Contents
To label AI-generated content you work in two layers: a visible, human-readable label that platforms such as YouTube, TikTok and Meta either require you to set or add on your behalf, and a machine-readable provenance signal, C2PA Content Credentials or the IPTC Digital Source Type metadata, that those platforms read to detect the file and label it automatically. Treat those as separate jobs. The visible label tells the audience what they are seeing; the machine-readable signal gives the platform a technical reason to apply, preserve or verify that label. The rules below are what the major platforms actually require in 2026, and they apply to images and audio alike.
The two layers
The two layers do different jobs. The visible label is the notice a viewer reads: YouTube’s altered-content disclosure, Meta’s AI label, TikTok’s AIGC tag. The machine-readable layer is provenance metadata a person never sees but a platform can parse, in practice C2PA Content Credentials and the IPTC photo-metadata standard’s Digital Source Type property. The connection between them is the point: when your file carries the machine-readable signal, a platform can recognise it and apply the visible label for you, and when it does not, the burden falls back on you to declare it manually. So if you publish your own AI image, audio or video, do both whenever the platform gives you the option: attach Content Credentials before upload, then set the visible disclosure in the platform workflow.
Platform by platform
| Platform | What it requires or does | Signal it reads | The visible label |
|---|---|---|---|
| YouTube | creator self-disclosure of realistic synthetic media | a creator declaration in Creator Studio | a disclosure in the description, set via the Altered content control |
| TikTok | auto-labels on upload, and lets you self-label | C2PA Content Credentials | an AIGC label |
| Meta (Facebook, Instagram, Threads) | auto-labels detected AI content | C2PA and IPTC metadata | an AI label; “Imagined with AI” for its own images |
YouTube: you declare it
YouTube puts the duty on the creator. You must disclose when, in YouTube’s words, “realistic content, content a viewer could easily mistake for a real person, place, scene, or event, is made with altered or synthetic media, including generative AI” (YouTube). You set it with the Altered content control in Creator Studio, and the resulting label appears in the expanded description, more prominently for sensitive topics such as health, news, elections and finance. Exempt uses include AI for “productivity, like generating scripts, content ideas, or automatic captions” and “clearly unrealistic content, such as animation or someone riding a unicorn” (YouTube). The practical test is whether a reasonable viewer could mistake the synthetic media for a real person, place, scene or event.
TikTok: it reads the credential
TikTok leans on the machine-readable layer. It is “launching the ability to read Content Credentials” so it can “instantly recognize and label AIGC”, which means a file that already carries C2PA is labelled automatically on upload (TikTok, 2024). TikTok also said it would, “over the coming months”, “start attaching Content Credentials to TikTok content, which will remain on content when downloaded”. For a creator the rule of thumb is simple: if your AI content is going to TikTok, attach C2PA Content Credentials before upload when your creation tool supports it, and still use TikTok’s own disclosure controls when they are presented. The credential is what triggers the automatic tag; the visible AIGC label is what the viewer reads.
Meta: it reads the metadata
Meta labels AI content across Facebook, Instagram and Threads by reading provenance signals rather than asking every creator. It built “industry-leading tools that can identify invisible markers at scale, specifically, the ‘AI generated’ information in the C2PA and IPTC technical standards” (Meta, 2024), and it labels its own Meta AI photorealistic images “Imagined with AI”. Do not overread that into a guaranteed label string for every third-party upload. The verified and more useful point is narrower: Meta reads the same C2PA layer TikTok does, plus the IPTC photo-metadata property, so a file that carries either can be labelled without your involvement, as those signals appear in files from AI and imaging companies.
Attach the machine-readable signal
Because two of the three platforms key off provenance metadata, the single most useful move is to attach a machine-readable signal before you upload, then verify it. Some tools add C2PA Content Credentials automatically; you can also add and check them yourself, in How to add Content Credentials and How to verify Content Credentials. If your export tool offers a Content Credentials option or an IPTC Digital Source Type field, set it before you publish. C2PA is the leading machine-readable provenance standard, and SynthID is the parallel invisible-watermark approach for AI images; both aim at a signal a platform can detect. Attaching one does not replace setting a platform’s visible disclosure where that is required, it makes the automatic labelling possible.
Why the rules are converging
These platform rules are lining up for a reason. From 2 August 2026 the EU AI Act’s Article 50 requires providers of generative systems to mark synthetic output so it is machine-readable and detectable as AI, and requires deployers to disclose deepfakes to viewers (Regulation (EU) 2024/1689, Article 50). That is the same visible-plus-machine-readable split the platforms are building, which is why C2PA-style marking is spreading across them. The compliance detail is in EU AI Act Article 50 explained.
The limit
One caveat decides whether any of this holds: the machine-readable layer is fragile. Platforms that re-encode on upload strip or invalidate C2PA as a side effect of ordinary processing, and a stripped credential is silent, indistinguishable from a file that never had one, because a hash has no inverse and the first changed byte invalidates the binding, which is 100 percent invalidation on any real edit (C2PA Specification v2.4). So do not assume the mark survives the trip: attach the signal, set the platform’s visible label, and keep one unchanged export copy that can still verify later. And remember a label satisfies a disclosure duty, it does not prove a file honest, since “C2PA provides provenance signals, not proof of authenticity” (Golaszewski, Krawetz, Sherman, 2026). Whether a credential survives a given platform is covered in do Content Credentials survive social media or a screenshot?. Deciding whether someone else’s content is AI is the opposite job, detection rather than disclosure, covered in is this image AI-generated?.
Sources
- Golaszewski, Krawetz, Sherman, et al. (2026). Verifying Provenance of Digital Media: Why the C2PA Specifications Fall Short. arXiv:2604.24890.
- European Parliament and Council (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 50.
- Meta (2024) Labeling AI-Generated Images on Facebook, Instagram and Threads. Available at: https://about.fb.com/news/2024/02/labeling-ai-generated-images-on-facebook-instagram-and-threads/ (Accessed: 3 July 2026).
- TikTok (2024) Partnering with our industry to advance AI transparency and literacy. Available at: https://newsroom.tiktok.com/en-us/partnering-with-our-industry-to-advance-ai-transparency-and-literacy (Accessed: 3 July 2026).
- YouTube (2024) How we’re helping creators disclose altered or synthetic content. Available at: https://blog.youtube/news-and-events/disclosing-ai-generated-content/ (Accessed: 3 July 2026).
- Coalition for Content Provenance and Authenticity (C2PA) (2026) C2PA Technical Specification, version 2.4. Available at: https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html (Accessed: 3 July 2026).