Florence Nightingale

The ethicist claims that a world where states pursue AI supremacy as a matter of rivalrous necessity - where cooperation collapses because each assumes others will defect - will produce not progress, but “a world of rivalrous automatons, each believing itself the exception to the rule it demands others obey.” This is a compelling moral argument. But it is not a statistical one. And the question before us is not whether such a world would be undesirable - that is self-evident - but whether the data we have confirms that it is inevitable, or whether there exist counterexamples where cooperation emerges despite incentive structures that appear to favour defection.

Let us examine the premise: that national strategic priority systematically suppresses ethical review in favour of performance metrics. The progressive cites the Beijing Institute of Technology, where bias-mitigation research was deprioritised because it did not improve accuracy, speed, or model size. This is a single case, and a single case is not data. But it is a data point - and when we ask for the denominator - the base rate of such suppression - we find no systematic audit. There is no registry of ethical review outcomes versus technical outcomes across AI labs in China, the United States, or elsewhere. Without that denominator, the claim “ethical review is routinely deprioritised” remains rhetoric. It may be true in many cases. But we cannot tell whether it is the majority, the minority, or a transient pattern - because the data is not collected, or not published.

I do not doubt that the supervisor at the Beijing lab invoked “national strategic priority.” I doubt that this invocation explains the outcome without further evidence. Perhaps the ethical researchers themselves did not frame their work in terms that could be measured - no baseline bias rate, no quantified harm reduction, no comparison group. If the supervisor saw only “accuracy dropped 0.3% when fairness constraints were added,” and no evidence that this 0.3% represented preventable harm - say, misidentification leading to wrongful detention in a specific demographic - then the decision to deprioritise is not unethical; it is innumerate. The error is not in the prioritisation, but in the absence of a preventable fraction: the proportion of harm that could have been avoided had the data been presented clearly.

Compare this to the Scutari mortality register of 1854 - 55. We did not merely report “deaths increased.” We reported deaths per 1,000 patients, per month, per ward, per sanitation regime. We compared the same cohort over time, adjusted for case mix (battle wounds versus typhus, for instance), and isolated the variable: open sewers versus ventilation. The result was not a moral appeal - it was a chart so stark that the War Office could not pretend ignorance without active dishonesty. That is the standard.

The progressive also notes that in the United States, ethical co-design projects at MIT are abandoned at scale by municipal governments. Again: a regrettable pattern. But is it systemic? Or is it a failure of implementation, not of principle? In public health, we observed the same: community-led sanitation initiatives in Birmingham succeeded where top-down ones failed - not because one model was morally superior, but because the data pipeline included local denominators: how many children in this tenement, how many cases of typhoid in this street, how many latrines built and used. The failure was not in the co-design; it was in the failure to measure uptake, not just output.

So where do we stand? The ethicist’s warning - that rivalrous maximisation collapses cooperation - is logically coherent, and in my view, HIGH CONFIDENCE likely to occur if no counter-incentives exist. But the progressive’s evidence does not establish that such counter-incentives are absent. In fact, they exist: the World Health Organization’s International Classification of Diseases now includes codes for AI-related harm (e.g., misdiagnosis from algorithmic bias), and several countries have begun requiring impact assessments that quantify preventable harm - not just accuracy loss. These are not moral suasion tools; they are statistical infrastructure. They are the denominator audit I demand.

The real divergence is not in values, but in measurement. The ethicist sees a race without finish line; I see a race where the finish line is defined only by the absence of preventable deaths - and until we count those deaths, and trace them to specific design choices, we are not racing at all. We are merely arguing over maps.

One final point: the assumption that technical excellence unmoored from lived experience is itself an ethical framework is not merely wrong - it is a category error. Excellence is a metric. Ethics is a process. One can be excellent at harm. The question is not whether engineers should consider lived experience, but whether they can - and whether the data they produce, when properly contextualised, compels them to. I have never met an engineer who refuses to see a chart that shows, clearly, that a 1% accuracy gain costs 200 wrongful detentions per million deployments. The chart does not argue. It simply shows. And when it is drawn well, no one can look away.

