The Quiet Standardization of AI Safety Practices
How red-teaming and model evaluation practices are converging across labs.
A shared playbook emerges
Despite competitive pressure between labs, safety evaluation practices — red-teaming for harmful outputs, adversarial testing, staged rollout — have converged into a broadly shared playbook across the industry.
Why convergence happened
Shared pressure from regulators, insurers, and enterprise customers pushed labs toward similar minimum practices, even without formal coordination between them.
What's still inconsistent
Depth and transparency of red-teaming still vary significantly between organizations, and independent verification of safety claims remains the biggest unresolved gap.