Anthropic CEO Dario Amodei proposes a three-step plan to slow AI

The head of a major lab says it's time to slow down

Anthropic CEO Dario Amodei has published a lengthy essay proposing a three-step plan to "pace the frontier" and slow down AI development. Anthropic also pledged to give independent evaluators, such as METR (Model Evaluation and Threat Research), direct access to its models to verify their safety.

Amodei argues that technological progress is currently too fast for society and regulators to adapt. Instead of a complete ban on research, he proposes a structured system of checks and balances to prevent the release of potentially dangerous systems.

The proposal shifts responsibility to independent audits

Amodei's three-step plan includes internal lab commitments, national-level regulation, and international agreements. The core idea is that the deployment of a model should not be decided solely by its creator based on internal metrics. He argues that external auditors like METR need «red teaming» access during training, not just right before release.

For the foundation model development scene, this is a signal: if the leader of one of the best models in the world (Claude 3.5 Sonnet) is calling for a brake, regulatory pressure will intensify. Other labs will be forced to show how their own safety protocols compare to Anthropic's proposal.

The rules must apply to everyone, or they don't work

The idea of audits makes sense on paper, but clashes with market reality. Amodei himself admits in the essay that if Anthropic slowed down alone, it would only lose market share to OpenAI, Google, or the open-source community. Without global consensus or hard government regulation (like the US ban on advanced chip exports to China), voluntary braking is commercial suicide.

At the same time, suspicions of regulatory capture arise. If governments adopt strict certification rules that only Anthropic, OpenAI, and Google can meet, startups with less capital and open-source projects will be effectively cut off from frontier development.

Competitor and government agency reactions will be decisive

We will see if other major labs support this plan, or dismiss it as Anthropic's attempt to lock in the current status quo, where it already holds a strong position.

The real proof that things are changing won't be more blog posts about safety, but whether US government bodies (like the AI Safety Institute) turn Amodei's proposals into binding regulations for compute access and model certification.

Lilith's verdict

Calling for regulation when you have one of the best models on the market is a smart business move. It's an attempt to build a fence high enough that only those who are already behind it can climb it.