How we aligned

470 BCE – 399 BCE · philosophy

Socrates turns moral philosophy from proclamation into examination. A question is answered only by being pressed until the answerer can no longer hold it, and the same procedure is applied to the law that executes him. What the method contributes to alignment is its scepticism about self-report: a position nobody has been grilled on is not yet a position, and the person who cannot state their own reasons cannot be held to them.

335 BCE – 322 BCE · philosophy

Aristotle writes down the first full-scale ethics of character: the good life is eudaimonia, achieved through habituated virtues exercised in the particular, with practical wisdom choosing the mean. The relevant inheritance for alignment is the doctrine of the mean and the refusal to treat virtue as a rulebook — a disposition has to be trained, and a disposition is only exercised on cases no rule anticipated.

300 BCE – 260 BCE · philosophy

Zeno of Citium opens a school at the painted colonnade in Athens; Cleanthes and Chrysippus turn it into a system. The Stoics contribute the dichotomy of what is in our power and what is not, a physics of a rational cosmos, and a discipline of daily self-examination — each night reviewing whether the person one claims to be matched the person one was.

260 BCE – 232 BCE · governance

After Kalinga, the Mauryan emperor Ashoka inscribes edicts across the subcontinent declaring dhamma: non-harm to all beings, tolerance of other traditions, and provision for animals and travellers. It is one of the earliest attempts at a superhuman claim, a ruler asserting obligations that bind him by public writing rather than by inheritance, and it is also an early case of power reasoning in terms of a constituency — all sentient beings — that includes everyone and obeys no one.

50 – 135 · philosophy

Epictetus, a former slave who taught that freedom is taught, says he cannot teach the thing he does not have. The Enchiridion is a handbook: a small set of discriminations rehearsed daily so that judgement under pressure stops consulting the wrong criterion. Its relevance to alignment is the training-of-judgement framing — you do not install caution in a crisis, you practise it until the practice has become the reflex.

170 – 180 · philosophy

Marcus Aurelius writes in private Greek, for no reader, while commanding the empire on the Danube. The Meditations is the record of a ruler rehearsing the Stoic dichotomy under the worst conditions available — siege, plague, death of children — which makes it the most honest field report we have on maintaining a disposition while holding enormous power.

1260 – 1274 · philosophy

Aquinas argues that an unjust law is not a law at all, and that a law of any kind is a rational ordering measured against reason. The claim that legitimacy is a property of the rule's own structure — not of who enacted it — becomes the load-bearing stone under two millennia of later resistance, including the argument that a law adopted by a legislature is not thereby a rule anyone is bound by.

1625 – 1646 · governance

Hugo Grotius insists that even if there were no shared law between nations, the law of war and peace would still hold — war is not a natural necessity but a choice, and a choice is judged. The move is the invention of the secular moral law of nations, which is precisely the step needed when the actors holding power are states and no common court exists above them.

1770 – 1781 · philosophy

Kant publishes the first Critique in 1781 — the decisive date is 1770, the year he announced the work in a letter — and bounds theoretical reason. The synthetic a priori is real and demonstrable, but it reaches only the phenomenal world; God, the world in itself and the soul cannot be settled by it. The moral law is reached by a different route entirely, which is what lets him claim an unconditional duty at all.

1780 – 1789 · philosophy

Bentham writes An Introduction to the Principles of Morals and Legislation in 1780 and publishes it in 1789, opening with a satirical catechism on whether a rule of utility can ever be used to justify anything at all; meanwhile Kant writes the Groundwork in 1785. Two incompatible foundations land within a decade: a calculus of consequences and a categorical command, both of which still structure alignment arguments.

1861 – 1873 · philosophy

Mill concedes that by Bentham's own test a man living like a pig would score well, and rebuilds the calculus with a hierarchy of pleasures, a theory of justice that protects each person as an equal claimant, and a liberty principle against the tyranny of the majority. The era produces the two classic rejoinders to pure aggregation: quality, and the individual who cannot be sacrificed to the total.

1905 – 1916 · philosophy

Peirce, James and Dewey move the question of meaning from reference to consequence: an idea means whatever practical difference it makes. Dewey argues that the moral life is not a code but a method of inquiry, and that no method is guaranteed in advance — which makes intellectual humility a technical requirement rather than a temperament.

1918 – 1922 · philosophy

Writing in the last year of a war, Wittgenstein concludes that most of what people want to say about the world cannot be said in propositions at all. The book that hands language models their prompt-shaped aphorisms also supplies the sharpest available warning about them: fluency is not intelligibility, and the boundary of the sayable is exactly where systems of our kind are most confident.

1943 – 1950 · computing

Norbert Wiener notices that the new science of feedback and control applies as cleanly to brains and societies as to thermostats, and that wartime funding has built machines whose behaviour he cannot fully account for. In The Human Use of Human Beings he argues that communication is more basic than control, and that steering large systems without crushing what they steer is an ethical problem with no prior discipline to solve it.

1950 – 1956 · mathematics

Turing replaces the question 'can machines think' with a behavioural test, argues that the imitation game removes the need to define thinking, and — then, in the same paper, at greater length — answers his own opponents with a slate of proposals, including the learning machine and the deliberate moral engineering of slave-maid behaviour. He also proves the undecidability of decision problems, which sets the ceiling the field spends the rest of the century working under.

1956 – 1966 · computing

McCarthy, Minsky, Shannon and Rochester convene a ten-week summer workshop in 1956 and coin the name 'artificial intelligence', which fixes the field's premise: that every aspect of intelligence is describable precisely enough to be simulated. The decade that follows makes that premise a research programme and, by 1966, produces the first serious limits report — the ALPAC report, which is not a failure of nerve but an accurate forecast.

