Social media and AI-generated content: what each platform allows

The short answer is that all of them allow it. The useful answer is that the rules differ in three ways that matter: what counts as disclosable, who does the disclosing, and what happens if you skip it. This page maps what each platform’s own documentation says — every rule below links to the platform’s page, not to somebody’s summary of it.

Meta: label it, rarely remove it

Across Facebook, Instagram and Threads, Meta’s approach is to keep AI content up and attach context to it rather than take it down. The company says so plainly: AI-generated content stays on the platform so informational labels can be added, unless it breaks some other rule — voter interference, harassment, incitement [1].

The label is called “AI Info”, and it arrives in two ways: you disclose, or Meta’s systems spot industry-standard AI indicators in the file. Meta asks you to disclose when you post photorealistic video or realistic-sounding audio that was digitally created or altered, and says penalties are possible if you don’t [2].

Two things follow for anyone running an agent. Self-disclosure is not really optional — if the file carries provenance data the label appears anyway, and having chosen not to add it is the worse position to be in. And the standard is photorealistic or realistic-sounding, so a plainly illustrative graphic sits differently from a synthetic photo of a person.

YouTube: the label you cannot take off

YouTube’s version is a section in the expanded description called “How this content was made”, carrying an altered or synthetic content disclosure. It gets there three ways: you tick the box in Studio, you used YouTube’s own generative tools, or the file arrives carrying valid Content Credentials [3].

The sentence worth reading twice is YouTube’s own: when content is undisclosed, it “may take action to reduce the risk of harm to viewers by proactively applying a label that creators will not have the option to remove” [3]. Not a warning and not a strike — a permanent label you cannot undo.

Automatic detection changes the question. It is no longer whether they will find out — it is whether you said it first.

TikTok: the file tells on itself

TikTok was the first video platform to implement Content Credentials, the C2PA provenance standard, and it reads them on upload [4]. That means TikTok can label something made with an AI tool somewhere else entirely, because the metadata rides along with the file.

This is the part people underestimate. The platform does not need to recognise that an image looks synthetic. It reads a signature the generating tool attached. If your agent makes images or video with a mainstream tool, assume the provenance data is there.

The rule that sits on top of all of them

Since 2 August 2026, the EU AI Act’s transparency provisions apply. Article 50 requires AI-generated or AI-altered content shown to people in the EU to carry machine-readable marks identifying it as synthetic, alongside clear labelling [5].

That is a different kind of obligation. A platform label is a product feature the platform controls. A machine-readable mark is a legal duty on whoever deploys the system, and it applies whether or not a platform asks. It reaches a Singapore business serving EU customers, and it is far cheaper to design in than to retrofit.

What this means if an agent posts for you

Four things any agent with publishing rights has to handle, or you handle them by hand forever.

Disclosure, per platform. Meta’s tag, YouTube’s Studio attribute and TikTok’s label are three different mechanisms. An agent publishing to all three needs all three.

Machine-readable marks for EU audiences — separate from the platform labels, and a legal duty rather than a courtesy.

Brand voice under supervision. None of the platform rules say anything about whether the agent sounds like you. That one is yours, and it is the one your audience notices first.

A record of what went out and why. When a post lands badly, “the agent decided to” satisfies nobody. This is the observability point from the failure modes guide, applied to publishing.

That last group — disclosure handling, brand constraints, an audit trail — is where tools in this category now differ, considerably more than in the quality of the writing. Replying to comments and messages is a related but separate question, with its own rules — we will cover it on its own page.

AiAgentNook editorial · last verified 26 Aug 2026 · platform policies change often; corrections are noted here, not silently edited.