a machine helped — inverts the purpose. The label was supposed to help the reader assess the work, not excuse the publisher from doing it.
The rule that follows is uncomfortable and clarifying: accountability does not transfer. Whatever the disclosure says, the byline is answerable for every claim in the piece, which is why the disclosure decision belongs at the human gate and is recorded there.
The four positions publishers take
Published policies vary widely, but the ones that have been made public cluster into four positions. They are described here as families, without attribution to any particular organisation, because specific policies are revised often and repeating a stale one accurately is worse than describing the shape.
| Position | Description |
|---|---|
| Prohibition | No generative output in published copy at all. Tools may be permitted for research, transcription and internal drafting, and nothing generated reaches the page. Common where the product is explicitly original reported work, and usually paired with a rule that any exception needs editor sign-off. |
| Assistive use, undisclosed | Treats the model as a spellchecker: mechanical assistance that does not change what the piece says needs no more disclosure than word processing does. Disclosure begins above a stated threshold of substantive contribution. |
| Per-piece labelling | Any generative involvement is labelled on the individual item, with a short note describing what the tool did and who reviewed it. The strong version names the reviewing human, which is the part that carries information. |
| Category-level disclosure | Whole content types are declared — automated results summaries, translated editions, machine-generated illustration — with a standing explanation of the process, while everything else follows ordinary byline rules. Suits high-volume structured output where per-item labelling would be noise. |
Most working policies combine two of these: category-level disclosure for automated formats, per-piece labelling above a threshold, nothing below it. The disagreement between them is almost entirely about where the threshold sits, which is why the next section is the useful one.
The threshold that decides which applies
One question separates the cases: did the tool influence what the piece claims? Not how it reads — what it claims.
- Below the line, no disclosure. Spelling and grammar correction. Transcribing your own recorded interview. Reformatting a table. Suggesting search terms. Summarising a document you then read yourself. None of that changes the substance, and labelling it devalues the label.
- Above the line, disclose. Generated prose that survives into the published text. A structure or an argument adopted from a model. Any claim that originated in output rather than in a source. Generated images, audio or video presented as depiction. A translation published without a fluent human check.
- Always disclose, regardless of extent. Anything representing a person: a synthesised voice, a composite quotation, an altered image of a real event. A reader’s default assumption about a depiction is that it happened, and correcting that assumption is exactly what disclosure is for.
The middle category catches the case people miss. If you asked a model how something works and that shaped a paragraph, the model supplied a claim — even though you typed every word yourself. Either trace the claim back to a source and cite that, or disclose it. This is also the point at which the disclosure question stops being separable from the attribution question: an unsourced claim you did not originate is a problem whether or not you label where it came from.
A template to adapt
Two parts: a per-piece note, and the standing policy it points at. Replace everything in brackets, and delete any clause you cannot honestly assert.
PER-PIECE NOTE (end of article, above the author bio) How this piece was made: [a first draft of the section on X was generated with a language model from an outline and sources compiled by the author / interview audio was machine-transcribed and checked against the recording / the illustration is machine-generated]. Every factual claim, quotation and figure was verified against a primary source by [NAME], who is responsible for the published text. STANDING POLICY (linked from the note; one page, plain language) 1. What we use these tools for We use them for [transcription, translation drafts, structural editing, summarising documents we have read, illustration for X]. We do not use them for [reporting, sourcing, quotations, images presented as depictions of real events]. 2. What we always do ourselves Finding and reading sources. Verifying every name, number, date and quotation. Deciding what the piece claims. Deciding what to publish. 3. When we tell you We label any piece where generated text or media reached publication, and any piece where a claim originated from a model rather than from a source. We do not label spelling correction or formatting. 4. Who is responsible The named author and the named editor, in every case, whatever tools were involved. If we get something wrong we correct it here: [link to corrections page]. 5. When this policy changed Last revised [DATE]. Previous versions: [link].
Point four does the work. A reader who knows a named person stands behind the claims has what they needed; a reader told only that “AI-assisted tools were used” has been handed a disclaimer dressed as transparency. Point five matters more than it looks, because a policy with no revision history is useless to anyone deciding whether to trust a three-year-old article.
Where it goes and how it is worded
- End of piece for text, on the asset for media. A generated image needs its label in the caption, because the image travels without the article — screenshots, social cards, aggregators. Text does not have that problem to the same degree.
- Say what the tool did, not which tool it was. “The first draft of the timeline was generated and then checked against the filings” is information. A product name and version number is trivia that dates the page within months.
- Name the verifier. The highest-value word in any disclosure is a person’s name.
- Keep it in the voice of the piece. Legal register signals that the note exists to protect the publisher, which is precisely the reading you are trying to avoid.
- Machine-readable provenance is a supplement. Content credentials and similar provenance metadata attached to a file are useful to downstream systems and invisible to a reader. You still need a sentence they can read.
Four ways disclosure goes wrong
- The universal footer. Applied to every page, it varies with nothing and tells a reader nothing, while giving the publisher the feeling of having disclosed.
- Disclosure as a lowered bar. If the label makes it acceptable to publish something less checked, the label has made the product worse. The standard is unchanged; only the reader’s information has changed.
- Retrospective silence. Adopting a policy without saying from what date it applies leaves an archive that the policy implies was unassisted.
- Disclosing the tool and hiding the failure. A note saying a model was used, on a piece where a fabricated statistic survived to publication, reads as an excuse. The correction matters more than the label, and the properties that make work untrustworthy are not repaired by admitting to them.