The Logic of Post-Hoc Prevention
Andrew Bailey, the Governor of the Bank of England, suggests that regulating artificial intelligence immediately might be a mistake. He advocates for a period of rigorous testing first, a technical approach to safety that prioritizes empirical data over legal constraints. While this sounds sensible in the context of high-frequency trading or liquidity ratios, applying this logic to frontier AI shifts the burden of proof away from the developers and onto the public. It assumes that we can wait for the system to break before we decide how to fix it.
History shows that this waiting game has a body count. In the early 20th century, chemical manufacturing operated under a similar ethos of experimentation. It wasn't until disasters like the Minamata disease outbreaks in Japan—caused by industrial mercury dumping—that stringent environmental regulations were enforced. In those cases, the "testing" phase happened in the lungs and livers of thousands of people. When tech executives call for a period of observation, they are essentially asking for permission to deploy unproven systems in the real world, treating human societies as the experimental group.
The Mirage of Morally Binding Agreements
Current discussions often revolve around "morally binding" agreements among top AI labs. These are voluntary pledges to prioritize safety and transparency. On paper, they look like progress. In practice, they function as a shield for massive capital. By creating a regime of self-regulation, tech hegemony ensures that the rules of the game are written by the players themselves. This bypasses the democratic process entirely, moving the conversation from the halls of parliament to the private boardrooms of Silicon Valley.
Self-regulation lacks the teeth of statutory law. A moral commitment does not carry the weight of a court order or a civil penalty. When an algorithm causes harm, a company can apologize, update its code, and move on. There is no mechanism for accountability that restores what was lost. This approach protects the business model while leaving the social consequences to be managed by the victims.
What is Left Out of the Safety Debate
The term "frontier AI" typically refers to the most advanced large language models (LLMs) and their theoretical risks, such as self-governance or autonomous agency. While these distant possibilities occupy the minds of policymakers, they distract from the tangible harms already occurring. The conversation is skewed toward hypothetical sci-fi scenarios, leaving behind the immediate reality of algorithmic bias and labor exploitation.
In the Global South, the human cost is already being tallied. Vast numbers of workers in countries like Kenya and the Philippines perform the grueling, repetitive task of data labeling to make AI models function. They are often paid pennies to filter out violent or traumatic content, enduring psychological distress without the protection of robust labor laws. Meanwhile, automated systems used in credit scoring, policing, and social welfare distribution are baking historical prejudices into digital stone. If we wait for "rigorous testing" to finish, we are essentially saying that these existing injustices are acceptable collateral damage in the pursuit of technical refinement.
Technological Safeguards versus Democratic Oversight
Tech companies often argue that technical safeguards—mathematical proofs of alignment or sandboxed testing environments—are sufficient to manage risk. This is a category error. A technical safeguard can prevent a model from saying something offensive, but it cannot decide whether that model should be used to automate mass layoffs or manipulate democratic elections. Technical fixes address the "how" of a machine, but they cannot answer the "why" or the "should" of societal governance.
The industry's "move fast and break things" ethos is fundamentally at odds with the slow, deliberate nature of human rights protections. Rights require deliberation, debate, and the inclusion of marginalized voices—people who are rarely present at high-level summits on AI safety. When we prioritize technical optimization over legal frameworks, we trade the stability of human dignity for the speed of software deployment. We must ask if the push for testing is a genuine safety measure or a tactical delay intended to prevent the implementation of rights-based legislation that would actually hold these companies accountable.
The Danger of the Data Point
When an AI system fails, the industry views it as a data point—a signal to be used for further training and optimization. But for a family denied a mortgage due to a biased algorithm, or a person whose job was replaced by an unvetted tool, the failure isn't data. It is a life-altering event. Treating human suffering as a mere feedback loop is a cold way to manage progress. By the time the "unexpected behaviors" are documented, the harm is already done. We cannot regulate through observation after the damage has become structural.
