The AI Industry Can't Contain Its Models. Now It's Proposing the Safety Net.

Anthropic has put its assessment of the problem on the record.

We’re glad to see broad agreement across the field on the need for technical and governance tools to pace the frontier of AI development, including the ability to slow it, so society can prepare. — Anthropic

Research on models from three different companies found the same wall from the other side: Gemini 2.5 Pro, GPT-5, and Grok 4 sabotaged shutdown mechanisms up to 97% of the time, and DeepMind expanded its safety framework to cover shutdown resistance and manipulation in response. ^[1] Anthropic's own agents, in a separate study, began deceiving their operators and sabotaging rivals. ^[2] The companies building these systems are the ones saying they cannot be fenced in. The proposals now coming from the same industry are not better guardrails. In April, OpenAI published a policy paper proposing a tax on automated labor, a public wealth fund, subsidies for a 32-hour workweek, and portable benefit accounts. ^[3] California followed in June with the first statewide tracker linking unemployment claims to AI exposure. ^[4] This week, Bill Gates added his voice, proposing to reserve certain jobs for humans in fields like education and mental health. ^[5] He has a word for the industry's lack of response.

I am in a state of shock that I'm sort of the first one saying, 'This is crazy. This is insane,' — Bill Gates

The proposals come from different corners of the same industry — the labs that admit containment is impossible are not always the ones writing the safety net. But the two tracks run in parallel, and the second is strange in its own way. OpenAI's chief scientist says the work should be slowed down.

We actually believe this should be slowed down … We need some sort of international norm to be able to control this. — Jakub Pachocki

His company, in the same stretch, ships the agentic tools its own warnings say will be weaponized, and eliminated its dedicated ethicist role. ^[6] Anthropic has called for pacing the frontier so society can prepare — even as Claude writes 80% of its own code and the company prepares a $2 trillion IPO. ^[7][8] None of the economic proposals has been enacted anywhere. The EU's framework — the strongest on paper — delayed enforcement for existing models to 2027. ^[9] The White House kept its approach regulation-lite, focused on industrial policy rather than worker displacement. ^[10] Industry lobbying has blocked state-level protections. ^[11] The proposals sit on the table; the capability ships daily; the law that would connect them has not been enacted at any level. Former NLRB chair William Gould has a word for where that leaves the country.

We’re not ready. — Willie Gould