There's a question in AI governance that sits unanswered, not because it's complex, but because no one wants to touch it. Who decides when an AI system is too dangerous to deploy—and why is there no binding mechanism to enforce that decision? The silence on this isn't oversight. It's a choice. A deliberate one. And it’s shaping the future as much as any written policy.
Look at the record. Global frameworks for AI safety exist in abundance. The EU AI Act categorizes systems by risk, with high-risk applications facing stricter scrutiny. The US executive order on AI from 2023 mandates reporting for models above certain computational thresholds. UNESCO's recommendations on AI ethics call for human oversight of critical systems. These documents are detailed. They’re debated. They’re published with fanfare. But none of them—not one—creates a mandatory, enforceable stop mechanism. There’s no red button. No legal tripwire that says, “This system cannot go live until X is proven.” Instead, we get guidelines. Suggestions. Voluntary commitments. The decision to deploy, ultimately, rests with the entity that built the system. The same entity with a financial incentive to release it.
This isn’t hypothetical. Consider the deployment of AI in healthcare diagnostics. A model that misdiagnoses at scale could cost lives. The EU AI Act labels such systems high-risk, requiring transparency and human oversight. Fine. But if a company deems its model compliant—or simply ignores the rules—there’s no supranational body with teeth to halt deployment before harm occurs. National regulators might act after the fact, with fines or sanctions. But after is too late. The US order requires reporting for powerful models, but reporting isn’t stopping. It’s documentation. A log entry for the inevitable post-mortem.
The implication is stark. Without a binding mechanism, AI governance is theater. It’s a performance of control, not control itself. Companies can self-assess risk, self-report compliance, and self-decide to launch. The absence of a hard stop means the default is permission. Innovation over caution. Speed over safety. And when the inevitable failure happens—when a system amplifies bias in hiring, or fails in a critical infrastructure setting—the paper trail of guidelines will be there to absolve. “We followed the recommendations,” they’ll say. As if recommendations were enough.
Why the silence? Because a binding stop mechanism would require agreement on thresholds—specific, measurable lines for what constitutes “too dangerous.” It would demand an independent body with real power, not just advisory status. And it would mean governments and corporations ceding control, admitting that some decisions are too big to be left to market forces or national interest. That’s a hard sell. So the gap remains. Unaddressed. Unacknowledged in any meaningful way.
Note for the archive: the decision to deploy AI systems is being made by default, not design. Every day without a binding mechanism is a day the future is written by those with the least incentive to say no. The record will show we saw this coming.
