The Fallacy of the Incremental Approach
Andrew Bailey, Governor of the Bank of England, argues that regulating AI before its full capabilities are understood is a mistake. His stance favors a "test-first, regulate-later" paradigm, suggesting that heavy-handed laws might stifle the very innovation needed to make the technology safe. He views the current era as a technical proving ground where mistakes serve as data points for future governance. However, this logic assumes that the errors produced during this testing phase are merely technical glitches or mathematical inefficiencies. It ignores the fact that when AI fails, the consequences are not just bits and bytes; they are lives, livelihoods, and fundamental human rights.
By framing the issue as a matter of technical maturity, Bailey shifts the conversation away from social impact and toward corporate efficiency. This approach treats AI development like a software beta test. But you cannot run a beta test on a society. You cannot pause a person's life to see if an algorithm makes a better decision regarding their medical care or their legal status. The delay inherent in Bailey's incrementalism creates a vacuum where power concentrates in a few hands while the risks are socialized among the most vulnerable.
The Hidden Labor of Frontier Models
Big Tech companies speak of "frontier AI" as if these models emerge from pure mathematical genius. They do not mention the human hands that clean the data. Beneath the polished interface of a Large Language Model (LLM) lies a massive, invisible workforce. In the Global South, thousands of workers in Kenya, the Philippines, and India spend ten hours a day labeling graphic imagery, toxic text, and violent content to train these systems. These workers often labor in precarious conditions with little to no psychological support for the trauma they absorb.
If we follow the "test-first" logic, we are essentially saying that the psychological toll on these data annotators is an acceptable cost of innovation. Self-regulation by tech giants offers no real protection for these workers. A "morally binding" agreement from a trillion-dollar corporation is not a substitute for enforceable labor laws. When companies prioritize speed to market, the supply chain of data becomes a new frontier of exploitation, one that a slow, incremental regulatory approach will likely miss entirely.
Wealth Disparity and the Illusion of Benefit
Bailey often points to the "immense benefits" AI might bring to the global economy. This is a broad, optimistic abstraction. For a hedge fund manager, a more efficient predictive algorithm is a tool for profit. For a gig worker, the same technology might be a tool for surveillance and wage suppression. The "benefits" of AI are not distributed evenly. Without immediate guardrails, the technology acts as a massive accelerant for existing wealth inequality.
We are seeing the erosion of labor rights in real-time. Automation does not just replace repetitive tasks; it optimizes away the human element that provides leverage to workers. If regulation lags behind deployment, by the time we decide how to tax or redistribute the gains from AI, the social fabric may already be torn. The gap between those who own the compute and those who are displaced by it is widening faster than any committee can meet.
Bias as a Weapon of Systematic Exclusion
Technologists often frame algorithmic bias as a "technical vulnerability" to be fixed with better datasets or more diverse training sets. This is a narrow view. Bias is not just a math problem; it is a human rights problem. When biased algorithms are used in predictive policing (using historical arrest data that reflects systemic racism) or in healthcare triaging, they become engines of discrimination.
In migration control, AI-driven facial recognition and sentiment analysis can determine who is "suspicious" and who is
