Most AI debates are anchored in the past or the present. Which model is ahead right now? Who missed the last release cycle? Dario Amodei, CEO of Anthropic, has chosen a different vantage point: What happens when the curve just keeps climbing, and no one in government is remotely prepared for it?
His essay „Policy on the AI Exponential” is not a hype document. It’s a sober attempt to name the gap between the pace of technological change and the capacity of political institutions to respond to it. And that gap, Amodei argues, is widening faster than most governments realize.
The central thesis is both straightforward and uncomfortable: we don’t need a brake driven by fear — we need one driven by reason. Evaluations, veto mechanisms, employment safety nets. Not to obstruct innovation, but to make it sustainable.
Amodei published the essay “The Adolescence of Technology” on January 26, 2026, on his website. (Dario Amodei)
The Exponential Curve as a Political Problem
Technological curves are usually an engineer’s problem. This time, they’re everyone’s.
Amodei’s core argument: AI systems don’t improve linearly — they improve exponentially. That means political institutions designed around annual cycles and legislative terms are structurally ill-equipped to keep up. By the time a bill has cleared three committees and two readings, the very model it was meant to regulate has already changed twice over.
This isn’t a new observation. What makes it significant is who’s making it — someone who runs one of the most capable AI laboratories in the world. Amodei isn’t watching from the sidelines. He’s both a participant and a warning voice.
What he’s calling for is not a moratorium. He proposes that governments introduce binding evaluation processes before certain capability thresholds are crossed. Anyone training a model that meets specific risk profiles should be required to have it tested — externally, not internally. That might sound like bureaucratic logic. It’s actually engineering logic: before you raise the scaffolding, you check the ground.
What Tests and Veto Rights Actually Mean in Practice
Amodei’s proposal centers on what he calls “pre-deployment evaluations” — assessments of models against specific risk categories before they are released, covering areas such as the potential to assist in developing biological weapons or compromising critical infrastructure.
In “Policy on the AI Exponential,” Amodei identifies four evaluation categories: “cybersecurity, biological weapons, loss of control of AI systems, and automated R&D that could accelerate these other risks.” (Dario Amodei)
This may read like security policy, but it’s equally innovation policy. Standardizing these tests creates clarity. Companies whose models don’t reach the relevant thresholds can deploy faster. Those that do must pause. That’s not a blanket prohibition — it’s a traffic light system.
The veto mechanism Amodei sketches goes a step further. Governments should have the ability to halt or delay certain deployments when risk assessments warrant it — not permanently, not arbitrarily, but as a last resort within a tiered framework.
Those who read this as government overreach are missing the essential point: without such instruments, no democratically elected body has any lever at all. The decision about when a model becomes too powerful would rest entirely with the labs themselves. That’s not a market. That’s an oligopoly operating on self-regulatory promises.
The Employment Safety Net Argument: Social Policy as a Safeguard for Innovation
The third pillar of Amodei’s thinking is perhaps the most unexpected. He argues that economic disruption caused by AI will generate political backlash if left unaddressed — not because disruption is inherently bad, but because when …
