Trump's AI Safety Accord Relies on Self-Policing—a Gamble Without Guardrails

The White House convened major AI executives this week to celebrate a new industry safety accord, with the president assuring observers that self-regulation would work. But this faith in voluntary compliance ignores a crucial historical lesson: when industries self-police, people tend to get hurt.

The Seat Belt Precedent#

In the 1960s, nearly 51,000 Americans died on highways annually, yet the automotive industry resisted adopting newly available safety measures. Rather than waiting for carmakers to do the right thing, Congress stepped in. The 1966 National Traffic and Motor Vehicle Safety Act gave federal authorities power to set binding safety standards, and seat belt requirements followed by law in 1968. Over six decades of vastly increased traffic, highway deaths have plummeted—not because companies voluntarily chose safety, but because regulators forced their hand.

The Current AI Accord's Weakness#

The recently announced AI safety accord essentially asks tech companies to monitor themselves. As one observer noted, the agreement lacks meaningful enforcement mechanisms. The accord depends on AI labs to "keep themselves in check," with limited external oversight or consequences for violations.

The administration's position is even starker. Officials have suggested the Department of Justice and FBI provide sufficient oversight, and that companies will naturally "police themselves" because healthy competition demands it. This echoes a bygone era in industrial regulation.

The Complexity of AI Risk#

The analogy between cars and AI systems has limits. There's no single technological fix—no "AI seat belt"—that demonstrably reduces harm overnight. Unlike highway fatalities, the nature and severity of AI-related risks remain contested: credible experts debate everything from near-term harms (such as AI systems generating false intelligence that nearly caused a military incident) to existential scenarios that critics dismiss as speculation.

Yet uncertainty itself is not an argument for inaction. The consensus among AI researchers and executives includes acknowledgment that some form of regulation is necessary. Bill Gates and others have issued warnings about potential catastrophic outcomes. Even OpenAI has delayed or canceled advanced model releases due to safety concerns—a step that voluntary compliance might encourage, but legal requirements would guarantee.