To disclose AI use to readers, tell them three things in plain language: which tool you used, what job you gave it, and how much of what they are reading it produced. Put that note where the content is, not buried on a policy page. That is the whole mechanism, and a small newsroom can run it in an afternoon.
Most newsroom AI policies handle the obvious cases well and stall on the grey ones. Was grammar cleanup an AI use? What about a headline you picked from ten machine-written options? Does a translation tool count if a fluent human reviewed every line?
The honest answer is that nobody has a settled taxonomy. What you can do is write down your own thresholds before you need them, then apply the same thresholds every time. Below is the process, the wording, and the awkward case of finding undisclosed AI work after it is already published.
Table of Contents
- What You Need
- Step-by-Step: How to Disclose AI Use to Readers
- Common Mistakes
- Frequently Asked Questions
- Does a news outlet have to tell readers when AI was used?
- Do I need to disclose AI if I only fixed grammar or spelling?
- What should be included when disclosing AI use in work?
- Where should an AI disclosure go in an article?
- Will disclosing AI use hurt engagement and readership?
- What do I do if I find undisclosed AI use after publication?
- Conclusion
What You Need
You need six things before the first notice goes live. None of them require a software purchase; all of them require somebody to write them down and put a name against them.
- A one-page editorial AI policy. Your own thresholds, in plain language, with the reasoning behind each one stated once.
- A short list of approved notice wordings. Three or four fixed sentences a writer can paste in, so nobody drafts disclosure language from scratch at 11pm.
- An owner. One named person — usually a standards editor or managing editor — who maintains the policy and signs off on tier assignments.
- An escalation contact readers can reach. A published address for questions about AI use in a specific piece, not just a general mailbox.
- A map of where notices appear. Byline note, story note under the headline, methods box, media credit, standards page. Write the list down so the same slot gets used every time.
- A use log. A field in the CMS, a line in the pitch doc, or a shared sheet: who used what tool, on what, at what tier. Without it you cannot answer a question about a story published eighteen months ago.
UCL’s guidance for academic work frames the same problem: acknowledge the tool, name the version, name the publisher, and say what you used it for in one sentence. That structure survives the jump from a thesis to a news story intact.
Step-by-Step: How to Disclose AI Use to Readers
Five steps, in the order they should happen. Most of the friction is at step one, and almost none of it is technical.
Identify What AI Contributed
Separate production assistance from material use. The working test: would a reasonable reader say the machine helped author, not just help operate, this piece? Brainstorming angles, summarising search results and tightening grammar are assistance. Words the reader sees that a model composed are authorship, however lightly you edited them afterwards.
Here is the grid most newsrooms end up with once they argue it out.
| Use case | Disclose? | Where the notice goes |
|---|---|---|
| Brainstorming angles, no model output in the story | No | Standards page only |
| Grammar and spelling pass on your own draft | No | Standards page only |
| Summarising documents for your own notes | No | Standards page only |
| Headline variants you choose between yourself | No | Standards page only |
| Interview transcription | Yes | Methods box and audio credit |
| Full translation of an interview or document | Yes | Byline note |
| Data analysis or chart generation | Yes | Figure caption and methods box |
| Synthetic or materially edited images | Yes | Caption on the image itself |
| Substantial passages of article text | Yes, prominently | Story note under the headline |
Notice the pattern in the top half of that table: if the reader would not see the difference, you do not owe them a note. Two cautions. First, transcription tools make mistakes, and a misheard name is a correction, so a human check on names, numbers and quotes is part of the disclosure, not an extra. Second, if you would be uncomfortable seeing the note on the story, that is a signal the tier is wrong — not that the note should be cut.
Choose the Disclosure Level
A single “AI was used” checkbox tells a reader and an editor nothing about how carefully to read the piece. Grade the involvement instead, by proportion of what the reader experiences.
- Tier 1, primarily human. Under roughly 20% of the output, or assistance that never reaches the page. A standards-page line covers it. No per-story note.
- Tier 2, mixed. Material AI output in a human frame — transcription, translation, data work, some graphics. Per-story notice, named editor, one extra review pass on facts and quotes.
