The history of artificial intelligence is a history of a moving target. The first machine decision was a thermostat closing a circuit. The field got its name at Dartmouth in 1956 and immediately wrote promissory notes it could not honor. Two winters followed, each one a correction of expectations rather than a verdict on the underlying ideas. Expert systems worked until the knowledge ran out. Statistics quietly won the middle decades. Deep learning won the last one. The transformer turned language into the interface, and suddenly everyone was having the same argument the field has had since 1956, using the same words, with the same confidence that this time the words finally mean what they say.
And through all of it, one pattern held. AI is whatever computers cannot reliably do yet. Once it works, we rename it software and stop being afraid of it.
That pattern is not trivia. It is the single most useful fact available to anyone trying to decide what to do about AI right now, because it tells us which parts of the current moment are permanent and which parts are an artifact of the label. The history does not answer every question. It answers four of them, and the four it answers are the ones policymakers, executives, and engineers keep getting wrong.
Govern the Decision, Not the Technology
Regulation aimed at a technology inherits that technology’s shelf life. Regulation aimed at a decision does not.
Every previous attempt to draw a legal boundary around “artificial intelligence” as a category has run into the naming problem within a few years. The systems the drafters had in mind stop being called AI. The systems that arrive next were not imagined when the definitions were written. A statute or framework that spends its first ten pages defining artificial intelligence has already conceded that its scope will be litigated forever, because the field itself has never sustained a stable definition for longer than a funding cycle. The EU AI Act runs headlong into this, sorting systems into risk tiers by technological category and then discovering, before the ink dried, that general-purpose models required a bolted-on regime of their own because the original taxonomy had no place for them.
The alternative has been sitting in American law for decades. The Equal Credit Opportunity Act does not care whether a loan denial came from a human officer, a scorecard, a random forest, or a language model. It regulates the decision: if you deny someone credit, you owe them specific and accurate reasons, full stop. The requirement was written before anyone involved had heard of backpropagation, and it applies to systems its drafters could not have imagined, precisely because it never tried to name the technology. The NIST AI Risk Management Framework leans the same direction, organizing itself around functions and consequences rather than around any particular architecture.
The lesson generalizes. Decide which decisions matter: credit, hiring, medical triage, sentencing, targeting. Attach obligations of explanation, contest, and accountability to those decisions wherever they are made and whatever makes them. The thermostat and the language model sit on the same continuum, and the law should treat the continuum, not the fashionable segment of it.1
Winters Are Corrections, Not Verdicts
Both AI winters followed the same script. Capability was real but narrower than advertised. Money arrived faster than understanding. The gap between demonstration and deployment went unexamined until it could not be, and then funding collapsed all at once. What died in each winter was the overpromise. What survived was whatever had quietly solved an actual problem: search techniques after the first winter, statistical methods and the neural network research program after the second.
The current boom will follow the same script, because the script is not about technology. It is about the economics of expectation. Somewhere between the capability that exists and the capability that has been sold, there is a gap, and the gap always closes from the sales side.
The practical discipline this history recommends is simple to state and rare in practice: fund and deploy against the problem, not against the technology. The deployments that survived previous corrections shared one trait. Someone could name the specific task being performed, the cost of performing it before, and the measured difference after. The deployments that evaporated shared the opposite trait: the technology was the point, and the problem was recruited afterward to justify it. A great deal of what is currently sold as AI transformation is the second kind, old solved problems wearing a new interface, and when the correction comes, the renaming machinery will run in reverse. Nobody will say the AI failed. They will say the product was bad, and the word AI will quietly detach itself from the wreckage and move on, as it always has.
Organizations that want to be standing after the correction should run the test now rather than waiting for the market to run it for them. If nobody can articulate what the system does that was not being done before, or the articulation collapses into the technology having been impressive, that is the tell.
Every Model Is a Position on How to Live
A system trained on text learns more than grammar. It learns the priors of the civilization that wrote the text: what counts as a good reason, what counts as harm, when deference is a virtue and when it is a failure, whether the individual or the relationship is the unit that matters. These are not parameters anyone set. They are the sediment of the training data, and they surface in exactly the places where the system is being trusted to exercise judgment.
