11.2 Four Platforms, Four Different Systems

The AI Act is one law, applied by authorities that take years to act. The platforms are four private rulebooks, applied by automated systems that act within hours. You are subject to all of them at once, and satisfying one buys you nothing on the next.

This is the layer that actually ends businesses. Nobody has lost an account to a market surveillance authority. Plenty of operators have lost an account, a reach profile, or a monetization program to a classifier.

Why platform rules bite harder than law

There is no court here. There is an Account Status page, a notification, and a decision. No due process, no appeal to a regulator, no proportionality test. Reach is throttled quietly, before you are ever told, and by the time you notice a decline you are looking at a distribution problem that started weeks earlier.

The second thing to internalize is what the penalty actually is. For a persona account, the scarce asset is not the content and not the account. It is the recommendation slot. A label costs you nothing. Missing a label costs you the slot.

Instagram and Meta

Meta renamed the declaration on 31 August 2026, from "AI creator" to "AI-generated profile." [AS OF 2026-10] The rename matters less than the enforcement structure underneath it, which has three parts.

Self-declaration, at two levels. There is an account-level setting that declares the profile as AI-generated, and there is a per-post label for content that depicts realistic AI-generated material. The account-level declaration functions as a blanket statement about who is behind the profile. The per-post label is what travels with a video into someone else's feed.

Detection-backed enforcement. If you do not declare and Meta's classifier flags the content anyway, the consequences attach to distribution: exclusion from recommendation surfaces, reduced distribution in Reels and Explore, and removal in repeated cases. The exposure is not the label — it is everything the algorithm stops doing for you.

The asymmetry, which is the single most useful fact in this chapter. Applying the label does not by itself reduce reach, according to Meta. [VENDOR CLAIM] That is the platform's own claim about its own ranking system, so treat it as a claim rather than a measurement. But the asymmetry holds even if the claim is only partly true: undeclared plus detected is strictly worse than declared. Declaring is the lower-risk branch in every scenario where detection is possible, and detection is improving.

Appeals. Violations surface on the Account Status page, and the appeal starts there. Appeals against an AI label are escalated to human review. False positives are real: authentic photographs have been labeled AI-generated, and the appeal path is slow relative to the damage to a launch. Chapter 11.5 covers what to do when the classifier is wrong about you.

Meta ads behave differently. Hard self-disclosure requirements apply to political and social issue advertising. Everything else gets a soft automatic label when Meta's detection flags the creative. The practical consequence: if you run paid traffic to a persona page, assume a label will be attached at some point, and put your own disclosure into the creative so the platform's label is confirmation rather than a surprise.

TikTok

TikTok's trigger is not "was this generated." It is realism — whether a viewer could mistake the result for a real person, place, or event. The policy works on four tiers:

Read the tiers as a statement about your visual direction, because that is what they are. A stylized, clearly illustrated persona sits in the fourth tier for most of its output. A photoreal persona sits in the third or first tier permanently. Your art direction decision is partly a compliance decision, and it is much cheaper to make early than to reverse after a thousand posts.

TikTok also applies its own labels automatically through content credentials where the provider embedded them. That helps you and does not discharge your obligation — chapter 11.5 explains why uploaded metadata does not survive a reupload.

Two absolute bans, with legal consequences attached. Political deepfakes are prohibited outright, and so is impersonation of a real person. These are not moderation outcomes; they are removal plus account termination plus referral to legal authorities. If your persona is a lookalike of a real public figure, or touches electoral content, the risk profile is different in kind from everything else in this book. Read chapters 11.4 and 11.6 before you publish anything in that shape.

YouTube

YouTube's rule is built around realistic content: you must disclose altered or synthetic content when it could be mistaken for a real person, place, or event. The mechanism is a disclosure checkbox at upload, which produces a public label.

YouTube publishes its exceptions rather than leaving them to inference, which makes it the most predictable of the four systems. The categories where disclosure is not required include clearly unrealistic or stylized content, production assistance such as scripting, captions, and color work, and content where the synthetic element is itself the point. Two details from that list deserve attention:

• Stated clearly unrealistic is a real exception. If nobody could mistake it for reality, the checkbox is not for you.
• Satire is not a blanket exception. Where the synthetic depiction could be mistaken for a real person saying something real, disclosure is required even when the intent is parody.

