AI Content Labels on Social Platforms: What Triggers Them

Three paper cards in graphite, mid grey and pale grey, each with a different small blank tab clipped to its top edge

The three inputs behind a label

Creator disclosure
The uploader ticks a box or adds a label when posting. Several platforms make this mandatory for realistic synthetic media and may penalise accounts that repeatedly fail to disclose.
Metadata in the file
Industry provenance data, such as a C2PA manifest or an IPTC digital source type, tells the platform that a generative tool made or changed the file. This is the input publishers influence most.
Platform detection
Classifiers and fingerprinting look for generated media without any metadata. Accuracy varies, and detection is the usual source of labels that surprise honest creators.

How the main platforms describe their approach

Platform	What its public policy says, in brief	Source
Meta (Facebook, Instagram, Threads)	Labels content when it detects industry-standard AI indicators or the user discloses it; asks users to disclose realistic video and audio	Meta newsroom
YouTube	Creators must disclose realistic altered or synthetic content; labels appear in the description or on the player for sensitive topics	YouTube Help
TikTok	Requires labels on realistic AI-generated content and can apply labels automatically from Content Credentials	TikTok Support
X	A synthetic and manipulated media policy that can label or limit misleading media	X Help Center

These summaries are deliberately short. Policies are revised often, and the exact thresholds (what counts as realistic, which topics are sensitive) are defined on the platforms' own pages, which you should read before a campaign.

Links versus native uploads

Most labelling happens on media uploaded directly to a platform. When a reader shares a link, the network builds a preview from your Open Graph tags and fetches the image named in og:image; whether it inspects that image for provenance data varies by network. Because Content Credentials often vanish when images are resized, a preview image rarely carries enough data to trigger a label either way. Your page is where disclosure has to live.

Regulation is catching up

In the European Union, Article 50 of the AI Act (Regulation 2024/1689) sets transparency obligations that apply from August 2026: providers of generative systems must mark output in a machine-readable way, and deployers who publish deepfakes, or generated text meant to inform the public on matters of public interest, must disclose it, with an exception for text that has undergone human review and for which a person holds editorial responsibility. That exception is one more reason to keep a named, accountable editor in your byline.

A checklist before you publish

Before a piece with generated or heavily edited media goes out, check three things. First, that the caption or disclosure line on the page matches what actually happened. Second, that the metadata in the file says the same thing, because a platform will trust the file over your caption. Third, that the share image named in og:image is the one you intended, not an automatically generated crop of a different picture. Most labelling disputes start with one of those three drifting apart.

Avoiding false labels

Check the metadata your editing tools write. Some editors mark any file touched by a generative feature, even a small background fill, and platforms may label the whole image.
If a label is wrong, use the platform's appeal or edit option and keep the camera original, which is your best evidence.
Do not strip metadata to dodge labels. It removes your own evidence and does not stop detection-based labels.
State your policy on the page, as described in what publishers should do about AI content in feeds.

The first two inputs are where honest publishers prevent surprises.

AI content label outcomes: a worked example

Suppose a publication uploads 30 images natively to one platform over a month and shares 30 links to posts with the same images. The counts are invented for this example. Of the 30 native uploads, 10 were generated with the source type field set, 10 were generated with all metadata stripped by an optimiser, and 10 were photographs. In the example, 9 of the 10 labelled files receive an AI label, none of the stripped files do, and 1 photograph is labelled because a retouching step had written a generative action into its metadata.

Of the 30 shared links, none receives a label on the card, because the preview images fetched by the platform were resized sub-sizes carrying no metadata. The two lessons are the ones this guide keeps returning to: the file's metadata decides the native label, and a link preview almost never carries enough data to trigger one. A publisher who wants the disclosure to reach card readers has to put it in the description, as the guide to disclosure lines that survive a share shows.

AI content label pitfalls

Reading a label as a verdict. A label says the platform found an indicator or a disclosure, and it can be wrong in both directions.
Stripping metadata to keep a photograph unlabelled after a small generative fill. The label may still arrive through detection, and the original evidence is gone.
Assuming one platform's policy matches another's. Thresholds for realistic and sensitive differ, and so do the surfaces where a label is shown.
Skipping the disclosure toggle because the file is already labelled. Some platforms treat the toggle as a separate duty and penalise its absence.
Relying on a preview image to carry disclosure. It typically arrives at the platform as a re-encoded copy with nothing to read.

What to record

Keep a label log for every post with generated or heavily edited media: one row per platform, noting what you uploaded, what the platform showed and what the file carried at each step. The entries below are examples in the same format as the share test log used across the provenance section.

Page	Network	Time (UTC)	Preview or post showed	File carried
/2026/09/street-scene/	Native upload	2026-09-18 16:40	AI label under the image	Source type: trained algorithmic media
/2026/09/street-scene/	Link share	2026-09-18 16:52	Card with title and description, no label	Sub-size, no metadata
/2026/09/portrait/	Native upload	2026-09-18 17:05	AI label on a real photograph (appealed)	Generative action from retouching
/2026/09/portrait/	Link share	2026-09-18 17:15	Card, no label	Sub-size, no metadata

The log is also your appeal evidence. When a real photograph is labelled, the row that records what the file carried tells you which editing step to change, and the guide to what platforms strip from shared images explains why the link rows stay empty.

Sources

Questions

Why was my real photo labelled as AI?

Usually because an editing step wrote generative metadata into the file, or a detector misread it. Appeal and keep the original.

Do link previews get AI labels?

Less often than native uploads. Preview images are usually resized copies without provenance data.

Does the EU AI Act apply to my blog?

Its transparency duties can apply to deployers publishing deepfakes or generated public-interest text in the EU, with an exception for human-reviewed text under editorial responsibility. Take advice for your situation.

Can I remove an AI label from my post?

Usually through the platform's edit or appeal route, if the label is wrong. Fix the metadata that triggered it first, or the next upload will be labelled again.

Does sharing a link avoid AI labels?

In practice the preview rarely carries provenance data, so link shares are seldom labelled. That is no reason to skip disclosure on the page, where readers and regulators look.

Hannah Voss, Editor. Checks every guide against a working WordPress install and the networks' current documentation. Last reviewed September 2026.

About the editor and our methods