Bernard Williams

In the [YOU] on AI Field Guide

The cycle that [YOU] on AI inaugurates refuses the clean verdict. It does not declare AI good or bad, liberating or dispossessing, the dawn of a new human potential or the twilight of depth. It holds the tension open, describes it with phenomenological precision, and asks what it means to act well in conditions that do not admit of a clean resolution. This refusal is philosophically significant: it is moral honesty about a situation that the morality system—Williams's term for the apparatus of modern ethical theory—is not equipped to describe. Williams's framework explains why the refusal is the right response, not the failure of intellectual nerve it might appear to be.

The specific moral phenomena that Williams named apply to the AI transition with a precision suggesting they were developed for exactly this kind of situation. Agent-regret names what the senior engineer feels who embraces AI, ships extraordinary products, and cannot shake the awareness that in building with tools that have transformed the conditions for developing the next generation's judgment, his hands are on the wheel. Not because his choice was wrong. Not because a reasonable person could criticize it. But because his agency has contributed to an outcome involving genuine loss that no productivity data can discharge. The utilitarian calculates the aggregate and closes the ledger; Williams insists the ledger does not close.

Moral luck names the phenomenon that the AI transition generates on an industrial scale: identical moral qualities—identical dedication, skill, intelligence, professional commitment—producing radically different outcomes depending on whether a practitioner's specific expertise happened to lie on the augmented side of the line or the replaced side. The triumphalist who tells the displaced practitioner to reskill is not wrong about practical necessity; but the moral framing is wrong, because it treats the difference in outcome as a consequence of choice when it is substantially a consequence of luck. The meritocratic mythology of the technology industry—the conviction that outcomes track talent and effort—is precisely the illusion that moral luck exposes.

Moral remainder names the accumulation of losses that survive the justification of the transition—the embodied knowledge lost, the temporal space compressed, the community of shared practice thinned, the habits of truthfulness eroded—losses that constitute a moral debt the justification does not discharge. Williams would have insisted on acknowledging the debt not as a therapeutic exercise but as a condition for adequate response: a builder who knows the weight of what was lost builds differently from one who sees only capability gains, not necessarily more slowly but with a different quality of attention to what the building serves and what it demands of the people who inhabit it.

Origin

Williams was born in Westcliff-on-Sea in 1929 and educated at Balliol College, Oxford, where he won the Craven Scholarship and the Ireland Scholarship before studying philosophy. His early career was shaped by the Oxford ordinary-language tradition, but he quickly diverged from it, developing an interest in ancient Greek ethics that would become the deepest intellectual anchor of his work. He taught at Oxford, Cambridge, and Berkeley, and served as Provost of King's College Cambridge from 1979 to 1987. His range was extraordinary: he wrote with authority on the philosophy of action, personal identity, the history of ethics, the philosophy of science, and political philosophy, as well as the moral philosophy for which he is best known.

His two most important books appeared in 1981 and 1985. Moral Luck collected the essays that introduced his central concepts to philosophical discussion—the paper on moral luck that he wrote jointly with Thomas Nagel, the paper on agent-regret, the paper on persons, character, and morality that introduced the concept of ground projects. Ethics and the Limits of Philosophy was his comprehensive statement of the critique of ethical theory—of what he called the morality system, its demand for completeness, its systematic blindness to the features of moral life that resist systematization. His last major work, Truth and Truthfulness (2002), examined the virtues of accuracy and sincerity as foundational to social life, a framework that reads, posthumously, as an almost uncanny anticipation of the credibility problems posed by language models that produce confident wrongness in the form of impeccable prose.

Williams died in 2003, thirteen years before the first versions of the technology that would make his vocabulary so urgently relevant. He could not have anticipated the specific phenomena of the AI transition. What he had developed, across a career of meticulous argument against the demand for clean resolution, was the philosophical apparatus that the transition demands: a vocabulary for naming genuine moral complexity without resolving it, and a defense of the importance of doing so.

Key Ideas

Agent-regret. The specifically first-personal regret an agent feels for outcomes her agency produced, even when her actions were justified and she would make the same choice again. The lorry driver who kills a child through no fault of his own feels something a bystander does not—not guilt for having done wrong, but agent-regret for having been the causal agent of an unavoidable catastrophe. A utilitarian who experiences no such regret when a justified action produces genuine loss is not admirably rational; she is morally deficient. The concept names precisely the moral weight that [YOU] on AI's builders carry—neither condemning their choice nor pretending the loss is absorbed by the justification.

Moral luck. Moral luck is the uncomfortable fact that moral assessment depends on factors beyond the agent's control. The two drivers who run a red light—one killing a pedestrian, one arriving safely—made identical choices with identical intentions, but the moral assessment is radically different. Williams regarded this as a feature of moral life that theory must acknowledge rather than theorize away. The AI transition instantiates moral luck at scale, distributing radically different outcomes to practitioners of identical moral quality on the basis of the arbitrary alignment of their skills with the technology's capabilities.

Moral remainder. The residue of value that survives the justification of an action—the loss that persists after the ledger has been balanced. When a government floods a village to save a city, the flooding is justified; the village's destruction is nevertheless a genuine loss that the justification does not dissolve. Moral remainder accumulates in any major transition. A theory that treats resolution as total—that declares the account closed when the aggregate is positive—has misdescribed what moral life actually looks like.

Thick ethical concepts. Williams distinguished between thin ethical concepts (“good,” “bad,” “right,” “wrong”) that carry evaluative content but minimal descriptive content, and thick concepts (“courageous,” “callous,” “gracious,” “treacherous”) that fuse description and evaluation, that tell you not only how to assess a situation but what it is. The AI discourse is conducted almost entirely in thin concepts. The moral complexity of the transition demands thick ones, which can only be developed through the communities of practice whose vocabulary is adequate to the specific texture of their specific experience.

Ground projects and integrity. Williams's structural account of integrity centers on the relationship between an agent and the deep commitments that give her life its character—what he called ground projects. These are not preferences that can be abandoned without cost to the self. They are constitutive commitments without which the agent would not recognize her life as her own. The demand that practitioners abandon the craft that constituted their professional identity and embrace the new paradigm without residue is not a demand for adaptability. It is a demand that they become someone else.