The AI Safety Menu

All laws

Pay for training data

Consent to train on copyrighted work. No training on copyrighted works or personal data without clear, affirmative consent.

Cuts lead by~3 monthsof 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 get clear permission, usually a paid license, before training on copyrighted books, articles, art or personal data. Anyone whose work is used without consent can sue.

Applies to: Training of generative AI systems on copyrighted works or personal data.

Where things stand

Labs have begun licensing some data from publishers and paid to settle past claims. Anthropic's $1.5 billion settlement with authors covered about 482,000 pirated books and only past conduct. Most training data is still used without explicit consent.

Why it costs ~3 months

Removing unlicensed data means training on less, or repeated, data. The model estimates about a month of lost progress a year in.

Biggest unknown: How much training data labs could license quickly, and how much capability depends on the rest.

Its effect on p(doom)

Protects writers, artists and other creators whose work trains AI.

A copyright and fairness rule. It doesn't target catastrophic risk.

The strongest case that it costs more

Chinese labs would not follow U.S. consent rules. If the rule removes a large share of high-quality text and code, the capability hit could persist for years, not a single training run.

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.