History
Anything that starts working stops being called artificial intelligence
Every task the field has solved was reclassified as ordinary engineering shortly afterwards, which leaves the label permanently attached to whatever remains unsolved and makes the record look like a series of failures.
By Manish Trivedi3 min read
A pattern that repeats without exception
Reading text on a page was an artificial intelligence problem, until it worked, at which point it became optical character recognition and moved into scanners. Playing chess at a high level was the canonical test of machine intelligence, until a machine did it, at which point it became search with a good evaluation function.
Understanding speech, routing packets, ranking documents, recommending products, detecting fraud, transcribing dictation and translating text have all made the same journey. Each was a hard research problem, then a solved research problem, then a feature, then infrastructure that nobody describes as intelligent at all.
The pattern is so consistent that it deserves a name, and it has one. A remark widely attributed to one of the field’s founders puts it bluntly: as soon as it works, nobody calls it intelligence any more.
Two readings, and both have something to them
The uncharitable reading is that critics move the goalposts. A capability is nominated as the test, the test is passed, and the response is that the test was never a good one. That is a real rhetorical move and it has been made often enough to be tiresome.
The charitable reading is that each success taught the field something genuine: that the task did not require what everybody assumed it required. Chess was thought to demand judgement of a sort only people possess, and it turned out to yield to a fast search with a good scoring function. That is a discovery about chess.
Both readings can be correct at once, and usually are. The retreat is partly defensive and it is also frequently justified by what the winning method actually turned out to be.
The field is therefore permanently defined by its failures
If everything that works is renamed, the label attaches only to the frontier, meaning the set of problems that have not yielded. A field defined this way can never accumulate visible successes, because success removes an item from the definition.
This produces a strange public record. Technologies descended directly from decades of research are embedded in everything — in phones, in logistics, in medicine, in every search box — while the general impression persists that the field has repeatedly overpromised and underdelivered.
It also distorts the funding cycle. Enthusiasm attaches to whatever is currently unsolved and mysterious, and departs the moment a thing becomes reliable enough to be boring. Reliability is what everybody wanted and it is not what attracts investment.
Researchers who spent careers making something dull and dependable therefore end up with the thanks that infrastructure engineers usually get, which is none at all until it stops working. The attention moves on to the next unsolved thing, and the record of what was achieved moves with it.
The label describes a research programme, not a technology
Once you see the pattern, the term stops looking like the name of a category of systems and starts looking like the name of an activity: attempting the tasks we cannot currently automate. That is a coherent thing to have a word for. It is just not a coherent thing to put on a product.
This explains why the phrase applied to a shipping product almost always resists definition. It might mean a large trained model, a small classifier, a set of hand-written rules, or a feature that somebody thought sounded modern. There is no test that separates them, because the term was never a technical one.
Some researchers prefer to name the method — machine learning, statistical modelling, a specific architecture — precisely to avoid this. The preference is sensible and it has never survived contact with marketing, which reliably prefers the vaguer word.
Whether to abandon the term is genuinely argued
One position holds that the phrase should be retired in technical writing because it means nothing precise and imports expectations from fiction that no system meets. On this view the term causes real harm by making it easy to claim more than was built.
The opposing position is that the phrase names a legitimate long-term aspiration, that it has held a research community together across two collapses in funding, and that abandoning it would fragment the field into method-specific specialisms with no shared conversation.
A modest observation to end on: if the pattern holds, the systems being argued about now will be renamed within a decade, and something else will have inherited the label. That is not a prediction anybody should bet on. It is what has happened every previous time.
Common questions
Is this just critics refusing to be satisfied?
Sometimes, and not always. Each reclassification has usually been accompanied by a genuine finding that the task required less than expected. The move is unfair when it dismisses a real achievement and reasonable when it reports what the achievement actually consisted of.
Does the pattern say anything about current systems?
Only that the label attached to them is unstable, which is worth remembering when reading claims. It says nothing about how capable they are. That has to be established the hard way, by examining what they do under conditions somebody else specified.
What would break the pattern?
A system whose successes kept being described as intelligent after they became reliable and widespread. There is no sign of that yet, and the more likely course is the usual one: today’s striking capability becomes next decade’s unremarkable utility, discussed in the same tone as spellcheck.
Consumer editor, AI Worth Knowing
Manish has written about how it works, in the world, limits & risks for most of the last decade and prefers a plain explanation to a clever one.
