Prove it's safe before launch
Safety case before deployment. Write a structured argument that the model is safe to release, and have it reviewed, before you release it.
What it does
Before release, the lab writes a structured argument, backed by evidence, that the model is safe enough to launch. A reviewer has to approve it.
Applies to: Frontier models before external deployment.
- Before external deployment, produce a safety case: a structured argument, with evidence, that the model's risks are below defined levels given its safeguards.
- Submit it to a regulator or independent reviewer and receive sign-off before release.
Where things stand
Anthropic's Risk Reports are the closest published artifact. Under RSP v3.4 they go to external reviewers within a week of the Board, and reviewers are asked to respond within 30 days. OpenAI's Capabilities and Safeguards Reports serve a similar internal role. No lab currently waits for an outside sign-off before launch.
Why it costs ~2 days
Writing and reviewing the case takes senior time and holds up the public launch. The lab's own progress loses under a day; if China depends on U.S. models, the delayed launch slows China more.
Biggest unknown: Nobody has published how long a frontier safety case takes to write or review.
Why it lowers p(doom) by ~0.3%
Makes a lab show a model is safe enough before release, instead of critics having to show it isn't.
Shifting the burden of proof is a big change: the lab has to argue why the model is safe enough, not just report that tests passed.
The strongest case that it costs more
We have no data on how long these take. AISI has written about safety cases as a method; no lab has published one for a frontier launch with a timeline. A regulator's review could take far longer than 30 days the first few times, and a lab that fails review restarts the clock. If the practical review tail is six weeks rather than two, this item costs a month and a half of release delay per generation, and if the lab holds internal rollout for the sign-off, it costs the same on the U.S. clock.
The debate
For
- Clymer, Krueger and colleagues proposed safety cases to guide AI deployment decisions, 2024.
- UK AI Security Institute builds safety case templates, 2024.
- GovAI argued safety cases can support both self-regulation and government oversight, 2024.
- Anthropic said some of its commitments take the form of affirmative safety cases, 2024.
Sources
- Anthropic RSP v3.4: external review of Risk Reports: shared within one week; comments within 30 days
- AISI, How can safety cases be used to help with frontier AI safety?
- OpenAI Preparedness Framework v2: Capabilities and Safeguards Reports
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.