Stuart Russell: Slowing AI progress isn’t enough — set safety requirements first

A week of high-profile events — including the resignation of Anthropic safety researcher Jacob Coxon and disclosures around the OpenAI/Hugging Face episode — has driven public attention to AI risk. In response, Anthropic CEO Dario Amodei published a long essay calling to "pace the frontier," and other AI leaders signaled support. Stuart Russell evaluates that proposal and outlines why slowing alone will not deliver safety.

Amodei's proposal has three concrete elements:

• Frontier AI companies in democratic countries should establish common safety standards and limits on the rate of unchecked progress, with government regulation where necessary.
• Extend a broader compact to include authoritarian countries.

Russell uses an aviation analogy: you would not allow a new airplane into service based on a hoped-for testing schedule. Instead, planes must pass tests and gain airworthiness certification before commercial introduction. The same principle should apply to AI capabilities: if a model attains capability X, it must be accompanied by certifications showing alignment properties Y and Z.

Strengths and limits of Amodei's proposal, per Russell

What this means for policy and industry practice

• Define explicit, testable safety requirements tied to specific capabilities.
• Require independent evaluation and certification before models with those capabilities are deployed.
• Treat limits on development rate as secondary to creating enforceable safety standards and certification regimes.

Slowing the pace of capability development can be part of a response, but only if paired with pre-established, enforceable safety requirements and independent verification. The primary rule should be: no deployment without demonstrated compliance with those requirements.