When to say you used AI and when it costs you

Disclose AI use where a person's trust in the source is the product, and stop disclosing where it is just a tool in a workflow nobody asked about.

The confusion comes from treating disclosure as one decision. It is two. There is what you tell a human reader or a journalist, which is a trust decision. And there is what your files tell a platform, which is not a decision at all any more, because the metadata travels whether you mention it or not.

The journalist number, and the contradiction inside it

Survey work across the PR and media industry puts it starkly: 78% of journalists say an AI-written pitch reduces their trust in the sender, while 54% of them use AI in their own workflow.

That is not hypocrisy. It is a precise signal about what the disclosure is for. A journalist using AI to condense a 40-page report is using a tool inside a process they control. A journalist receiving an AI-written pitch is being told, implicitly, that the sender did not care enough to do the thinking. The objection is not to the technology. It is to the outsourcing of the part that was supposed to be about them.

Apply that test to every disclosure question and it resolves quickly. Did AI touch the part where your judgement, your access or your relationship was the value? Then it matters, and hiding it is a risk. Did it touch the mechanical part? Then announcing it is noise, and slightly odd noise at that, like disclosing that you used spellcheck.

Where to disclose

Pitches and personal outreach. Never send an AI-drafted pitch as if it were written for that journalist. If you use AI in the process, use it on the research and write the pitch yourself. Do not disclose. Just do not send the thing they can spot.

Editorial and opinion under a named byline. If the byline is a person, the opinions must be theirs. A drafted-from-interview piece is fine and needs no label. A piece the named expert never spoke to is a trust liability, whether or not anyone finds out.

Anything with faces, voices or footage that could be mistaken for a record of something real. This is the hard line. Synthetic imagery presented as documentary is a different category of problem from AI-assisted prose, and it is the one that ends careers.

Regulated and safety-relevant content. Where a claim carries legal or physical consequences, note the review process rather than the drafting tool. Reviewed by a licensed engineer is the disclosure that matters.

Where not to

Ordinary marketing copy, drafted with assistance and edited by a person who knows the subject. Google's spam policies do not care what typed a page, only whether it exists to help or to manipulate. Ahrefs analysing 1.9 million URLs cited in AI Overviews found 87.8% of cited pages were mixed human and AI, versus 8.6% pure human. The mixed page is now the normal cited page. Labelling it tells the reader nothing useful and signals insecurity about your own work.

Internal documents, transcripts, summaries, first drafts, translations, alt text. Tools.

The part that is no longer your call

Content Credentials, the C2PA standard, embed provenance data in the file at export. Adobe, Microsoft, OpenAI, Google and others attach them. TikTok adopted C2PA in 2024 and reads those credentials on upload to auto-label AI-generated content, and it attaches invisible watermarking to content made with its own AI tools. Other platforms are moving the same way.

The practical consequences are worth spelling out for a marketing team.

Your exports carry provenance whether you intend it or not. If a designer generates a background in one tool, composites in another and exports, the credentials can survive the chain. Assume they do.

Platform labels are applied by the platform, not requested by you. A post can appear with an AI label you did not choose, on content you consider human-made with a minor generated element. Plan for that in creative review rather than being surprised by it in comments.

Stripping metadata is available and is a bad idea. It is trivially detectable in aggregate, it looks like concealment, and it removes your ability to prove provenance on the day you actually need to.

Check what your pipeline emits. Take one finished asset, run it through a Content Credentials inspector, and see what a platform sees. Most teams have never looked, and it takes ten minutes.

Write a one-page policy, not a philosophy

The teams that handle this badly are the ones still debating it. Decide, write it down, tell the team.

Four lines is enough. Named human bylines mean a human wrote or dictated the substance. No synthetic imagery of people, places or events presented as real, ever. Outreach to individuals is written by the person sending it. Everything else uses whatever tools do the job, edited by someone who knows the subject, with no label.

Then add the operational one that actually saves you: every claim in published content traces to a source your team can produce on request. That single rule does more for trust than any disclosure statement, because the failure people are really worried about is not that a machine wrote it. It is that nobody checked.

The cost of getting it wrong in each direction

Over-disclosing costs you credibility in a small way, repeatedly. A label on ordinary copy invites the reader to discount it, and they will.

Under-disclosing where it counts costs you a relationship, once, permanently. Journalists talk. So do clients.

Given the asymmetry, the rule of thumb is easy. If a person's trust in you personally is what the content trades on, do the work yourself. Everywhere else, use the tools and stop announcing it.