Competing AI firms tend to release frontier models prematurely, creating safety externalities that markets ignore; minimum quality standards can restore the social optimum, whereas forced release delays can perversely worsen outcomes by shifting preemption to earlier announcement choices.

Main Finding

Firms racing to be first to deploy frontier AI models choose release times that are earlier than socially optimal because they capture first-mover advantages while not internalizing safety harms their premature releases impose on society. Some policy tools (minimum quality standards, credible voluntary commitments) can restore cooperative outcomes; others (mandatory delays) can backfire by shifting the race to the announcement stage; Pigouvian safety taxes help but cannot fully remove the preemption distortion on their own.

Key Points

• Model setup: competing firms improve model quality with additional development time; releasing earlier yields first-mover benefits but increases societal safety risk (a negative externality not borne fully by firms).
• Equilibrium vs. social optimum: the symmetric Nash equilibrium of the preemption game yields strictly earlier releases than the social planner’s optimum.
• Mechanism: firms trade off incremental quality improvements against the risk of being preempted; because benefits of being first concentrate to the firm while costs are socialized, equilibrium is biased toward earlier release.
• Policy interventions analyzed:
• Minimum quality standards (performance/safety thresholds) — can implement the first-best outcome when designed to restrict releases below the socially inefficient preemption point.
• Mandatory release delays — can reduce realized model quality by moving the competitive pressure to the announcement decision (firms announce earlier and lock in lower quality before the mandated delay ends).
• Voluntary safety commitments — can sustain cooperative (delayed, safer) equilibria if commitments are observable and credible (enforceable reputationally or contractually).
• Pigouvian safety taxes — internalize part of the externality and mitigate premature release, but cannot fully eliminate the strategic preemption distortion in general.
• Generality: results apply beyond AI to technologies where quality increases with development time, first-mover advantages exist, and premature deployment imposes safety externalities (e.g., certain biotech or robotics applications).
• Comparative statics (informal): stronger first-mover advantages, larger safety externalities, or weaker observability/credibility of commitments increase the gap between private equilibrium and social optimum.

Data & Methods

• Approach: theoretical game-theoretic model of development and release timing with symmetric competing firms. Quality is an increasing function of development time; firms choose when (and possibly whether) to release.
• Strategic structure: a preemption (race) game with an equilibrium concept of symmetric Nash equilibrium in release timing; also consider announcement-stage strategies and commitment devices.
• Analytical results: characterization of equilibrium release timing and comparison with social planner’s solution; proposition proofs show equilibrium is strictly earlier than the social optimum and derive conditions under which each policy instrument achieves which outcomes (including proofs that delays can worsen outcomes by shifting preemption).
• Policy analysis: modeled interventions as constraints or additional payoff terms (minimum quality cutoff, mandated waiting periods between announcement and deployment, instruments that make commitments observable/credible, and taxes proportional to expected societal safety cost).
• Robustness: arguments and variants demonstrate results do not rely on knife-edge parameter choices; qualitative conclusions hold across reasonable functional forms and information structures considered in the paper.

Implications for AI Economics

• Regulatory design matters: simple instruments can have counterintuitive effects in strategic settings. Minimum quality/safety standards are potentially powerful, but must be set correctly and enforceable; blanket delays risk perverse incentives unless paired with measures preventing announcement-stage preemption.
• Importance of observability and credibility: policies or market institutions that make safety commitments observable and enforceable (auditable benchmarks, third-party certification, binding contracts) can facilitate cooperative delay and higher deployed quality.
• Partial interventions: Pigouvian-style taxes or fees are useful but, by themselves, unlikely to fully correct the strategic race; combining taxes with standards or commitment mechanisms will be more effective.
• Empirical predictions and monitoring: expect clustering of releases, strategic pre-announcements, and pre-announcement quality locking when mandatory delays exist. These are observable markers regulators and researchers can use to detect harmful preemption dynamics.
• Broader relevance: the framework applies to other frontier technologies—insights transfer to regulation of risky innovations where first-mover advantage and safety externalities coexist.
• Policy takeaway: prioritize enforceable minimum quality/safety thresholds and mechanisms that make commitments credible; be cautious relying solely on delays or taxes, and monitor for announcement-stage strategic behavior.