I made a version of this argument about driving earlier this year: an autonomous vehicle’s split-second choices encode somebody’s ethics, chosen upstream, invisible at the point of use. The argument scales. A model built primarily from one philosophical tradition will weigh social harmony against individual assertion differently than a model built from another. It will draw the line between guidance and intrusion in a different place. Two systems can be equally capable, equally aligned by their builders’ standards, and give systematically different answers to the same question, because the question was never technical.
This is why the phrase algorithmic sovereignty is going to matter more every year. When a country, a company, or a person adopts a model, they are importing a value system along with the capability, usually without an inventory of what is in the container. The history of AI offers no precedent for this, because previous generations of the technology did not speak. A decision tree has biases; it does not have a worldview. Language models have worldviews, plural, and choosing among providers is now partly a choice among philosophical traditions embedded at a depth no procurement checklist currently reaches. The question to ask of any consequential deployment is no longer only whether the model is accurate. It is whose values are running, and whether anyone checked.
Three Professions, Each Half Right
The AI argument is really three professions talking past each other, and the history shows each of them holding one correct insight and one correctable error.
The engineers are right that the capability is real. Anyone who lived through the pattern-recognition plateau of the 1990s and then watched ImageNet fall knows the difference between hype and a phase change, and this is a phase change. The engineering error is believing that because the capability is real, deployment is an engineering problem. It is not. Deployment is a systems problem, an incentives problem, and a values problem, and the model is the component of that system easiest to build and hardest to blame.
The economists are right that AI decomposes work into tasks rather than consuming jobs whole, and that substitution and complementarity will do what they have done through every previous technology transition, unevenly and with real casualties whose pain is not refuted by the aggregate statistics. The economic error is treating intelligence as a homogeneous input, a quantity of which machines now supply more at lower cost. The history argues otherwise: every era of AI automated a different narrow slice of cognition and left the rest untouched, and there is no reason to believe the current slice, wide as it is, is the whole loaf.
The philosophers are right that the value questions are load bearing and cannot be delegated to the loss function. Alignment, fairness, and accountability are not engineering constraints to be satisfied; they are contested moral terrain that the systems now occupy whether anyone resolved the contests or not. The philosophical error is the tempo. Deployment is not waiting for the seminar to conclude. The values questions are being settled right now, by default, in the training data and the product decisions, and a discipline that prefers its answers rigorous and late is ceding the field to answers that are sloppy and early.
The useful posture takes the correct half from each: the capability is real, the economics are task-level, and the values are load bearing, and any analysis missing one of the three legs falls over.
The Pattern Is the Instruction
I have spent a career deploying these systems in domains where being wrong has consequences: forecasting political instability, modeling epidemics, assessing whether cryptographic implementations complied with federal standards. The consistent lesson from that work is not that the models were weak. It is that the model was never the deliverable. The deliverable was a decision someone could defend afterward, and everything that made the decision defensible, the validation, the documented limits, the human who could explain the reasoning to a skeptic, lived outside the model.
That is what the seventy-year history keeps teaching, winter after winter, renaming after renaming. The technology changes. The discipline required to use it responsibly does not. Govern decisions, because the technology label will not hold still long enough to govern. Fund problems, because the correction always comes for the deployments that were only ever demonstrations. Inventory the values, because the systems now carry them whether inventoried or not. And keep all three professions in the room, because each one can see exactly the failure mode the other two are walking into.
The question this series set out to answer was what artificial intelligence is. The answer turned out to be a mirror: AI is the name we give to the frontier of our own uncertainty, and it retreats exactly as fast as our understanding advances. Fifty years from now, most of what alarms us today will be plumbing, unremarked and load bearing, and the people of that time will be alarmed about something else wearing the same name. They will be convinced, as every generation before them, that this time the machines have finally crossed the line. The most valuable thing the history can give them is the same thing it gives us: the knowledge that the feeling of unprecedentedness is itself the most precedented thing about the whole story.2
Footnotes
-
The thermostat has been making decisions on your behalf, without explanation or appeal, for a century. Nobody has demanded an impact assessment. This is either evidence that we regulate by novelty rather than by consequence, or evidence that the thermostat has excellent lobbyists. ↩
-
Plumbing is meant as a compliment. Civilization is mostly plumbing: infrastructure that works so reliably it has forfeited the right to be interesting. The highest honor a technology can earn is to bore us. AI has been collecting that honor in installments since 1956, one renamed capability at a time. ↩