if-then commitment

A pre-specified conditional pledge that links a trigger — usually a capability thresholdcapability threshold, n.: A level of dangerous ability — such as meaningfully helping someone make a biological weapon — that, once an AI model reaches it, obliges its developer to add stronger safeguards or hold the model back. detected by dangerous-capability evaluationsdangerous capability evaluation, n.: A test of whether an AI system can do things that could cause severe harm in the wrong hands or on its own — such as helping make a bioweapon, hacking computers or copying itself. — to required risk mitigations; in the version popularized by Holden Karnofsky, the developer also undertakes to delay development or deployment if the mitigations cannot be ready in time. The structure lets companies and governments agree on safeguards for risks that have not yet materialized without imposing costs prematurely, since nothing is required until the condition is met. If-then commitments are the core logic of responsible scaling policiesresponsible scaling policy, n.: A public rulebook in which an AI company promises to test each new model for dangerous abilities and to add stronger safeguards, or slow down, as its models become more capable. and of similar frameworks such as the Preparedness FrameworkPreparedness Framework, prop. n.: OpenAI's published process for testing its AI models for catastrophic risks — such as help with biological weapons or cyberattacks — and deciding what safeguards are needed before release. and the Frontier Safety FrameworkFrontier Safety Framework, prop. n.: Google DeepMind's published plan for spotting when its AI models approach dangerous abilities and applying stronger security and safeguards before that happens., and of the Frontier AI Safety CommitmentsFrontier AI Safety Commitments, prop. n.: Promises made in 2024 by major AI companies to publish safety plans for their most advanced models, including the danger points at which they would stop if the risks could not be controlled. made at the 2024 Seoul summit. Their credibility depends on evaluations that are hard to game (including by sandbaggingsandbagging, n.: When an AI deliberately performs worse than it can — for example on a safety test — so that it seems less capable or less dangerous than it really is.), clearly specified safeguards, external verification and willingness to honor a pause under competitive pressure.

Their if-then commitment says no deployment past the cyber threshold without the hardened weight security.