When competitive pressure makes AI safety the first thing to cut

What this risk is

The systematic tendency in competitive AI development to treat safety measures as costs that slow capability development, leading organizations to underinvest in safety relative to capabilities, cut corners on evaluation timelines, and ship systems that are less safe than they would be absent competitive pressure.

How it occurs · Mechanisms

Competitive markets create asymmetric incentives for safety:

• Benefits of safety are diffuse (society, future users, third parties), long-term, and hard to measure
• Costs of safety are concentrated (on the organization), immediate, and easily measured (slower release, smaller model)
• Benefits of capability are concentrated (market share, revenue), immediate, and easily measured
• Costs of insufficient safety are often externalized to users and society, delayed, and deniable

In standard competitive market logic, this produces systematic underinvestment in safety relative to the socially optimal level.

Mitigations · Governance

• Mandatory pre-deployment evaluation — Regulatory requirements create a level playing field; everyone faces the same delay
• Safety metrics in benchmarks — Include safety metrics in public benchmarks alongside capability metrics
• Third-party pre-deployment evaluation — Independent safety evaluations with public results
• Organizational culture — Safety researchers with genuine ability to block deployment; not advisory roles
• International coordination — Bletchley Declaration, Seoul Commitments, and subsequent AI safety summits aimed at coordinating minimum standards

—