Failing to disclose does not automatically remove the video. YouTube may apply the label itself when it detects synthetic content, and repeated undisclosed realistic synthetic content moves you into the misinformation policies, where strikes and monetization consequences live. That is the escalation path worth avoiding, and it starts with a checkbox that takes two seconds.

Google Ads

Google Ads surfaces a "How this ad was made" panel on ads created or significantly edited with AI. It is informative for the viewer and it is not something you configure. The operative consequence for an advertiser is the same as on Meta: expect the disclosure to appear, and make sure your own creative does not contradict it. An ad that reads as a testimonial from a real customer, labeled as AI-generated by the platform, is a credibility problem you created by not declaring first.

A platform label is not a legal disclosure

These two things look identical on screen and are not interchangeable. A platform-applied "AI-generated" label is a provenance signal produced by a classifier or by content credentials. The Article 50(4) disclosure from chapter 11.1 is something you give, clearly and visibly, at first exposure.

Three reasons the label does not discharge the obligation:

• It may be applied without your participation. A duty you owe cannot be satisfied by an action a third party takes on your behalf, especially when it is triggered by detection rather than by your declaration.
• It may not be visible where exposure happens. A label that lives behind a tap or in an overflow menu is not clear and visible at first exposure.
• It may not appear at all. Detection is imperfect, and a provider watermark stripped by a reupload removes the trigger entirely.

The sentence to remember: the label satisfies the platform, and your own disclosure satisfies the law. Do both, because they cost the same two seconds.

One context point about what you are trading against. Organic reach on Instagram has been falling across the board — a widely cited figure puts the decline at roughly 35% year over year. [UNVERIFIED] Whatever the exact number, the direction is not in dispute, and it changes the arithmetic of these rules. A distribution penalty that removes you from recommendations is no longer a haircut on a growing baseline. It is a subtraction from a shrinking one. That is why the platform layer, not the statute, is the one to take seriously first.

The comparison

One flow, four labels

The efficient way to satisfy four systems is to stop thinking platform by platform and build the disclosure into production once:

1. Declare at account level wherever the setting exists. One toggle, permanent, and it removes the "you hid it" reading from every future enforcement decision.
2. Declare per asset when the content is photorealistic synthetic media. This is the tier that travels into feeds and cannot be fixed retroactively.
3. Put the disclosure where it travels. Caption text and on-frame marking survive reposting; a bio does not. Chapter 11.3 gives the wording.
4. Log the declaration. Which asset, which platform, which label, when. Chapter 11.5 turns this into the register that answers the question you will eventually be asked.

Then check the four rulebooks on a schedule, because the rules move. Instagram renamed its category twice in under two years, and a rename is the kind of change that never announces itself.

Where this goes wrong

Assuming the label is the penalty and hiding it. The verified asymmetry runs the other way. Undeclared and detected is worse than declared, on every platform where detection exists.

Assuming compliance transfers. Four rulebooks, four definitions of the trigger. The AI Act does not preempt any of them.

Treating visual style as purely a design decision. On TikTok and YouTube, realism is the trigger. Style is a compliance variable with a cost attached.

Never opening Account Status until reach drops. Throttling is quiet by design. Make the page part of the weekly loop from chapter 12.2.

Reading the policy pages once. The changes that catch operators are terminology shifts and tier redefinitions, not headline bans. Chapter 12.5 schedules the re-read.

Accepting platform claims about their own ranking as measurement. "The label does not reduce reach" is a statement from the party whose classifier applies it. Plan for the asymmetry, not for the reassurance.

The gate

1. Where is your account-level AI declaration, and is it on?
2. Which of your recurring formats are photorealistic synthetic media, and which are stylized?
3. Where does your disclosure live on each of the four surfaces you publish to?
4. When did you last open Account Status on each platform?
5. What is your appeal path if a real photograph of a human collaborator is labeled AI-generated?

What you leave with

Four rulebooks, their triggers, their penalties, and one production flow that satisfies all four without a separate compliance pass per platform. The wording to paste into that flow is next.