content production.automatic labelling doesn't work: declaring it is our job.

On 4 September 2026 a test by The Verge showed Instagram's "AI Content" label getting it wrong both ways: barely retouched real photos, including edits made in the phone's own photo app, came out labelled, while material generated with everyday commercial tools went through unmarked.

The false negative has a technical explanation. Generated images usually travel with content credentials —the C2PA standard— or an invisible watermark such as SynthID, and in the test those signals did not trigger the label. The false positive is simpler: any computational edit leaves a trace that looks like generation, and the classifier does not separate retouching from synthesis.

The aggravating part is procedural, not technical: for a wrongly labelled post there is no published appeal mechanism. The badge appears, warns nobody and is not up for discussion.

There are two different labels and they are worth keeping apart. One sits on the post, and that is the one that misfires. The other, since 31 August 2026, sits on the profile: Instagram renamed its "AI creator" label to AI-generated profile and started limiting the reach of accounts that show a generated person without declaring it. The account drops out of Reels, Explore and suggestions.

The operational difference is the level: it applies to the entire account, not to the post. A synthetic avatar published from the brand account drags the whole account with it, including the content featuring real people. Which is why where an avatar lives —a declared brand account or a separate one— is decided before the first asset is produced, not when reach drops.

There an appeal does exist, through account status. But it is an appeal: you file it and you wait, and distribution is already limited in the meantime.

Three moves in recent days point the same way. Snapchat rewrote its community guidelines on 4 September 2026 with a focus on declaring AI-made content. In Europe, Article 50 of the AI Regulation has required algorithm-generated content to be declared and marked in machine-readable form since 2 August 2026. And on 5 September 2026 the Dutch advertising association VIA published the first operational guidance for agencies: four cases where disclosure is mandatory, one of them public-health copy with no human editorial review.

Together they move where disclosure lives. It stops being a good habit inside the creative team and becomes a process control, with a record and an owner. If the platform can no longer tell how an asset was made, the only reliable version of that fact is the one written down by whoever made it.

Two cheap habits follow, and they cover almost everything: keep the untouched original of every real photo that goes through editing, which is the only evidence available if the label lands wrong, and declare AI use in the copy instead of waiting for the system to do it.

The argument against disclosure is almost always commercial: saying it scares the audience off. The available data says otherwise. A FreeWheel study published on 5 September 2026, covering 2,496 adult viewers with fieldwork between November 2025 and April 2026, found that 1 in 10 rejects AI-made creative, while media buyers estimate nearly 5 in 10. It comes from a company inside the advertising industry and should be read as such, but what that 38-point gap overstates is the buyer's fear, not the public's.

The market is still testing the opposite route. Digiday reported on 9 September 2026 the rise of ghost creators: brands hiring people to post under a pseudonym. Hiding the author is the inverse bet to declaring, and it is the one that does not survive a single leak.

One uncomfortable closing note for anyone writing with AI and not saying so. An Orbit Media survey released on 10 September 2026 found that 92% of marketers already write with AI assistance while only 14% report good results from their blog. Using it is not the differentiator; human editing and original research still are.

The note explains the problem; the services explain the work.