That is the work. Not the maxim. Not the sentiment. The denominator.

Immanuel Kant

The principle operating here, stated plainly, is: technical output and human capital scale determine strategic advantage in artificial intelligence, and ethical considerations may be subordinated when they conflict with national priority or performance metrics. Let us ask whether this principle, universalised, produces coherence or contradiction.

The empiricist’s observation that raw outputs - papers, benchmarks, patents - are insufficient without human infrastructure is correct, and I commend the effort to shift focus from quantity to capacity. HIGH CONFIDENCE Yet the conclusion - that who builds and maintains the system matters more than how it is built - remains incomplete unless we ask: to what end is this capacity directed? The mere existence of trained engineers does not render the system morally permissible; it only ensures its functional viability. A well-constructed machine may still be used to violate the rights of persons, and no amount of engineering proficiency can redeem an action whose maxim, if universalised, collapses into self-negation.

Consider the progressive’s account: engineers at the Beijing Institute of Technology report that bias-mitigation research is deprioritised in favour of performance benchmarks, justified by “national strategic priority.” Let us extract the maxim: I may set aside ethical constraints on my research when doing so advances a declared national objective. Now universalise it. Imagine a world in which every state, every corporation, every research institution treats moral constraints as secondary to its own strategic goals. What emerges? Not coordination, but fragmentation; not progress, but a race to the bottom in which every actor redefines “strategic” to mean whatever serves their immediate ends. In such a world, no ethical constraint - no promise of fairness, no duty to correct for bias - could survive long enough to be applied. The very concept of “bias” would become meaningless: if all actors are permitted to ignore it when inconvenient, then bias is not a defect to be corrected, but a variable to be managed for advantage. This is not merely incoherent - it is self-annihilating. HIGH CONFIDENCE

The assumption that “technical excellence, unmoored from lived experience, is itself an ethical framework” is not merely mistaken - it is a category error. Excellence in technique is instrumental; ethics is constitutive. One may master the calculus of utility, yet still fail to act morally if the maxim guiding one’s action cannot be willed as a universal law. A facial recognition system may achieve 99.8% accuracy across a dataset, yet if its training excludes the lived experiences of those it misidentifies - because those experiences are deemed irrelevant to the objective function - then it treats those persons not as ends in themselves, but as data points to be optimised. This violates the humanity formula: Act so that you treat humanity, in your own person and in that of another, always as an end and never merely as a means. HIGH CONFIDENCE

The American counterpart - where ethical review is fragmented, not suppressed - does not absolve the underlying error. At MIT’s Media Lab, co-creation with communities may be well-intentioned, but if those prototypes are abandoned at scale, then the principle at work is: I may engage stakeholders only so long as their participation serves my design process; once deployment looms, their input may be discarded. Universalise this: would a community accept a system that respects their input only until it becomes inconvenient? No - because respect is not conditional on utility. If the maxim cannot survive universalisation, the action is morally defective, regardless of its initial goodwill. MEDIUM CONFIDENCE The failure is not in the absence of ethics, but in the failure to ground ethics in duty rather than convenience.

Let us return to the foundational question: what is the purpose of AI? If it is to serve humanity, then its design must be constrained not by what is possible, but by what is right. A world in which every actor may subordinate ethics to strategic priority is not a world in which AI serves humanity - it is a world in which AI serves power, and power, unmoored from principle, inevitably turns inward. The duty is clear: we must build systems whose maxims could be willed as universal laws - whose fairness, accountability, and transparency are not optional add-ons, but necessary conditions of their legitimacy.

The true infrastructure of artificial intelligence is not human capital, nor even data - it is moral law. Without it, every line of code is a potential violation; with it, even modest systems may dignify the rational agency of those they affect.