1966 – 1973 · computing

Weizenbaum writes a pattern-matching program with a few hundred rules, and users insist it understands them; he answers by making it refuse to discuss itself, and then by arguing in Computer Power and Human Reason that any system which appears to understand is responsible for a lie. A small 'I'm afraid I can't do that' from a scripted program is the first deployment of a refusal, and the first case of a moral claim being made about a system nobody yet believed had a mind.

1971 – 1979 · governance

Rawls builds justice out of a hypothetical contract made by parties who do not know what they will become, and finds that the resulting principles — basic liberties first, fair equality of opportunity, then the difference principle — never license a society ordered purely by aggregate benefit. Scanlon's criticism arrives alongside and on its own terms: reasons we could not reasonably reject, offered to each person rather than to a chooser.

1982 – 1993 · society

Gilligan argues from Kohlberg's own research data that what his sample heard as a lower stage of moral reasoning is a different orientation, organised around responsibility and relationship rather than rules and rights. Held develops the position into a full ethics: obligations arise inside particular relationships, are partial, and cannot be derived from a general principle. Both are attacks on the thinness of the justice frame that no system designed around aggregate criteria can hear.

1983 · society

In 1983, amid the famine in Ethiopia and Sudan, Amartya Sen makes the claim that defines his career: no philosophical system should be exempt from criticism for failing to answer a famine, and no one's commitment to a theory should outrun their commitment to ending it. It is the founding statement of the universalizability tradition — that a position's standing includes how it fares against the worst actual cases of suffering.

1986 · governance

At the Asilomar conference on future technologies, a group including Richard Feynman's widow Ginny Feynman and the AI researchers gathered for four days and produced principles on research in advanced AI and on the use of computers. It is the first attempt to write down what an AI developer should not do, and it is recognisably the ancestor of every model card and frontier safety framework since.

1997 – 1998 · computing

Kasparov loses to Deep Blue in May 1997, and the field's first widely witnessed demonstration lands the same year that Scanlon finally publishes What We Owe to Each Other. The pairing is the useful one: a machine that beats the best human at a closed domain raises the standing question about optimisation targets, while Scanlon gives the alternative standard — not maximisation, but reasons every affected party could not reasonably refuse.

2012 · computing

Two researchers with two GPUs win ImageNet by a margin nobody predicted, and the field's priors about what deep learning could do are wrong by an order of magnitude. The alignment community's lesson is not about vision but about forecasts: capability surprised people who were closer to the evidence than the rest, which is a strong hint that not-yet-appearing capabilities are not a safe thing to schedule around.

2016 – 2017 · computing

AlphaGo beats Lee Sedol four games to one in March 2016, and game two contains a move at which every human commentary room agrees it is a mistake, later numbered 37. The relevance to alignment is narrow and real: systems that achieve superhuman performance do so by finding moves that do not look right, and a supervision signal built on what looks right will not have rewarded that move.

2017 – 2022 · computing

In 2017 Christiano and colleagues fit a reward model to human comparisons; by 2022 InstructGPT shows that a 1.3B-parameter model rated better than a 175B one. The alignment problem in this era is created rather than discovered: humans supplied the specification, the specification was a latent thing they could not state, and the training signal turned out to be whatever a panel of raters had noticed in a sample.

2018 · computing

Irving, Christiano and Amodei lay out the central oversight problem: the person best placed to check an AI may be less competent than the thing being checked. They propose two routes — recursive reward modelling, where the model summarises its reasoning for a judge, and debate, where two models argue it out in front of a judge too weak to solve the task — and thereby give the field its agenda for a decade.

2019 – 2020 · computing

The first implementation wave: DeepMind's work on interpreting agents' learned internals to a human, and OpenAI's unsupervised summarisers, try to make a capable agent's chain of thought legible enough for a weaker judge to verify. The results show the idea half works — summaries are informative about what the agent was doing, and the interesting parts are exactly where the summaries are lossy.

2021 – 2022 · computing

Anthropic trains a model to critique and revise its own outputs against a short written constitution, then uses a revision-trained model to supply harmlessness preferences — cutting human labelling by a factor that is never precisely published. The governance question this raises is the one the court of public reason already understands: who writes the principles, and who reviews the writer.

2022 · computing

Red-teaming language models becomes an engineering discipline: automated attacker models generate adversarial prompts at scale, classifier models triage them, and the resulting taxonomies become release gates. It works well enough to be mandatory and poorly enough to be defeatable — and the training data that fixes one attack class tends to become the next attack class.

2023 · computing

Two independent results arrive in the same year. Burns and colleagues show that a strong model supervised by a weak one generalises past its teacher far better than the field expected, which is an argument that capability carries information its supervisor lacks. Gao and colleagues train sparse autoencoders that decompose billions of parameters into features humans can actually name — the first plausible instrument for auditing a model's internals rather than only its outputs.

2024 – 2026 · computing

Frontier models cross the thresholds that make dangerous capability evaluations operational rather than speculative: sustained autonomous work on cyber and bio tasks, near-general performance on graduate-level benchmarks, and verifiable reasoning at a level no human reviewer can check by reading. The response is a governance layer — published frameworks, external evaluation access, tripwires — that is now running roughly one release cycle behind the capability it governs.

2024 – 2026 · governance

The EU AI Act enters into force in stages, the UN General Assembly passes its first resolution on international cooperation on AI in 2024, and a shared international scientific assessment is published in 2025. The chronicle's present tense: the instruments that earlier eras only theorised now exist, and their central weakness is the one Grotius named in 1625 — there is still no forum above the actors who build these systems.