Ten AI Predictions That Were Confidently Wrong

Confident, dated, wrong predictions are the most reliably entertaining part of AI history and the most frequently fabricated. This list contains only entries whose author, date and substance can be checked, and it says explicitly where the wording is a paraphrase rather than a quotation.

What is and is not on this list

Three rules, applied to every candidate:

• The author and date must be identifiable. No “experts predicted”. Several well-travelled items were dropped at this stage because every source for them cites another secondary source.
• Quoted wording is used only where it is well attested. Where the substance is certain but the exact sentence is not, the entry paraphrases and says so. A page about people getting facts wrong is a bad place to invent a quotation.
• The prediction must have had a deadline that has passed. “Machines will one day think” is not falsifiable on any timetable. Ray Kurzweil’s prediction that a machine would pass a Turing test by 2029 is not on this list because 2029 has not arrived.

Nine that were too optimistic

1. 1955 — the Dartmouth proposal. McCarthy, Minsky, Rochester and Shannon wrote that a significant advance could be made on problems including natural language use, abstraction formation and self-improvement if a carefully selected group worked together for a summer. What happened: natural language use took about sixty years; self-improvement in the sense meant is still open. The full text is discussed on the Dartmouth page.
2. 1954 — machine translation in three to five years. After the Georgetown–IBM demonstration in January 1954, which translated about sixty prepared sentences with a 250-word vocabulary, the project’s leadership indicated that machine translation would be an accomplished fact within three to five years. This is a paraphrase of the claim made in the surrounding press coverage. What happened: the ALPAC report of 1966 concluded that the output was slower and more expensive than human translation, and American funding was cut for about twenty years.
3. 1958 — the perceptron will be conscious of its existence. Reporting a US Navy demonstration in July 1958, the New York Times described the machine as the embryo of a computer that the Navy expected would be able to walk, talk, see, write, reproduce itself and be conscious of its existence. What happened: the machine sorted marked cards. Eleven years later Minsky and Papert’s book set out its representational limits, and the funding went.
4. 1957 — a computer will be world chess champion within ten years. Herbert Simon made four ten-year predictions in this period, published with Allen Newell in 1958: a chess champion, the discovery and proof of an important new mathematical theorem, music accepted by critics as having aesthetic value, and most theories in psychology taking the form of computer programs. What happened: the chess prediction was right in substance and wrong by thirty years — Deep Blue beat Kasparov in 1997.
5. 1965 — machines will do any work a man can do, within twenty years. Herbert Simon, in The Shape of Automation for Men and Management, wrote that machines would be capable, within twenty years, of doing any work a man can do. What happened: 1985 was in the middle of the run-up to the second AI winter. The claim remains open sixty years on.
6. 1967 — the problem will be substantially solved within a generation. Marvin Minsky, in Computation: Finite and Infinite Machines, wrote that within a generation the problem of creating artificial intelligence would be substantially solved. What happened: a generation later, in the early 1990s, the field was at the bottom of its second funding collapse.
7. 1970 — three to eight years to average human general intelligence. Life magazine, in a November 1970 article about Shakey the robot, quoted Minsky as saying that in from three to eight years we would have a machine with the general intelligence of an average human being. What happened: nothing resembling it by 1978, and Minsky subsequently disputed that the article represented his views accurately. The entry is included because the quotation is genuinely in the magazine and is endlessly recirculated; the dispute is included because a page about accuracy should carry it.
8. 2016 — stop training radiologists. Geoffrey Hinton said in a 2016 talk, and repeated in interviews and in a widely read 2017 magazine profile, that we should stop training radiologists now, because it was completely obvious that within five years — or perhaps ten — deep learning would do the job better. What happened: ten years on, radiology residency places have increased rather than fallen and several health systems report radiologist shortages. Image models are genuinely used in radiology workflows, as tools alongside radiologists, which is the outcome the prediction did not describe. Hinton has since revisited the claim publicly.
9. 2016 — cars that need no human driver, by 2021. Several manufacturers announced production vehicles without steering controls for 2021, and Tesla stated in October 2016 that its cars were being built with the hardware needed for full self-driving, with a coast-to-coast autonomous demonstration drive to follow by the end of 2017. What happened: the demonstration drive did not take place, the 2021 targets were missed by every manufacturer that set them, and driverless operation in 2026 exists as a geofenced service in a limited number of cities rather than as a consumer vehicle capability.

A tenth candidate belongs here on substance but is worth stating carefully. After IBM’s Watson won Jeopardy! in February 2011, IBM described medicine, and oncology specifically, as the next application. What is documented rather than merely reported: the MD Anderson Cancer Center project was halted after a University of Texas System audit in 2017 recorded expenditure of roughly $62 million, and IBM sold the Watson Health data assets in 2022. The general claim that Watson would transform oncology is well attested; the exact wording of any single prediction is not, so this entry is a paraphrase.

Two that were too pessimistic

A list of only over-optimism teaches the wrong lesson, which is that scepticism is the safe position. It is not.

1. 1965 — no computer will play good chess. Hubert Dreyfus argued in his RAND memorandum Alchemy and Artificial Intelligence, and in What Computers Can’t Do in 1972, that computers could not achieve competent chess play because the relevant human ability was not rule-following. In 1967 he played the MacHack program at MIT and lost, which the laboratory publicised with some enthusiasm. What happened: the strong version of the prediction failed within two years. Dreyfus’s deeper philosophical argument — that human competence rests on embodied know-how rather than explicit rules — held up considerably better than his chess forecast, and is one of the reasons the symbolic programme stalled.
2. 1950 — 10⁹ bits of storage will be enough. Alan Turing predicted that a computer with about 10⁹ bits of storage — roughly 119 megabytes — would suffice to play the imitation game well enough that an average interrogator would be right no more than 70 per cent of the time after five minutes. What happened: that much memory arrived in the 1980s and was nowhere near sufficient. Turing under-predicted the resource requirement by many orders of magnitude, which is a much rarer error in this field than over-predicting the timeline.

The two failure modes

Nearly every entry above is an instance of one of two errors, and naming them is more useful than the list itself.

There is also a documented meta-result. A 2014 analysis of a large collection of published AI predictions found that forecasts of human-level AI cluster in the range of fifteen to twenty-five years from whenever the forecast is made, largely independently of the decade, and that predictions by AI researchers were not noticeably more accurate than those by others. Fifteen to twenty-five years is far enough away to be unfalsifiable within a career and near enough to be motivating, which is what you would expect a psychologically determined interval to look like rather than a technically determined one.

How to read a prediction being made now

The list above is only useful if it changes how you read the next one. Four questions, each of which would have flagged several entries above:

1. What exactly would count as the prediction coming true? “AI will transform healthcare” cannot be wrong. “A machine will be world chess champion by 1967” can, which is why Simon’s prediction is remembered and a thousand vaguer ones are not. Prefer the falsifiable one even when it is the one that turns out wrong.
2. What are the conditions on the demonstration it is based on? If they are not stated, that is the finding. Ask for the test set, the failure rate, and what happens outside the conditions.
3. Is the remaining work the same kind as the work already done? Going from 90 to 99 per cent is usually a different problem from going from 50 to 90, not more of the same. This is where nearly all the self-driving forecasts went.
4. Who bears the cost of being wrong? A prediction that raises money for the person making it is evidence about their incentives before it is evidence about the world. That is not a reason to dismiss it, and it is a reason to ask what the same person said five years ago and what happened.