What the platforms require now

The layer that reaches you first

Legal duties are covered in the final module. This lesson is about the rules that actually bite: the ones written by the services where your work appears, enforced by their systems, with account-level consequences and slow appeals.

The pattern across major services, described structurally because the details change:

Self-declaration at upload. A checkbox or field asking whether the content is synthetic or has been meaningfully altered. Video platforms introduced requirements of this kind for realistic altered content, with a stated obligation to disclose when a viewer might mistake it for real.

Automatic detection and labelling. Services read embedded watermarks and provenance metadata and apply their own label. This is why watermarking exists commercially: it makes labelling possible at scale without asking anyone.

Category-specific rules. Stricter requirements for political and election content, for health claims, and for anything depicting a real person. Some services prohibit synthetic political advertising outright; others require prominent disclosure.

Prohibitions rather than labels for the worst categories: synthetic sexual content of real people, fabricated statements by public figures presented as real, impersonation for fraud. These are removals, not disclosures.

What "meaningfully altered" tends to mean

The threshold most services have converged on is close to the interpretive standard from the previous lesson: disclosure is required when a viewer could reasonably mistake synthetic content for a real recording of a real event or person.

So generally not required: colour correction, stylistic filters, obviously unreal or fantastical imagery, background cleanup, production assistance such as scripting or editing help.

Generally required: a realistic depiction of a real person saying or doing something they did not; a realistic-looking event that did not happen; a real place altered to look materially different; synthetic voice presented as a real person's.

The frictions to expect

Automatic labels you did not ask for. A photograph edited in software with generative features can pick up provenance metadata that a platform reads as AI, producing a label on a real photograph. Contest it, and expect it to take time.

Labels that do not survive. A label applied by one platform disappears when the file is downloaded and reposted elsewhere. Anything you rely on being disclosed should be disclosed in the content, not around it.

Inconsistency. The same piece can carry a label on one service and none on another. There is no shared threshold, and there is no sign of one arriving soon.

Advertising review. Advertising policies are stricter and enforced harder than organic content rules, with faster consequences. If the work is an advertisement, read that policy specifically, because it is a different document.

What to do

1. Read the actual policy for each service you publish on, once, and re-read when you hear it has changed. They are short.
2. Declare accurately at upload. A false declaration is a terms breach that can cost the account, and the upside of a false one is nothing.
3. Put the disclosure in the content as well as in the platform's field, so it survives redistribution.
4. Keep the generation record, which is what you will need if a label is applied wrongly.
5. Assume the strictest rule among the platforms you publish to, and make one version.

The honest assessment

These rules are new, unevenly enforced, and largely dependent on self-declaration and watermark reading — both of which the previous lessons showed can be avoided by anyone who wants to. They are therefore doing very little about deliberate misuse.

What they are doing is establishing a norm for the vastly larger volume of ordinary, non-malicious synthetic content, and creating a contractual hook that lets platforms act. That is worth something and it is not what it is often described as.

One more thing to know about how these rules are enforced, because it changes how you should respond. Enforcement is overwhelmingly complaint-driven and automated. Nobody is reviewing your upload; a system checks a watermark, a rights holder's matching system checks a fingerprint, or another user reports it. That means the decisions arrive without context and are appealed into a queue rather than to a person. It also means the single most effective protection is to be boring: accurate declarations, a clear in-content disclosure, no bulk uploading, and a record you can produce quickly. Accounts that never generate a signal are not the ones that get caught in the machinery.

For you, the practical summary is unromantic: the rules that constrain your work are contractual, they vary by service, they change without notice, and complying accurately costs you a minute per upload. Do it, keep the record, and put the disclosure where it travels with the file.