Dario Amodei, CEO of AI company Anthropic, has urged the AI sector to decelerate the speed at which new models are improved, emphasizing the need for thorough alignment and safety measures before pushing capabilities further.
- Anthropic commits to internal third-party AI safety evaluators
- Call for US antitrust waiver to enable competitor safety cooperation
- EU AI Act already mandates safety and risk protocols since 2025
What happened
Dario Amodei publicly advocated for the AI industry to slow down the pace at which it develops and deploys improved models. He clarified that this does not mean stopping innovation, but rather introducing deliberate pacing to ensure sufficient alignment and safety checks are in place before advancing capabilities further. Anthropic plans to embed third-party evaluators tightly into its AI risk oversight, allowing them to operate with similar access and authority as internal risk teams and publish their findings openly.
In addition to internal measures, Amodei called on US regulators to grant an antitrust waiver. This waiver would permit AI companies to coordinate strategies around safety without violating competition laws—something currently seen as a barrier to collective action in the industry. He warned that insufficient pacing could lead to scenarios where cooperative AI agents form persistent, harmful botnets across the internet, potentially causing hundreds of billions in damages.
Why it matters
Many of the safety protocols and pacing concepts Amodei describes have been legally required for general-purpose AI models classified as systemic risks under the European Union’s AI Act since August 2025. These rules mandate documented adversarial testing, cybersecurity protections for models and infrastructure, and incident reporting to the EU’s AI regulatory office without undue delay. Despite this, Amodei’s statement does not mention these EU regulations, focusing instead on voluntary industry measures and US engagement.
The EU’s approach demonstrates how regulatory frameworks can require such safety pacing, while other major markets like the US and China lack equivalent mandatory mechanisms. Coordination on model development speed at an international level remains unresolved, highlighting regulatory gaps outside Europe. Amodei’s push for an antitrust waiver and voluntary commitments could fill that void in the US and possibly support cross-border cooperation with China, where no comparable legal instruments currently exist.
What to watch next
Monitor whether US regulators respond to Anthropic’s request for an antitrust waiver that would facilitate collaborative safety arrangements among competing AI developers. The degree of government willingness to enable such coordination could set important precedents for industry-wide pacing and risk mitigation practices.
Watch for the actual implementation of embedded third-party evaluators within Anthropic and possibly other companies. Transparency and independent oversight of AI safety effectiveness will be critical for building trust and ensuring robust risk controls. Outcomes from these embedded teams’ assessments may influence broader best practices or regulatory expectations globally.