AI Content Labeling and Disclosure

AI content labeling is the set of rules and technologies that tell people when images, audio or video were generated or substantially altered by AI: visible labels that platforms show to viewers, disclosures that creators must make when they publish, and invisible, machine-readable marks embedded in the media itself. Until recently it was a matter of platform policy. As of October 2026 it is also law in several jurisdictions, with the EU AI Act's transparency obligations applying since 2 August 2026, California's AI Transparency Act in force, and China's labeling rules in effect since September 2025. For anyone in the creator economy publishing synthetic media, labeling is now part of delivery, like captions or loudness.

Platform Rules

YouTube has required creators to disclose realistic altered or synthetic content since March 2024, using a setting in Creator Studio at upload. The test is whether a viewer could mistake the content for a real person, place, scene or event: replacing someone's face, generating a person's voice to narrate a video, or showing a real building on fire all need disclosure. YouTube explicitly exempts production help, such as AI-generated scripts, ideas or automatic captions, along with clearly unrealistic content like animation, and ordinary enhancements such as color correction, background blur or beauty filters. Its current help page says the label appears on the player for photorealistic AI content and AI-generated music, and in the description for other AI content, and that YouTube applies labels itself based on C2PA metadata, its own generative tools and its own detection systems.

TikTok has its own label for AI-generated content, and in May 2024 became the first video platform to implement C2PA Content Credentials, automatically labeling uploads that carry them. In November 2025 it said it had labeled more than 1.3 billion videos, began adding invisible watermarks to AI content made with its own tools and to uploads labeled through C2PA, and started testing a control that lets people see less AI-generated content in their feeds.

Meta labels AI content on Facebook, Instagram and Threads with an "AI info" label, which replaced "Made with AI" in July 2024. It relies on C2PA and IPTC metadata and on users' own disclosures; since September 2024 the label sits in the post menu for content that was only edited with AI, and stays visible on content that was generated.

The Law

The EU AI Act's Article 50 sets two duties that matter for video. Providers of AI systems that generate synthetic audio, images, video or text must ensure the output is "marked in a machine-readable format and detectable as artificially generated or manipulated." Deployers who use AI to create a deepfake must disclose that it was artificially generated or manipulated. For "evidently artistic, creative, satirical, fictional or analogous" works, the duty is limited to disclosing the AI content "in an appropriate manner that does not hamper the display or enjoyment of the work." The obligations apply from 2 August 2026; according to Cooley's analysis, generative systems already on the market before that date have until 2 December 2026 to meet the marking requirement. Breaches can be fined up to €15 million or 3 percent of worldwide annual turnover, whichever is higher. The Commission's Code of Practice on transparency of AI-generated content, finalized on 10 June 2026, describes how providers and deployers can comply.

In the United States, California's AI Transparency Act (SB 942, as amended by AB 853) requires large generative AI providers to offer a free AI-detection tool, to embed latent (hidden) disclosures in AI-generated content, and to give users the option of a visible disclosure, with penalties of $5,000 per violation. Its main provisions became operative on 2 August 2026, with obligations for large online platforms following on 1 January 2027 and for makers of capture devices on 1 January 2028. China's Measures for Labeling AI-Generated and Synthetic Content, in effect since 1 September 2025, require both explicit (visible) and implicit (metadata or watermark) labels, and require users to declare synthetic content when they publish it.

Credentials and Watermarks

Two technologies carry the machine-readable half of labeling. C2PA Content Credentials are signed metadata that record how a piece of media was made and edited; the specification reached version 2.4 in April 2026. Credentials carry rich, verifiable provenance, but metadata can be stripped by a screenshot, a re-encode or a platform that discards it. Watermarks such as Google's SynthID embed an imperceptible signal in the pixels or audio itself, which survives many transformations but can only be read by tools that know the scheme; Google said in November 2025 that SynthID had watermarked more than 20 billion pieces of AI-generated content. The two are complementary, and platforms increasingly read both, falling back on their own detection models when neither is present.

Labels tell viewers what they are seeing; they don't protect the person being depicted. Consent is a separate control, and some tools build it in. The audio-to-video platform Koyal, for example, requires a person to complete CHARCHA, a check of randomized live actions on a webcam such as turning their head or smiling, before their likeness can be used to create a character, so that one person can't upload another's photos. Laws on digital replicas, covered in AI dubbing, increasingly require such consent for voices and faces.

A Practical Checklist

  • Disclose realistic synthetic people, places and events at upload, using each platform's own setting, even when the platform might detect it automatically.
  • Don't over-label production help. AI-written scripts, captions, cleanup and color correction generally don't need a label; synthetic faces, voices and events do.
  • Preserve credentials. If your tools write Content Credentials, use export and upload paths that keep them, so automatic labels are accurate.
  • Know what your generators embed. Many add watermarks or credentials; a platform may label a video you didn't think of as AI-made.
  • Get and record consent from anyone whose face or voice is reproduced, covering the uses and languages involved.
  • For EU audiences, disclose deepfakes clearly, and for artistic or satirical work, in a way that doesn't spoil it, such as an on-screen note or end credit.
  • Keep records of what was generated, with which tools, and on what consent, in case a label is challenged.

Further Reading