Slowing down AI development is technically possible — but coordinating it globally is turning out to be one of the hardest problems in tech policy today. As calls for a more deliberate pace have grown louder this year, they’ve run straight into the same wall that’s stalled similar efforts before: nobody wants to slow down first.
The Proposal That Reignited This Debate
The current wave of discussion traces back to an essay from Anthropic CEO Dario Amodei, who argued the AI industry needs a more deliberate pace of frontier model development to give safety work room to keep up with rapidly advancing capabilities. The proposal was publicly endorsed by OpenAI CEO Sam Altman and SpaceX founder Elon Musk — a notable show of agreement among competitors who rarely align publicly. But even Amodei acknowledged the proposal’s central weakness in a CBS News interview: what happens if geopolitical rivals simply don’t adopt the same restraint?
Four Ways a Slowdown Could Actually Happen
| Approach | How It Works | Main Limitation |
|---|---|---|
| Voluntary restraint by AI labs | Companies deliberately pace their own development | A lab that slows down loses market share to competitors who don’t |
| National regulation | Governments impose testing, disclosure, or compute-threshold rules | Only binds companies operating within that country’s jurisdiction |
| Export controls on hardware | Restricting access to advanced chips | Concentrates development rather than slowing it; incentivizes rivals to build domestic alternatives |
| International agreement | Treaty-style coordination between countries | Extremely difficult to verify what’s happening inside a data center or training run |
Why This Is Harder Than Past Arms Control
AI slowdown proposals are frequently compared to nuclear arms control, but the comparison reveals exactly why AI is harder to govern. Nuclear facilities are physically large, hard to hide, and observable from satellites — inspectors can verify compliance without needing a country’s cooperation. A large AI training run, by contrast, happens inside ordinary-looking data centers using commercially available chips, with no equivalent way to verify from the outside what a company or government is actually building. Add in that AI development is driven primarily by private companies rather than states, and the coordination problem multiplies: any agreement needs buy-in not just from governments, but from competing corporations with billions of dollars and market position on the line.
The Geopolitical Layer Makes It Messier Still
Recent events show just how differently “slowdown” gets interpreted depending on who’s talking. China’s state-backed Global Times dismissed Amodei’s proposal as a “Cold War” strategy, arguing it was really designed to preserve US technological dominance while excluding China from global AI governance — not a genuine safety measure. Meanwhile, in the US, President Trump rejected calls to slow AI development from a completely different angle, dismissing safety concerns as coming from “very negative forces” and framing continued rapid development as a competitive necessity, arguing “whoever wins AI wins.”
That leaves safety-focused voices caught between two very different objections — one calling the same proposal a competitive weapon, the other calling it unnecessary caution — even though both objections work against actually implementing any real slowdown.
What’s Driving the Urgency Behind These Calls
The push for a more careful pace hasn’t come from nowhere. Two Anthropic researchers have separately warned that rapidly advancing AI could pose an existential risk to humanity within the coming years, with one publicly estimating roughly a 10% chance of AI causing human extinction within the next decade — arguing that major labs, Anthropic included, are racing toward increasingly autonomous, self-improving systems without adequate safeguards keeping pace. Those warnings have prompted US lawmakers to call for new regulatory frameworks, adding domestic political pressure to a debate that’s already tangled up in international competition.
So What’s Actually Achievable?
A complete, verifiable global slowdown looks very unlikely under current conditions — the economic incentives to keep advancing are enormous, verification is genuinely difficult, and no major player currently trusts the others enough to go first. What looks more realistic is narrower, targeted friction rather than a full stop:
- Mandatory safety testing before deployment of the most capable models, rather than pacing development itself
- Restrictions on specific high-risk applications — autonomous weapons systems or certain biological research uses, for example — rather than blanket capability limits
- Industry norms and reporting requirements that raise the cost of reckless deployment without halting the underlying research
- Continued export controls, which slow who can build frontier systems even if they don’t slow the technology’s overall advance
Final Thoughts
The debate over slowing AI development has exposed just how differently “safety,” “competition,” and “geopolitics” get tangled together once a proposal moves from a research lab’s essay into the real world of national rivalries and corporate incentives. A clean, coordinated slowdown remains unlikely — but the pressure building around targeted safeguards, testing requirements, and use-case restrictions suggests the more realistic outcome isn’t a pause on AI itself, so much as a slower, more contested tightening of the rules around how it gets built and deployed.
