Platform labels compared: who labels AI images, how they decide, and what each label proves
Meta, TikTok and YouTube each describe an AI label on their own pages, and LinkedIn describes a C2PA icon. They differ in what sets it off: a creator's disclosure, signals inside the file, or the platform's own tools and systems. We found no statement on these four pages that a label proves how an image was made.
Which platforms label AI images, and what triggers each label?
We include only platforms whose own page we opened on October 30, 2026. We leave out X, Pinterest and the online marketplaces, whose rule pages we did not open. Rules change often. Some pages cover video as well as images.
| Platform and page | Label or indicator | What the page says triggers it | What the page does not say |
|---|---|---|---|
| Meta, approach to labeling (published April 5, 2024; last update listed October 23, 2025) | "AI info"; "Imagined with AI" on photorealistic images from its Meta AI feature | Meta detecting "industry standard AI image indicators", or a person disclosing an AI upload | That a label verifies whether content is true or false |
| TikTok newsroom post (November 19, 2025) | Labels on AI-generated content, with added context on whether a label came from detection, creator labels or TikTok AI tools | Creator disclosure, TikTok's own detection models, and Content Credentials in the file | That a label survives a re-upload; TikTok says labels "may be removed" then |
| YouTube help page "Disclosing use of GenAI content" | An AI label, shown in the player or in the expanded description | A creator answering "Yes" under "AI use" in YouTube Studio; YouTube may also apply it for its own generative tools, content with C2PA metadata, or content its systems flag | What a label proves beyond the disclosure or the signal that set it off |
| LinkedIn help page "Content credentials" | A "C2PA icon" on images and videos | Content cryptographically signed with C2PA Content Credentials | What viewers see on unsigned content, or how LinkedIn checks credentials |
How do Meta and TikTok decide?
Meta says it labels when it detects industry-shared signals that other companies add to content made with their tools, or when people disclose that they are uploading AI-generated content. On July 1, 2024 it renamed "Made with AI" to "AI info", saying indicator-based labels were not always aligned with people's expectations.
TikTok says it requires people to label realistic AI-generated content, uses its own detection models, and reads Content Credentials. It adds that these labels may be removed when content is reuploaded or edited elsewhere. The post says TikTok was therefore testing an invisible watermark that only it can read, and that it would start adding it "over the coming weeks" to content made with its own tools and content uploaded with Content Credentials.
How do YouTube and LinkedIn decide?
YouTube requires creators to disclose realistic content that is generated or meaningfully altered with AI. It says it may apply a label automatically for content made with its own tools, content that contains C2PA metadata, or content its internal systems detect. Creators are notified, and in most cases can change the disclosure if YouTube's systems err. The page says they cannot change one on content made with YouTube's tools, content with C2PA metadata, or content labeled after manual review.
LinkedIn describes a different design. It shows its icon when an image or video is signed with Content Credentials. Clicking it shows the credential, including an assertion about whether AI was used, which the page says comes from the entity that signed it. For Sponsored Content, the page says the icon appears only on single image or video ads, when the media is uploaded.
What does each label prove?
None of the four pages says its label proves how an image was made. YouTube says it adds a label after a creator selects "Yes", or automatically in some cases. LinkedIn says the AI assertion in its credential view comes from the entity that signed it.
The C2PA explainer speaks to the signal behind some of these labels: it says Content Credentials do not provide value judgments about whether provenance data is "true". So read a label as a report of its trigger, such as "the creator said so", not as a statement about the pixels. An image with no label shows only that no label was applied.
How much can a trigger inside the file see?
Several triggers depend on a signal carried inside the file. We measured how often such signals are present and what a screenshot-style re-render does to them. No platform was contacted.
Our measurement, October 30, 2026. We ran 31 real files filed as AI-made: 29 from Wikimedia Commons categories named "Images generated by" a product, and 2 vendor samples. Nine carry a signed Content Credentials record (8) or an unsigned IPTC source declaration (1). The other 22 carry neither; one of them has only a tool name in its metadata. Third parties uploaded the 29 Commons files, so we cannot know how each was handled before upload.
We then re-rendered the 9 real files in our set that carry a record (those 8 and the C2PA test image) through a browser canvas, as a screenshot or an export does, once as PNG and once as JPEG. After either, 0 of 9 still carried a record. This is a canvas, not any platform. We tested one browser and one JPEG quality, and we did not test a resize. Our accuracy page has the real-file counts and the re-render counts.
The result is narrow. A trigger that depends on a signal inside the file sees nothing in a re-rendered file. It does not say what any platform does with an upload.
What this means when you check an image
Open the checker, choose the file, and look at the headline. If a copy has lost every signal we read, it shows "No supported provenance signals found". It means we found no signed record, no source label in the metadata and no Stable Diffusion watermark pattern. The line under it says this does not mean the image is real or human-made.
If the file kept a record, the headline depends on the signer. Of the 9 real files in our set with a record, 1 reads "Trusted Content Credentials declare AI use", 5 read "Content Credentials found; legacy trust-list match", and 3 read "Content Credentials found; signer trust not established". The last group is two older OpenAI files with expired certificates and the C2PA test image. It means a signed record is present and no trust list we load vouches for the signing certificate. It does not tell you who really signed or whether the record's claims are true.
No headline tells you whether a platform labeled the post. We do not read a platform's labels, and we do not check Google's SynthID or Meta's own watermark. Our page on visible AI watermarks covers the marks you can see in the picture itself.
Guides on this topic
- Visible AI watermarks is the main page on this topic.
- Meta's AI info label covers the first row.
- TikTok's AI-generated label covers the second.
- YouTube's altered or synthetic content label covers the third.
- What LinkedIn's Content Credentials label means covers the fourth.
Frequently asked questions
- Does a missing label mean an image is real?
- No. A missing label shows only that no label was applied. LinkedIn's page says it is not possible to identify and label all AI-generated and modified content.
- Does an AI label mean the platform checked the image?
- Not necessarily. The pages describe several triggers, and a creator's own disclosure is one of them. Meta, TikTok and YouTube all say a creator's disclosure can set a label off.
- Do labels survive a screenshot or a re-upload?
- TikTok says its labels may be removed when content is reuploaded or edited elsewhere. The other three pages do not say. In our canvas test, 0 of 9 Content Credentials records survived a re-render.