On 23 July, the Associated Press published new standards for the use of artificial intelligence in its newsroom. AI will be able to assist journalists with preliminary research, document summarisation, transcription, translation, headline and summary suggestions, shotlist preparation, language editing, search engine optimisation, and software development. Every output, AP specifies, must be reviewed and edited by a journalist before publication.
The list makes clear where AI may enter the journalistic workflow. It leaves open a harder question: when readers need to be told.
AP says it has introduced disclosure standards for cases in which generative artificial intelligence plays a "material role in published content." The announcement, however, contains no operational definition of that materiality. It does not explain whether an AI-suggested headline, summary, translation, or documentary research must always be disclosed, never disclosed, or disclosed only in certain circumstances.
Not all automation affects the final product in the same way. A spell-checker may alter the meaning of an article very little. A translation can change its wording, tone, and precision. A headline selects the angle presented to the reader; a summary decides which elements should appear essential. A shotlist organises the visual representation of a story.
It is therefore plausible to read AP's chosen threshold as a distinction between simple processing tools and interventions capable of meaningfully changing what the public receives. That is an interpretation, though — not a definition supplied by the agency.
The most useful precedent dates to May 2024, when AP had launched three experiments: translation of English-language articles into Spanish, creation of summaries, and headline suggestions. All outputs would be reviewed by journalists. Only for translations, however, did the agency explicitly announce it would indicate the use of the technology, noting that translation can represent a considerable change for a story. No equivalent public commitment was made for summaries and headlines.
That precedent does not establish a hierarchy valid in every situation, but it makes the editorial problem visible. What matters is not simply that software entered the workflow — it is how much it changed what the public reads, hears, or sees.
The new policy also marks a shift from the rules published in August 2023. At that point AP defined any AI-generated output as unverified source material and allowed employees to experiment cautiously, but not to use generative tools to produce publishable content.
By 2024 a handful of bounded experiments had arrived. By 2026 AI assistance enters explicitly into a larger portion of ordinary work, provided that final control, verification, and accountability remain human. AP summarises the principle by stating that AI does not replace reporting, source-gathering, editorial judgement, or fact-checking.
The ban on images also requires precision. AP continues to prohibit the use of generative AI to create, alter, or enhance journalistic photographs. This is not, however, an absolute ban on every synthetic image. Already in 2023 the agency acknowledged that an AI-generated illustration or artwork could be shown when it was itself the subject of the news, provided it was clearly identified and placed in the proper context. The distinction is between manipulating a record of reality and displaying a synthetic object in order to report on it.
What remains unclear is who determines when AI has played a material role. The announcement reaffirms that editorial judgement, verification, and accountability rest with journalists, but does not specify whether the disclosure decision falls to the author, an editor, the Standards department, product teams, or some other internal structure.
Nor is it clarified whether the threshold depends on the volume of text generated, the function performed, the risk of error, or the concrete effect on the final product. All that can be said is that AP has adopted internal standards and that the operational criteria are not contained in the public announcement.
The problem is not AP's alone. In an analysis published by the agency itself in February 2026, Alex Mahadevan of the Poynter Institute observed that generically disclosing any use of artificial intelligence in the newsgathering process could even become a disservice: AI may appear at many stages, from research to publication, without always carrying the same editorial weight.
The solution is not necessarily to attach an undifferentiated label to every technical operation. Poynter proposes that a useful disclosure explain precisely how AI was used and how a person verified the absence of factual errors or distortions. Transparency, from this perspective, consists not merely in acknowledging the presence of a technology, but in making the division of labour between system and journalist comprehensible.
Meanwhile, a study of 186,000 articles published in the summer of 2025 by approximately 1,500 US newspapers suggests that AI use is far more widespread than the disclosures attached to it. In a manual review of one hundred articles classified by the researchers' system as likely AI-generated or AI-assisted, only five contained a disclosure. The figure should be treated with caution: the classification relies on an automated detector and does not prove with certainty the origin of each text. It does, however, show the gap that can open between the adoption of tools and transparency toward the public.
The story, then, does not end with the expansion of permitted uses. AP has specified where AI may be employed, but has not made equally visible the criterion by which it will decide to inform the public.
We do not know whether the disclosure promised in 2024 for translations remains automatic or has been folded into the new materiality threshold. We do not know where any notice will appear: alongside the article, in a methodological note, in metadata, or only in products distributed to clients. The rules for coding assistants also remain to be clarified — tools that can affect how a story is built or presented without touching its words directly.
This does not mean AP lacks more detailed criteria. It means those criteria are not available in the public announcement.
The editorial principle that emerges from this case — and that KANSEI puts forward without attributing it to AP as an already-established rule — is to disclose AI use when it no longer functions as a simple processing tool but substantially affects the content the public receives.
That is the direction suggested by the phrase "material role." We do not yet know, however, how AP will translate it into practice, who will be responsible for enforcing it, or which interventions will fall inside or outside the disclosure obligation.
Today it is relatively easy to list every point at which artificial intelligence can enter a newsroom. Harder is deciding when readers have a right to know.
