Do AI Laws Require a Watermark on AI Images and Video?
Chapter 5, part 1 of 5 · 5 min read — by HALLMARK.AI
Mostly no. Laws in the EU, China, South Korea and India require AI-made images and video to be labelled or marked. But in the sources we reviewed, none clearly requires an invisible watermark specifically. Metadata that a machine can read satisfies the statutes. The duty usually sits on the company that builds or runs the AI tool, not on the person who posts the result. (This is general information drawn from the published texts and law-firm summaries. It is not legal advice.)
What "machine-readable marking" means
A machine-readable mark is information software can detect without a person looking at the picture. The two common forms are metadata, a record attached to the file such as C2PA Content Credentials, and an invisible watermark, a signal hidden in the pixels or audio. The laws ask for the first outcome, detectability, and mostly do not say which form to use. Metadata is easy to strip when a file is re-uploaded, which is why some people argue for both. We compare them in C2PA vs invisible watermarking.
Where the law stands, by place
- EU. Article 50 of the AI Act requires providers of generative systems to mark output in a machine-readable way, "effective, interoperable, robust and reliable" as far as feasible. It applies from 2 August 2026 for new systems and 2 December 2026 for systems already on the market, though one law firm disputes that grace date. Fines go up to EUR 15 million or 3%, and small companies pay the lower figure. Our Article 50 post covers the text.
- California. SB 942 covers generators with over one million monthly users: a free detection tool and a hidden disclosure in image, video and audio, from 2 August 2026. A second law, AB 853, makes large platforms read and show provenance data from 1 January 2027.
- New York. Advertisers must visibly disclose a synthetic performer in an ad, since 9 June 2026.
- China. Generators and platforms must add a visible label plus metadata. A watermark is asked for only "where feasible". In force since 1 September 2025.
- South Korea. AI businesses, including overseas ones, must label generative output. The label may be machine-readable, but realistic deepfakes need a mark people can clearly see or hear. Fines of up to KRW 30 million come after a grace period of at least a year.
- India. Platforms must show a prominent label, add metadata where technically feasible, and not strip it. The penalty is loss of safe harbour, the legal protection platforms have for user content.
- Vietnam and Kazakhstan are also reported to make a machine-readable mark mandatory for providers in 2026. We could not confirm that against the laws themselves, so treat it as unconfirmed.
The closest thing to a watermark rule
The EU Code of Practice on AI-generated content, finalised on 10 June 2026, comes nearest. It is voluntary. It says that where one marking method is not enough, providers should use at least two layers, for example signed metadata plus a watermark. It also expects a free detection tool and interoperable watermark detection by 2 February 2027. Published signatory counts conflict. A voluntary code is not a mandate.
Who the duty falls on
In the EU, the marking duty is on the provider of the generative system, any size, including companies outside the EU if the output is used there. A separate duty on deployers, anyone using AI commercially, is to disclose deepfakes. Personal use is exempt.
The tool makers do not always take it on. Higgsfield's terms say it "may embed" marks and does not warrant that any marking will be applied to, or persist in, every output. Its terms also make the user solely responsible for disclosure. Model-hosting services such as Replicate, fal.ai and Together say in their terms that they have no marking duty and the customer is responsible. So check what your tool actually does rather than assuming it marks your files.
Where enforcement has happened
Only in China, so far. Authorities issued warnings, rectification orders and app-store removals, and named apps including CapCut in April 2026. No fines were disclosed. In the EU we found no Article 50 fine, order or investigation, and five member states were reported to still have no enforcing authority in October. We found none in the US, Korea or India. That can change, and the first EU action would be the signal to watch.
What a creator using AI tools should do
- Use the disclosure toggle when a platform asks. YouTube, TikTok, Meta and Pinterest all offer a free one or label automatically. TikTok rejects or restricts ads with undisclosed AI content.
- If you run ads with synthetic people or deepfake-style content, disclose it visibly. Do not rely on a hidden mark.
- Keep records: which tool, which prompt or source files, and the date. Clients increasingly ask for this in contracts.
- Do not sign or tag filmed footage as AI-made. Metadata can trigger automatic AI labels, and YouTube says so.
- Know that a mark helps prove where a file came from. It does not stop anyone copying it. See do watermarks survive AI editing.