Guide to AI-content labels on Instagram, TikTok, YouTube and more — what each label means and where it falls short.
Free with no sign-up, and the check runs inside your browser, so what you enter never leaves your device. The result is a signal to help you decide, not proof for money or legal decisions.
Instagram, TikTok and YouTube have started labeling AI content — but the labels mean less than people think, and "no label = real" is never true. Here is how each platform labels, where to look, and how far to trust it.
Instagram/Threads (Meta): the "AI info" label — applied via creator disclosure plus metadata (C2PA-style) auto-detection, shown near the account name or in the post menu. If an editing app strips the metadata, detection misses it. TikTok: the "AI-generated" label — disclosure plus own detection; automatic for TikTok's in-app AI effects, but externally made videos rely on disclosure. YouTube: the "altered or synthetic content" notice — disclosure is mandatory for realistic synthetic media, shown in the description (or on-screen for sensitive topics).
Shared limits: much of the system relies on self-disclosure — deceivers simply do not disclose; auto-detection is metadata-based, so screenshots and re-encodes evade it; and a label does not always mean deepfake — beauty filters can qualify too. Read labels as "useful when present, meaningless when absent". When in doubt, save the file and run it through this site's pixel and provenance checks yourself.
Much automatic platform labelling works by reading a provenance record inside the file. C2PA is the standard for that record, carrying who made the content, when, with which tool, and what editing followed — cryptographically signed, so it is hard to alter quietly.
The trouble is how easily it disappears. Screenshot the content and you have a redrawn image with no record at all. Re-compress it, or run an edit that tidies metadata, and it goes. Platforms re-encoding uploads to their own specification drop it too.
So the record is strong evidence when present and no information whatsoever when absent. Anyone intent on deceiving can erase it with a single screenshot, which means treating its absence as evidence inverts your judgement.
Labels rarely sit where you are looking. They go inside an overflow menu, at the bottom of a description, or as a small marker beside the account name — one step in from the surface. You will not catch them scrolling a feed.
For a post that matters, fix a sequence. Open the post menu and look for AI-related markings; if there are none, expand the full description and read it; if still nothing, save the file and put it through the provenance check on this site to read what is inside it. A file can carry a record even when the screen shows no label.
Know the reverse case too. A label does not mean deepfake. Beauty filters and partial use of generative editing tools attract the same label. It says an edit happened, not what was edited or how far.
The labelling regime tells you what is there; it certifies nothing about what is not. So practice does not end at checking for a label. Treat its absence as starting with one fewer piece of information.
The order runs like this: save the file and check its provenance, then move to the pixel-level check, and for photographs of people, run a reverse image search to see whether the image appears elsewhere. If all three come back empty, record that fact. Nothing at all is a common shape for recent generated content.
Then look at context. Who posted it, when, where, and what that account has posted before is frequently stronger evidence than any image analysis. Labels support that judgement; they do not replace it.
Uploading a file straight from a generator lets the platform read its C2PA credentials and attach an "AI info" style label with no action from the uploader. These labels are comparatively trustworthy, because they rest on a signed record inside the file rather than on someone’s declaration.
YouTube’s "altered or synthetic content" notice appears only when the uploader ticks the box, which they are required to do for realistic synthetic scenes. No disclosure, no label. So this type is useful information when present and means nothing when absent.
Screenshotting and re-uploading, passing the file through an editor, or stripping metadata all remove the C2PA record so no label attaches. Self-disclosed labels simply never appear if nobody discloses. A label carries meaning only when it is there. Unlabelled content still has to be judged on its own — pixel analysis, provenance data, reverse image search.
No. The system depends on self-disclosure and metadata detection, both easy to evade. A missing label is "no information", not "verified real".
For realistic AI content, turning on the platform's disclosure setting (e.g. YouTube's altered-content disclosure) is both the rule and basic respect for viewers. Getting caught without it can mean account penalties.
For realistic scenes, yes, as a rule. Platforms provide disclosure settings and many require notice for realistic synthetic media. Posting without it risks removal or account penalties, and more simply, it is basic courtesy to the people looking.
No. A label indicates the rules were followed. The problem case is realistic synthetic media with no label at all.