Disclose it when AI wrote most of the release and no person meaningfully edited it. You do not need to flag a grammar pass or a round of AI-assisted copyediting. That line, full drafting versus light editing, is roughly where the newswires' own guidance lands, and as of August 2026 it is also roughly where EU law draws it for content aimed at European audiences. Almost nobody is actually drawing that line in practice. Pangram Labs scanned 178,672 PR Newswire releases published between August 2025 and July 2026 and found that the share containing AI-generated text rose from 22% to 48% over that year, with no public tag on any of them telling a reader which was which.
Source: Pangram Labs' analysis of 178,672 PR Newswire releases, linked in the article.
How much of what goes out on the wires is actually AI-written now?
Enough that treating a press release as a human-written document is no longer a safe assumption. Pangram's detector flagged 61,523 of the 178,672 releases as containing AI-generated text, roughly a third of the sample. Among those flagged releases, the median was 52% AI-authored, and 15,205 of them, 25% of the flagged set, were fully AI-generated start to finish. The rate was not even across industries. Advertising releases ran hottest, at 64% containing AI text and 41% fully AI-generated, with one single advertising issuer in the dataset publishing 2,265 releases that were 97% fully AI-written. Finance, energy and biotechnology sat at the other end, with fully AI-generated shares under 5%, which tracks with how much those industries have to lose if a generated number turns out wrong.
The releases also got longer over the same period. Median length rose from 488 words to 562 words, about 15%, which is the opposite of what you might expect if AI were mainly being used to compress or simplify. More AI use correlated with more padding, not less.
Does any newswire actually require disclosure?
No. PR Newswire itself sells an AI drafting tool and does not publicly tag which releases used it, and nothing in the standard distribution agreements with the major wires requires a sender to flag AI involvement. The wires are distribution infrastructure, not editorial gatekeepers, and they have left the disclosure decision entirely to whoever is sending the release.
That gap between "nobody requires it" and "nobody does it" is worth sitting with for a second, because the wires' own published advice disagrees with the silence their own data shows.
What does the guidance actually say you should disclose?
PR Newswire's own published guidance on disclosing AI draws a specific line. Small, mechanical edits, fixing spelling, grammar or syntax, do not need to be flagged. Neither does using AI purely for internal brainstorming or research that never makes it into the public text. What does need a disclosure, by the same guidance, is a release that was substantially or fully AI-drafted, or one built on AI-driven data analysis that shaped what the release says. The suggested approach is a single sentence, placed near the end of the release before the boilerplate, something to the effect of noting that an AI tool assisted in drafting and that a human reviewed and fact-checked the result. You are not expected to name the specific AI platform or version.
That guidance exists. The 48% figure says almost nobody is following it.
Is there now an actual legal requirement to disclose?
For content reaching the European Union, yes, in a specific and narrower way than a lot of the coverage suggests. Article 50 of the EU AI Act, which took effect on August 2, 2026, requires that anyone deploying an AI system to generate text on matters of public interest disclose that the content was artificially generated or manipulated. There is a specific carve-out: disclosure is not required if the text underwent substantive human editorial review, with a named person taking editorial responsibility for it. In practice, that exemption is wide enough to cover most legitimate press release workflows, a draft generated with AI assistance and then genuinely reviewed by a person who stands behind it does not trigger the rule. What it does not cover is a release generated and sent with no real review, which under the EU framework is no longer just a quality problem. It can carry fines of up to 15 million euros or 3% of worldwide annual turnover, whichever is higher, for companies operating in scope of the regulation.
Outside the EU, there is no equivalent statute yet. The newswires' own guidance is the only standard most companies are working against, which is exactly why the 48% figure matters: it is a measure of how many companies are ignoring a voluntary standard, not just an unenforced legal one.
What actually happens if you skip disclosure and someone notices?
The practical risk has changed, not just the ethical one. AI-text detectors like the one Pangram used to produce its dataset are accurate enough now that the tool caught just one false positive among 21,625 pre-ChatGPT releases it tested, a false positive rate of 0.0046%. That level of accuracy means an undisclosed, fully AI-written release is checkable by any journalist, competitor or researcher who wants to look, the same way Pangram's own researchers checked 178,672 of them for a public report. A release that reads as AI-generated and was not disclosed as such reads, to anyone who checks, as a company that either did not review its own public statement or chose not to say so. Neither is the impression you want attached to a document meant to be a factual, citable record.
There is also a quieter cost. Generative engines and AI search tools weigh earned coverage far more heavily than a brand's own self-published text when deciding what to cite. A press release padded by AI into a longer, denser document does not become more citable for the padding. If anything, it looks more like the kind of interchangeable wire content these systems already discount.
So what should a company actually do?
Use the same line the newswires already drew, and actually follow it. If AI helped draft the release and a real person then edited, fact-checked and takes responsibility for the final version, you do not need a disclosure line, that is standard tool use, not deception. If AI wrote most or all of the release with no meaningful human pass, either get a person to genuinely review it before it goes out, which is the better fix, or add the one-sentence disclosure the wires already suggest. A clean, factual release still only has one real job: being the accurate, dated record a reporter, a regulator or an AI system can check against. A disclosure sentence does not weaken that. An undisclosed AI release that nobody reviewed, caught after the fact, does.
The comparison to look at, if your company already handles disclosure for other kinds of promotional content, is how the FTC treats sponsored content labeling: the standard there is also that a reasonable reader should not be misled about what they are looking at. AI disclosure in a press release is the same question asked about authorship instead of payment.
The direct answer, for anyone who read this far
Disclose AI involvement in a press release when it did most or all of the drafting and no person substantively reviewed the result before it went out. You do not need to disclose routine AI-assisted editing, and you do not need to name the tool. Nearly half of PR Newswire's releases now contain AI-generated text with zero public disclosure, which is a bigger gap between practice and the newswires' own published standard than most companies realize exists. If your own process already includes a real human review before release, you are almost certainly already compliant with both the voluntary guidance and the EU's narrower legal requirement, and the only thing left to decide is whether White Ocean Media or your own team handles the distribution once the review is done. If you are not sure where your process actually stands, talk to us.
