TikTok says it labels AI-generated content. Instagram says it does too. Facebook has the same policy. If the platforms are labeling AI video, why are deepfakes still going viral on all three?
Because the label system has a flaw that all three platforms share. The label is embedded in the file at the point of creation. It is not burned into the video itself. Anyone who wants to remove it can do so in under a minute.
Here is how each platform's system works, what it can catch, and what it cannot.
How TikTok, Instagram, and Facebook Label AI Content
The technical standard behind all three platforms' labeling systems is C2PA (Coalition for Content Provenance and Authenticity). C2PA is a metadata standard developed by Adobe, Microsoft, Intel, and others. When a video is created by an AI tool that supports C2PA, the tool embeds a cryptographic tag in the file's metadata declaring that it was AI-generated.
When TikTok, Instagram, or Facebook receives a video with a C2PA tag, the platform reads that tag and adds a visible label to the post. The label typically reads something like "AI-generated content" or has a small icon indicating synthetic origin.
This system works correctly when every step functions as designed: the AI tool embeds the tag, the creator does not remove it, the video is uploaded directly without processing, and the platform reads the tag accurately.
In practice, none of those conditions are guaranteed.
TikTok
TikTok uses C2PA metadata detection combined with self-disclosure requirements. Creators who use AI tools to generate video are required by TikTok's community guidelines to disclose this in the post. TikTok also uses automated systems to detect AI-generated content and apply labels independently of creator disclosure.
In its Q1 2026 transparency report, TikTok stated it removed 2.3 million AI-generated videos in three months. That figure covers videos the platform's systems caught. It does not represent the total volume of AI video uploaded during that period, which is orders of magnitude larger.
TikTok's labeling has two practical weaknesses. First, the C2PA tag is stripped if the video is processed after creation: run through a screen recorder, a video editor, a compression tool, or any third-party app that does not preserve metadata. A video that was labeled at creation arrives on TikTok without any tag, and TikTok's automated detection has to identify the AI origin from the visual content alone, which it cannot always do. Second, TikTok's self-disclosure requirement depends on creator honesty. A creator running a fraud operation will not disclose.
Instagram's approach is similar. Meta developed its own AI detection system and combined it with C2PA support. Instagram applies labels automatically when it detects AI-generated imagery or video, and requires creators to self-disclose when posting realistic AI content of people.
Meta has been more public about the limitations of its system than TikTok. The company has acknowledged that detection accuracy varies by content type and that the system is less reliable on video than on still images. The self-disclosure policy carries the same enforceability problem: it catches compliant creators, not bad actors.
Instagram's labeling also applies inconsistently to Reels compared to static posts. Video detection requires more computational resources and the system updates more slowly as new AI generators are released.
Facebook uses the same underlying Meta detection infrastructure as Instagram, applied to posts and video in the main feed and in Groups. The system performs similarly to Instagram's.
Facebook's AI content policy is slightly older and has had more revision cycles. The current version requires disclosure for synthetic media that "realistically depicts a real person saying or doing something they did not say or do." This language covers deepfakes of real people but does not cover fully synthetic people who do not exist, which is the category that covers most AI influencer fraud.
The shared flaw: metadata stripping
Every platform's C2PA-based labeling system depends on the metadata surviving from creation to upload. It frequently does not.
The most common removal methods require no technical skill:
Screen recording: Record the AI-generated video on a second device while it plays. The output is a new video file with no embedded metadata from the original. It takes 60 seconds.
Re-encoding: Run the file through any video editor or compression tool that does not explicitly preserve C2PA metadata. Most consumer apps do not. The output file has no AI origin tag.
Third-party download and re-upload: Download the video from one platform, upload it to another. The receiving platform's metadata handling determines whether the original tag survives. In most cases it does not.
A creator who wants to post AI-generated content without a label does not need to be technically sophisticated. They need to record their screen.
What the label tells you when it is there
A present label is a reliable positive signal. If TikTok, Instagram, or Facebook has added an AI content label to a video, it means one of two things: the platform's automated detection identified AI generation, or the creator disclosed it. Either way, the label indicates genuine AI involvement at some point in the production chain.
The label is not reliable as a negative signal. A missing label does not mean the video is real. It means the platform did not detect AI origin through the automated system and the creator did not disclose. Both of those conditions can be false while the video is still AI-generated.
This is the gap that community-based verification fills. When a video circulates without a platform label, the question of whether it is AI-generated does not have an authoritative answer from the platform. It has an answer from the people who have seen it, analyzed it, and reported it.
What the platforms should do differently
The structural fix for metadata stripping is to move from embedded metadata to perceptual hashing: a fingerprint derived from the visual content of the video itself, not from the file's metadata. Perceptual hashing survives re-encoding and screen recording because it is based on what the video looks like, not what the file says about itself.
Several enterprise tools use perceptual hashing for detection. Platform-level implementation would require cooperation on a shared hash database, which raises competitive and privacy concerns. As of April 2026, no major social platform has committed to a perceptual hashing standard for AI content labeling.
Until that changes, the C2PA system remains a first layer, not a complete one.
How to use platform labels correctly
Treat platform labels as one signal among several, not as the final answer.
If a label is present: The video is AI-generated or AI-assisted. Trust the label.
If a label is absent: Check the video against what the community has flagged. Look for the visual tells that AI generators cannot yet hide. A missing label is not clearance.
If the account posts consistently AI-generated content: An account with a pattern of AI video is more likely to post new AI video. Account-level patterns are more reliable than per-video labels. Ledger checks at the account level for TikTok, which is why account-level verdicts are available separately from video-level checks.
Understanding the label system is part of understanding what a deepfake actually is and how synthetic content moves through platforms without detection. The label is a tool. Know what it can and cannot do.
Related Posts
- What Is a Deepfake? A Plain-English Guide for Social Media Users: how synthetic video is generated and why certain platform defenses fail against specific generation methods
- How to Tell If a TikTok Video Is AI-Generated: 7 Signs to Check Right Now: the visual detection guide for when platform labels are absent or unreliable
- What to Do When You Find a Deepfake on TikTok or Instagram: reporting steps across all three platforms covered in this post
