Anthropic CEO Dario Amodei has called for a deliberate slowdown in AI development, announcing that the company will grant third-party evaluators access to its models as part of a three-step plan to ensure safety practices keep pace with innovation. In a detailed essay, Amodei urged the industry to 'pace the frontier' — slowing training and development to allow time for safeguards and regulation.
Anthropic CEO Dario Amodei has taken an unusual step in the fast-moving AI sector, publicly urging a deceleration of development. In a winding essay published this week, Amodei proposed a three-step framework he calls 'pacing the frontier,' a term he uses to describe slowing the pace of training and deployment to give companies and regulators time to build adequate safety measures.
The first step, which Anthropic is already implementing unilaterally, involves giving third-party evaluators like METR (Model Evaluation and Threat Research) wide-ranging access to its models. This access is intended to verify the company's 'adherence to safety practices and commitments,' according to Amodei.
The second step would require broader industry coordination, likely with government involvement, to establish shared safety standards and evaluation protocols. The third step remains unspecified in available reports but is part of a broader push for collective action.
Amodei's call for restraint comes amid a period of intense competition in AI development, with companies racing to release more powerful models. The Anthropic CEO's position contrasts with that of some rivals who have pushed for rapid advancement, arguing that speed is necessary to remain competitive and address pressing global challenges.
The Verge notes that the announcement follows a week in which Anthropic faced scrutiny over its cybersecurity practices, adding context to Amodei's emphasis on safety transparency.
Analysis
Why This Matters
- This could set a precedent for AI companies voluntarily slowing development and opening up to external scrutiny, shifting the industry away from an unbridled 'race to the top'.
- If adopted widely, it may influence how regulators approach AI governance, potentially leading to mandatory third-party evaluations.
- The proposal challenges the dominant narrative that speed of innovation is paramount, introducing safety and deliberation as core competitive values.
Background
The AI industry has grown rapidly, with companies like OpenAI, Google, and Anthropic releasing increasingly capable models. Concerns about safety, bias, and misuse have prompted calls for regulation, but industry adoption of voluntary safeguards has been uneven. Anthropic, founded by former OpenAI employees, has positioned itself as a safety-focused AI company, though it has also faced its own controversies, including recent cybersecurity concerns.
Key Perspectives
[Anthropic and its CEO Dario Amodei]: Argues that slowing down is necessary for safety, allowing time for proper testing and oversight. Sees third-party evaluation as a way to build trust and accountability.
[AI industry rivals (e.g., OpenAI, Google, Meta)]: May view slowdowns as competitive disadvantages, especially if competitors in other countries (e.g., China) continue rapid development. Could push back against voluntary restraints without global coordination.
[Regulators and safety advocates]: Likely to welcome the proposal as a step toward responsible development, but may question whether voluntary measures are enough without binding regulation.
What to Watch
- Whether other major AI labs follow Anthropic's lead and also grant third-party evaluators access to their models.
- Any government response, particularly from the U.S. or EU, that codifies these voluntary measures into law or regulation.
- The timeline for step two (industry-wide coordination) and whether it materializes before or after the next major model release from a leading lab.