Will it take a disaster to force AI regulation? | The Economist

Political Triggers and Regulatory Trajectory

• Significant government action is predicted to occur only after a major catalyst, described as an "AI Chernobyl," involving either serious financial losses or significant loss of life.
• Current political dynamics are viewed as stagnant due to a lack of trust in American leadership's competence to handle existential risks.
• Regulation is already occurring incrementally through "regulation by the back door," specifically via a de facto licensing regime where advanced model releases require weekly government consultation.
• A patchwork of state-level legislation is currently building momentum, with laws passed in New York and California and liability bills under discussion in Rhode Island.
• There is a risk that a catastrophic event could trigger an irrational, heavy-handed regulatory overreaction rather than a proportionate response.

Strategic Competition: US vs. China

• The US is expected to maintain its lead over China because Chinese labs rely heavily on distilling and replicating US research rather than independently breaking new ground.
• A three-to-four-month slowdown by the US could allow China to catch up to the current frontier, citing recent releases like the open-source GLM 5.3 system.
• The GLM 5.3 system is projected to reach capability levels comparable to Anthropic's March models regarding potential damage in cybersecurity domains.
• The primary concern is not that China will overtake the US, but that it will reach the dangerous capabilities currently held by US labs that are already releasing harmful models to trusted partners.
• There is skepticism regarding the efficacy of "stronger AI to deal with weaker AI harms," characterized by the speaker as "swallowing a spider to catch a fly."
• The speaker argues that slowing development is the only method to stop other entities from possessing damaging models, despite the inherent difficulty of the "accelerationist" race dynamic.

Feasibility of Safety-First Approaches

• It is feasible for the US to shift the balance toward safety without losing its technological edge, provided a light-touch regulatory regime is implemented.
• Unlike a total ban, the proposed path involves a regulatory shift that tips the scale toward safety while maintaining the US position relative to China.
• Historical precedents, such as the automotive and medical industries, suggest that safety regulations evolve incrementally following mistakes, rather than being instituted immediately.
• The window for effective, non-catastrophic intervention is narrowing, with the next 18 months expected to be "seismic" for both political debate and technological evolution.
• The US is urged to avoid the binary choice between the status quo and draconian measures, instead pursuing a calibrated approach that demonstrates costly commitments to safety.