The AI Safety Menu

All laws

Watermark AI content

Watermarks on AI-generated content. Hidden labels on AI-generated images, video and audio, plus a free detection tool.

Lead costAlmost noneof America's lead over China, over 3 years
No p(doom) effectNonefrom a 5% starting estimate
Holds up○ Training○ Lab's own use○ Public release
Enact it on the menuSee the findings

What it does

AI companies must embed hidden, machine-readable labels in the images, video and audio their models generate, offer a visible label, and provide a free tool to check whether content is AI-made.

Applies to: Generative AI providers with large user bases; in the EU, all general-purpose model providers.

Where things stand

Major labs sit on the steering committee of C2PA, the industry provenance standard, and many already attach Content Credentials to generated images.

Why it costs almost no lead

Labs that generate media add labels using standards they already help build. The work displaces under a day of research and holds up no model.

Biggest unknown: How much engineering reliable watermarking takes, especially for text in the EU.

Its effect on p(doom)

Helps people tell real photos, video and audio from AI fakes.

It targets fraud and misinformation, not catastrophic risk.

The strongest case that it costs more

Robust watermarking of text is an unsolved problem, and the EU rule covers text. If regulators demand watermarks that survive editing, labs could spend far more engineering time than a few days.

The debate

For

Against

Sources

Rough starting points, not precise forecasts. Lead costs assume China doesn't depend on U.S. models, the case least favorable to safety laws, and count 3 years. On the menu you can change every assumption and put in your own numbers. Last priced 2026-09-26.