Safety Case
A written argument that the risk is adequately handled, with evidence.
A structured argument that a system is safe enough for its intended use, made of explicit claims, the evidence supporting them, and the assumptions being relied on. Borrowed from safety-critical engineering and now written for frontier models.
The useful property is falsifiability: because claims and assumptions are written down, they can be attacked individually. External scrutiny then targets the weakest link rather than the whole document, which is the point of the format.
The most useful thing a reader can do with one is attack the assumptions, because that is where safety arguments are typically weakest: an assumption that a capability is absent, or that monitoring would detect a behaviour, is often the load-bearing claim.
Finding the load-bearing assumption
Reviewing a safety case before approving a deployment.
For each claim, ask what has to be true for the evidence to support it, then ask how that was established. The weak link is usually a statement about what the system cannot do, which is substantially harder to evidence than what it can.
- reportAnnouncing our updated Responsible Scaling Policy ↗
Capability thresholds with graduated safeguards and documentation duties, written in the safety-case style: here is the evidence that this risk is adequately handled.
- reportUpdating the Frontier Safety Framework ↗
DeepMind’s front door for its frontier-risk framework of critical capability levels and mitigations. Treat as the official entry point and read the linked report for thresholds.
Responsible Scaling Policy · RSP
Capability thresholds that trigger heavier safeguards before release.
Red Teaming
Adversarially trying to make the system do the wrong thing.
Model Card
The documentation shipped with a model.
Evaluation · Evals
How you know the system is actually working.
Interpretability · Explainability
Working out why the model did that.