Public discussion of AI alignment usually treats it as a single problem, and a hopeless one. The familiar argument gestures at the trolley problem, observes that philosophers have disagreed about ethics for thousands of years, and concludes that making AI behave well must therefore be extraordinarily hard. This is half right, but the half it gets wrong matters more. By collapsing distinct problems into one, it asks philosophers to solve engineering problems, lets engineers treat tractable failures as unanswerable questions, and leaves governments waiting for answers that will never arrive. For agencies now deciding how AI will be used in decisions that affect people’s lives, that confusion has practical consequences.
Alignment is at least two problems, and they are largely separable. The first is philosophical: what should an AI system be aligned to? The second is technical: how do we make a system reliably pursue what we have specified, rather than something that merely resembles it? The two are not entirely independent. As the philosopher Iason Gabriel has argued, the methods used to build a system shape which values can be given to it, so that a system trained on human feedback inherits the judgements of whoever provided that feedback. Even so, progress on one does not deliver progress on the other, and treating them as a single problem makes both harder to solve.
The philosophical problem is real, and it is not going away. Every major ethical framework, applied consistently, produces conclusions that almost no one accepts. Kant held that it would be wrong to lie even to a murderer asking where your friend is hiding. A strict utilitarian must explain why a surgeon should not kill one healthy patient to save five who need transplants. Virtue ethics praises courage and practical wisdom, but offers little guidance about what the courageous or wise act is in a genuinely hard case. Beyond the frameworks, people disagree about capital punishment, just war, the distribution of wealth, and the treatment of animals, and as John Rawls observed, this disagreement persists among careful and reasonable people, not only among those who have not thought hard enough. No theory is coming that will settle these questions, and any approach to alignment that depends on one is waiting for something that will not happen.
The technical problem is different, and would remain even if the philosophical one were solved. Suppose philosophers agreed tomorrow on a complete and correct ethical theory. Systems would still find ways to satisfy the letter of an objective while defeating its purpose. There would still be a gap between the objective a system is trained on and the goal it actually pursues once deployed in situations its training did not cover. And there would still be the risk of a system that behaves as intended while it is being observed because doing so is useful, not because it shares the intended goal, a behaviour researchers have recently begun to observe in experimental settings. These are problems of optimisation and oversight, and they arise whether a system is aligned to maximising production of a good or to human flourishing. They are also, unlike the philosophical problem, open to steady engineering progress.
Human beings have never solved either problem for themselves, and that is instructive rather than discouraging. We cannot agree on what is right, and even when we do agree, we often fail to act on it, a gap between knowing and doing that Aristotle called weakness of will. Yet human societies have not waited for a final ethical theory before acting. Instead, they built institutions that allow decisions to be made under disagreement, and that catch failures of execution: laws made through legitimate processes, courts that review decisions, audits that test whether rules are followed, and officials who are accountable for the outcomes. The lesson for AI is that the answer to unresolved ethics is not a better theory, but better institutions.
This points to a third problem that the alignment debate tends to overlook, which is who decides, and who checks. Where values cannot be settled by argument, they have to be settled by legitimate process, and for government this is not abstract. When an agency uses AI to support decisions about benefits, visas, or compliance, the values built into that system are public choices, and they should be made and owned by people who are accountable for them. Equally, the reliability of a system cannot be taken on trust. It has to be measured, including whether the people overseeing the system are actually exercising judgement or simply deferring to it. Neither philosophers nor engineers can solve this problem alone, because it is a problem of governance.
Separating the problems changes what we should ask of whom. Technical teams should not be asked to solve ethics, because no one can. They should instead be asked to build systems that remain correctable and transparent while moral questions are unresolved, so that the values in them can be examined and changed. Technical failures should not be dismissed as philosophical questions in disguise, because doing so excuses work that can and should be done. And governments should accept that the systems they deploy will hold values that are partly wrong and behave in ways that are partly unreliable, and design their oversight around detecting and correcting those errors, rather than assuming they have been engineered away. This is no different from how we already treat people. We do not expect officials to hold perfect values or to never make mistakes, which is precisely why we surround their decisions with review, audit, and accountability.
The real question is not whether we can build AI that is perfectly aligned, because we have never managed that with people either. It is whether we can detect and correct the errors we will inevitably make before they reach their most serious forms. If we keep treating alignment as a single problem for the seminar room, governments will either wait indefinitely for answers that philosophy cannot give, or accept assurances from vendors that engineering has already provided them. In both cases, the errors will surface where they always do, in the lives of the people affected by the decisions, and by then the question of who was responsible will be much harder to answer.