AI-generated content must announce itself before anyone acts on it

What it is

A product photo in a feed, lighting perfect, caption “shot on location.” Nothing tells the viewer the image was generated. Generated-content identifiability requires that a recipient, before using or passing the material on, can know that what they are looking at is not a direct record of a person, a scene, or a work, but a model synthesis.

Visible badges and watermarks are two technical channels with different failure modes. The question here is whether the recipient can know the origin class, not which channel carried it.

Why it happens

People calibrate belief by origin class: a photograph as evidence, an illustration as a claim, a generated artefact as a synthesis. Strip the class and the calibration channel closes; the material is treated as whichever class it most resembles — a photo-like generation as a live shot, a reporter-voiced paragraph as reporting. A generator’s objective does not include “look synthetic”; realism lowers identifiability.

Identifiability is a state on the recipient’s side, not a feeling on the producer’s. The person who pressed generate knows; the person they forwarded it to, and readers after that, do not. The label’s object is every later recipient.

Studying it

The same material labelled “live / generated / unspecified”; measure evidential weight, willingness to share, willingness to cite as fact. Independent variables: presence of a class mark, realism of the material. Dependent variables: accuracy of origin-class judgment, rate of treating generated artefacts as records.

Realism must be a factor. Misjudgment under high realism and no mark is the measure of identifiability failure; a low-realism piece guessed correctly is not the label working.

Where it stops holding

When the user just generated the piece and has not left this interface, the action itself supplies identifiability; an extra badge is a reminder, not the only channel. Explicitly fictional genres (a game portrait, parody) will not be taken as records anyway. In accessible settings, a colour-only badge is invisible to some users; identifiability must be readable by a screen reader. This entry does not treat whether a label still exists after a screenshot, nor how absence pollutes downstream corpora.

Applying it

• Put the class at first presentation of the generated artefact, readable by a recipient who does not know the product — not only a settings line that “this product uses AI.”
• Class language should name origin (“model synthesis”), not quality (“may be inaccurate”). A quality warning is not a class.
• Raise salience for modalities that can pass as evidence (photographs, voices, official layouts).
• Check: hand the output to someone who has never seen the product and ask “record or synthesis.” Every answer of “record” is a failure of identifiability.

Related

• Same group: L3.04.2 Labels must survive after the content circulates · L3.04.3 Missing labels pollute the later information environment
• Nearby: L3.10 Labeling and Watermarking AI Content · L3.05 Copyright and Material Provenance
• Search terms: generated-content identifiability · machine-origin disclosure · synthetic media label