- Tier 3, primarily AI. Around 80% or more of the output, or a one-shot generation with no substantial human edit, or an agentic run that gathered and drafted without a reporter shaping it. Visible notice in the first screen of text, explicit methods box, and a named human editor against it.
Tier 2 is the grey zone and it is where most editorial time goes. When you cannot tell, ask what a skeptical reader would need to know to trust the piece. If the answer is a paragraph, give them a paragraph.
Write a Clear Reader-Facing Notice
Specificity is what makes a notice read as honest. Vague boilerplate reads as a stunt or a cover-up, because it is indistinguishable from both. Write in plain language, name the tool and version, name the task, and state the human check that followed.
Story note, for text the model drafted or substantially shaped:
“This story was produced with the assistance of [tool name, version], which [generated a first draft / restructured our notes / drafted pull quotes]. All reporting, interviews, facts and quotes were checked by [editor’s name].”
Methods box, for data, transcription and translation:
“Interview audio was transcribed using [tool, version] and reviewed line by line against the tape. Names, titles and figures were corrected after review. Charts were generated in [tool] from [dataset]; the underlying data and calculations are available on request.”
Caption label, for images:
“Illustration generated with [tool] and prompted by [name]; edited by [name]. Not a photograph of the events described.”
Translation note:
“Translated from [language] with [tool, version] and reviewed by a native speaker at [publication]. Quotations were retranslated after review.”
When AI use changes after publication:
“Update, [date]: This story has been amended to disclose that [tool] was used for [task]. The text was re-reviewed and [what changed]. The original version did not disclose this use.”
On the standards page itself:
“We use AI tools for [list]. We disclose use in the story itself whenever a tool contributed text, translation, transcription, data analysis or imagery that a reader would encounter. Minor assistance — brainstorming, grammar, summarising documents for internal notes — is covered by this page rather than a per-story note. Questions: [address].”
IEEE’s author guidance goes further than any of these, asking for the specific sections that contain AI-generated content to be identified individually. That is workable in a feature of a few thousand words and unworkable in a rolling news file; most newsrooms settle for one notice that names the sections.
Place the Notice Where Readers Expect It
Proximity is the rule. A disclosure next to the content it describes beats a link to a global standards page every time, because most readers never follow the link.
- Text: a story note directly under the headline, or a byline note if the use touches the whole piece.
- Images and video: the caption, not the file name and not the footer.
- Audio and podcasts: the episode description, plus a spoken mention in the first ninety seconds for anything tier 2 or above.
- Data and interactives: a methods note inside the graphic, where the numbers are.
- Newsletters and social posts: a line at the end of the post for tier 2, and for tier 3 in the first line, because those get clipped and reposted without context.
The formats nobody covers are audio, video and social. Those are also the formats most likely to be stripped of their captions, which is exactly why the label has to live inside the asset.
Review and Update the Disclosure
Before publication, an editor checks five things: the tool and version are named, the task matches what the log says was done, the human check is stated and true, the notice sits in the agreed slot, and the wording matches the notices used on the last ten stories.
Consistency is doing more work than it looks. Audiences read the same label as a signal about the outlet, and a different phrase each week reads as improvisation.
Tool versions churn, so log the version at the time of use rather than linking to a product page that renames itself later. If your CMS can store it, store the whole tool name and version in a field. That habit is what makes an audit trail possible when a reader emails eight months later asking what produced a chart.
When you find undisclosed AI use in an already-published piece, work through this in order:
- Stop and verify. Confirm the use with the reporter before changing anything. Automated detectors produce false positives, and a wrong accusation is its own failure.
- Assess the tier. Ask whether a reader’s understanding of the piece changes once they know.
- Add the notice without removing the original. Annotate the disclosure date and describe what was used and for what. Readers value seeing the process.
- Correct anything factually wrong that came out of it, and say plainly that the correction exists because of the undisclosed use.
- Notify the desk, not just the author. If one person skipped the notice, the policy probably lacks a clear threshold, and the fix is the threshold.
- Log the incident and check whether other pieces by the same desk carry the same gap.
