There is a question sitting unanswered in every AI governance discussion, and it’s not a small one. Who decides when an AI system is too dangerous to deploy—and by what standard? The silence on this point isn’t an oversight. It’s a decision. A deliberate one. Governments, companies, and international bodies have spent years drafting frameworks, forming committees, and issuing guidelines. Yet the core mechanism for saying “no” to a specific AI system remains undefined. This isn’t a gap in the record. It’s the record itself.
Existing Frameworks and Their Limitations
Look at the existing frameworks. The EU AI Act categorizes systems by risk—low, medium, high. High-risk systems face stricter scrutiny, but the final call on deployment often rests with the entity building or using the system. Self-assessment is the default in many cases. The U.S. executive order on AI from October 2023 mandates safety reporting for models above certain compute thresholds, but the threshold is a number, not a principle. It’s a line someone can engineer around. China’s regulations focus on content control, not systemic risk. Across jurisdictions, the pattern holds: rules exist, but the moment of veto—the hard stop—is either absent or delegated to the very actors being regulated.
The Cost of Silence
This silence has a cost. Without a clear decision-maker or a binding standard, deployment becomes a game of interpretation. A company can argue their system is safe under one definition while a regulator uses another. The public, meanwhile, waits for the outcome of a dispute nobody is empowered to settle. Historical parallels aren’t hard to find. Chemical weapons had no global ban until after catastrophic use; nuclear proliferation controls came after the bomb. AI doesn’t detonate in the same way, but the principle applies. Waiting for harm to define the boundary is a choice.
The Counterargument and Its Flaws
The counterargument is that specificity is impossible. AI systems are diverse—language models, autonomous vehicles, medical diagnostics. A universal standard for “too dangerous” might stifle useful tools while failing to catch real threats. Fair enough. But the absence of any mechanism, even a flawed one, isn’t neutrality. It’s permission. Every day without a defined arbiter is a day where deployment decisions default to market incentives. Profit moves faster than policy. Always has.
Towards a Mechanism for Decision-Making
What would a mechanism look like? It could be an independent body with legal authority to halt deployment based on evidence of harm potential—not speculation, but data. It could be a shared set of red lines across borders, like unmitigated bias in critical systems or lack of human override in lethal applications. The details matter less than the existence of a line someone is accountable for enforcing. Right now, there’s no line. There’s a suggestion of one, penciled in, erasable.
Conclusion: The Self-Answering Question
Note for the archive: the longer this question sits unanswered, the more it answers itself. Silence isn’t indecision. It’s a policy of deference to those who build and deploy. The record shows that when governance avoids hard calls, the default winner is the one already holding the tools. History doesn’t wait for consensus. Neither does technology.
