OpenAI scientist urges slowdown in AI development

OpenAI’s chief scientist warned in recent public remarks that AI labs may need to slow development so safety and alignment research can catch up. He described a deliberate pause or slowdown as a way to lower risks from misuse, unexpected behavior and an unregulated race to scale compute and model size.

He urged companies and research teams to adopt longer testing periods, stronger red-teaming, phased rollouts and closer coordination with regulators and independent auditors.

He warned that moving straight to ever-larger models without sufficient safety checks increases the chance of harms such as automated disinformation, cyberattacks and unforeseen model behavior.

He framed the issue as a tradeoff between rapid capability gains and the time needed to develop reliable alignment techniques. Minor failures in current systems can be managed, he noted, but very large models operating at scale can create new categories of risk that are harder to foresee.

He recommended expanding investment in alignment research, publishing model evaluations and creating shared testing standards that labs can use to assess harms before wide release.

He identified pressures that push labs toward faster development: competition for talent, incentives tied to product releases and the value placed on headline performance benchmarks. He called for clearer government rules setting minimum safety requirements and for industry practices that reduce incentives for destructive acceleration.

Technical proposals included better interpretability tools, more robust adversarial testing and stronger methods to ensure models follow intended goals. Procedural proposals included staged deployment plans, independent audits of high-risk systems and formal mechanisms for sharing information about failure modes and near misses. He suggested that smaller, focused pauses tied to specific testing milestones would be more practical than indefinite moratoriums.

He acknowledged the economic and scientific costs of slowing development, noting many organizations rely on ongoing model improvements to fund research and operations. He recommended targeted, time-limited slowdowns focused on high-risk systems and clear criteria for when development can resume, such as demonstrated advances in alignment or completion of standard safety tests.

Companies, universities and governments have used staged releases, safety reviews and voluntary guidelines to reduce immediate harms while continuing research. Regulators in several countries are exploring legal frameworks for high-risk AI systems. OpenAI has used phased access and safety mitigations while investing in internal safety teams, and practices across the industry vary on openness versus pre-release evaluation.

He concluded by describing slowing development as a way to buy time to build safety tools and governance, and he urged work on practical, verifiable safeguards so advanced systems are deployed with clearer limits and known risks.