What precisely would a slowdown of artificial intelligence look like — and who would enforce it — are increasingly central questions as policymakers, industry leaders and researchers debate how best to limit potential harms from rapidly improving AI systems.
Proposals for slowing development are varied and imprecise. Some advocates talk about temporary moratoria on releasing the most capable models, others suggest caps on compute or training budgets, and still others call for longer review and safety testing before public deployment. Each framing reflects different assumptions about where the risks lie and how they can be mitigated, but none offers a clear, universally accepted blueprint.
Measuring and verifying a slowdown
A core technical obstacle is measurement: determining when a model is ‘‘more capable’’ than previous systems is not straightforward. Capabilities grow along many dimensions — raw compute, model architecture, training data quality and downstream fine-tuning — and there is no single metric that captures all of them. That complicates any attempt to define thresholds for limits, to verify compliance, or to detect clandestine development.
Even if robust metrics could be agreed, practical enforcement poses further problems. AI research is globally dispersed across private companies, startups, universities and nation-state projects. Coordinated international agreements would be required to prevent migration of activity to jurisdictions with weaker rules. Observability is also limited: much of the most sophisticated work happens behind closed doors, making audits and inspections politically and technically difficult.
Economic and strategic incentives
There are strong commercial and strategic incentives to continue aggressive development. Firms invest heavily in AI to secure competitive advantage and market share, while governments view advanced AI as a matter of national competitiveness and security. Those incentives create pressure to keep pushing even if voluntary restraints are proposed. Critics warn that unilateral pauses by some actors could simply hand advantage to others that do not comply.
At the same time, proponents of pacing argue that a slowdown — if effectively implemented — could buy time to develop safety standards, testing regimes and regulatory frameworks that are currently lacking. But turning that theoretical benefit into practice would require credible monitoring, enforceable penalties for violations and international buy-in, all of which are hard to achieve quickly.
Alternatives to blunt pauses
Because of these difficulties, many experts are urging more targeted approaches: increased transparency about model capabilities and training processes, mandatory risk assessments, stronger incident reporting, third‑party audits, and restrictions on certain high‑risk applications rather than blanket halts. Others argue for stepped deployment and staged approvals similar to those used in regulated industries.
The conversation is shifting from whether slowing AI makes sense in principle to how a slowdown could be operationalized without creating perverse incentives or leaving gaps that bad actors could exploit. Policymakers face a delicate trade-off between encouraging innovation and protecting the public from harms that are hard to anticipate.
The appeal of a rapid, simple fix remains strong, but the reality is that pacing AI development presents a thicket of definitional, measurement, enforcement and geopolitical challenges. Any effective approach will likely be complex, incremental and dependent on sustained international cooperation rather than a single, one‑off pause.