Common Mistakes
Vague labels. “AI was used in the production of this article” tells a reader nothing and signals nothing. Fix: name the tool, the task and the human check, in one or two sentences.
Burying the notice. A disclosure in a site-wide footer or behind a “learn more” link satisfies a policy and not a reader. Fix: put it in the first screen of the story, and keep the standards page as the backup rather than the primary.
Treating all AI use as identical. One checkbox collapses a grammar pass and a machine-written draft into the same category, so readers cannot calibrate their trust. Fix: tiers tied to proportion, with a different review bar and a different notice for each.
Never updating. A note written at publication can be wrong a month later when the reporter learns the tool also drafted a section. Fix: build the update note into the correction template and treat disclosure changes the same way you treat factual corrections.
Treating disclosure as accountability. A label is not a byline. A notice that names no human editor reads as the model doing the work. Fix: attach a named person to every tier 2 and tier 3 piece.
Assuming disclosure costs you readers. The pattern reported from audience research is the opposite of the usual fear: undisclosed AI use erodes trust, and audiences consistently favour a visible human hand over an efficient-looking one. Specificity is what buys credibility — a blanket disclaimer on every story teaches readers to ignore the label entirely, which is the worst outcome available to you.
One more practical note. Some platforms now estimate AI use for readers directly, and Substack began surfacing reader-facing detection estimates for longer posts in July. Platform-native signals change the calculus in a hurry: the day a reader can see an estimate next to your story, a quiet policy page is no longer enough.
Frequently Asked Questions
Does a news outlet have to tell readers when AI was used?
Often yes, though the trigger depends on where you publish. Academic and research publishers usually require disclosure for any AI-generated text. EU AI Act transparency obligations cover synthetic content and deepfakes, and a growing number of state synthetic-media laws and FTC enforcement actions make undisclosed use a legal risk rather than just an editorial one. Most newsroom standards bodies publish their own rules. Publish your own policy regardless; it is the fastest way to meet whichever standard applies.
Do I need to disclose AI if I only fixed grammar or spelling?
No, not per story. Grammar, spelling and style corrections on text you wrote are treated as assistance by most guidance, including UCL’s, along with brainstorming and summarising search results for your own notes. Cover them once on a standards page instead of cluttering every story. The line moves when the tool rewrote sentences rather than flagging them, or when you accepted a suggestion you would not have written yourself. Judge each case on whether a reader would consider the machine a co-author.
What should be included when disclosing AI use in work?
Five things: the name of the tool, its version, the publisher, what you used it for, and the extent of its contribution. A useful extension is the human check that followed — who reviewed the output and how. UCL requires the first four plus a one-sentence context line, and IEEE asks in addition that specific sections containing AI-generated content be identified. Log the version at the time of use, because product pages rename themselves later.
Where should an AI disclosure go in an article?
Next to the content it describes. For text that is a story note under the headline or a byline note. For images and video, the caption. For audio and podcasts, the episode description and a spoken mention early in the segment. For charts and interactives, a methods note inside the graphic. A link to a global standards page works as a backup, not as the main notice, because most readers never follow it.
Will disclosing AI use hurt engagement and readership?
The audience research runs opposite to the fear. Studies covered by the Newspaper Association of America found AI-generated journalism provokes distrust, with audiences favouring a visible human touch. Specific, proportionate disclosure lets readers judge the work on its merits instead of guessing. The engagement risk comes from over-disclosure — a generic banner on every story trains readers to skip the label, so keep notices proportional and tied to actual involvement.
What do I do if I find undisclosed AI use after publication?
Verify it with the reporter first, since automated detectors produce false positives. Then assign a tier, add a dated disclosure without deleting the original, and correct anything factually wrong the tool introduced. Tell the desk rather than only the author, because if one person missed the threshold the policy probably needs a clearer one. Log the incident and audit other work from the same desk before you close it out.
Conclusion
Start with an inventory, not a policy. List everywhere AI already touches your output, agree the thresholds in one meeting, publish the standards page in plain language, then test the wording on a real story before you need it again.
That last step is the one that changes things. Wording that has survived a live story survives anything easier.
