This opinion piece argues that as artificial intelligence systems become more powerful, the safety testing used to probe their risks can itself cause real-world harm, and that current accountability mechanisms are not equipped to deal with this properly. The author warns that reflexively punishing testers for damage caused during experiments could backfire, much as forcing carmakers to crash-test vehicles only at low speeds would limit understanding of how they behave in genuinely dangerous conditions.

The piece draws a direct analogy between AI safety testing and automotive crash testing, suggesting that overly cautious constraints on how frontier AI systems are tested could prevent researchers from learning how these systems behave under realistic, high-risk conditions. It calls for a proper framework, akin to an insurance model, to manage liability when AI tests go wrong, allowing rigorous safety testing to continue while ensuring victims of any resulting damage are fairly compensated and testers remain accountable.