The US needs to tighten its approach to regulating artificial intelligence, shifting from a disclosure regime to mandatory safety checks for the most powerful AI models. This proposal was made by Anthropic CEO Dario Amodei in his essay.
According to him, it’s time to move from transparency to more serious and mandatory AI regulation. Alongside the essay, Anthropic presented two documents: the Advanced AI Framework on oversight of frontier models and the Economic Policy Framework on preparing the economy for potential AI-driven worker displacement.
External Checks and Right to Block
Amodei proposed mandatory testing for models that exceed a certain threshold of computational resources. In Anthropic’s framework, this refers to systems trained using more than 10²⁵ floating-point operations if the developer earns more than $500 million annually from AI or spends over $1 billion on research and development in this area.
According to his plan, qualified external organizations should conduct the checks. Testing should cover four risk categories:
- cybersecurity;
- creation of biological weapons;
- loss of control over AI systems;
- automated research and development that could accelerate these threats.
If the assessment reveals an unacceptable risk, authorities should have the right to block the launch or restrain the deployment of the model.
Amodei compared the proposed regime to aviation oversight:
“The most powerful AI models, like airplanes, should undergo technical tests and audits.”
Additionally, developers should be required to protect model weights and training infrastructure, conduct regular safety checks, publish risk reports, and promptly report serious incidents.
Labor Market: From Retraining to Basic Income
In the essay, Amodei warned that AI could cause more extensive and prolonged disruptions in the labor market than past technological shifts. Three scenarios are likely: unemployment around 5%, 10%, and an unprecedented level.
Among possible measures, he listed improved data collection on worker displacement, incentives for job retention, retraining programs, and, in more severe scenarios, long-term income support, including unconditional basic income.
According to the company’s document, Anthropic will allocate $200 million to the Economic Futures Research Fund for research and evaluation of economic policy, and will establish a national scholarship program of $150 million for early-career specialists.
Civil Liberties
A separate section of Amodei’s essay focused on civil liberties. He proposed establishing accountability rules for fully autonomous weapons, banning their use within the US, and closing the loophole of purchasing data from brokers for mass surveillance.
Another idea is to ensure citizens have access to AI that is not inferior to that used by the government against them in regulatory or legal procedures. According to Amodei, without such balance, AI could enhance the authorities’ advantage.
The essay was published the day after the launch of Claude Fable 5 and Claude Mythos 5. Anthropic described Fable 5 as its most powerful publicly available model, while Mythos 5 is a version of the same base model with relaxed restrictions in certain areas for a limited group of Project Glasswing partners.
Back in June, members of the Anthropic team observed signs of approaching recursive self-improvement in AI. The company suggested that society might need a mechanism to slow down the development of advanced models so that institutions and safety research can keep pace with progress.
