The Future Is Late Again: AI and the Business of Being Certain

I wanted a list of all the wonderful things artificial intelligence was supposed to have done by now, and all the terrible things it was supposed to have done to us. Something to keep beside the next announcement. The computer will do everything. Nobody will have a job. Everybody will be rich. Everybody will be dead. Depending on which gentleman has the microphone, sometimes we manage several of these before lunch.

There is a considerable history here. There is also a considerable temptation to make the history say whatever would be most satisfying today, and I know that temptation, I would quite like to assemble the distinguished gentlemen in chronological order and make them stand beside their expired predictions. Unfortunately, even this small pleasure requires checking the fucking dates.

Some forecasts failed. Some arrived late. Some were promises about products that were already being sold, except the products did not do what the advertisements said. And some warnings have not been disproved at all. They describe something that might happen under conditions that have not yet occurred. Putting those in a collection of failed prophecies would make the collection more impressive and me less honest.

I would prefer to retain at least one of those disadvantages.

In November 1957, Herbert Simon and Allen Newell predicted that a computer would become world chess champion within ten years. The paper appeared in 1958, which is one reason the date wanders about in retellings. Their clock started with the 1957 talk. By 1967, the reigning human champion had not been relieved of his responsibilities by an electronic brain.

Deep Blue defeated Garry Kasparov in a match in 1997. That matters. The achievement was real, and the forecast was late by thirty years. I don’t see why acknowledging one should require swallowing the other. If I promise to return your money next Thursday and do it thirty years later, you may appreciate receiving it without retrospectively praising my punctuality.

Simon went considerably further in 1960. Machines, he expected, would be technically capable of doing any work a human could do within twenty years. That gives us 1980. The quotation is often dated 1965, when it was republished, handing the machines another five years without their even asking.

He did not predict that every human would be unemployed. He specifically distinguished technical capability from economic advantage and expected people to retain advantages in certain kinds of work. It is an interesting experience to find a sixty-six-year-old discussion making a distinction that still gets trampled every time someone turns a chart of automatable tasks into a funeral announcement for the entire labor force.

But the capability forecast failed. In 1980, machines could not do every kind of work a human could do. Saying so does not require declaring Simon an idiot. He helped build the subject. That makes the error worth understanding. His authority did not make the extrapolation come true.

I. J. Good gave the future an even more vigorous shove. In a 1962 essay, he imagined machines reaching extraordinary intelligence, learning their own construction, and designing better successors. Science would accelerate, disease would be eliminated, efficient robots would create unemployment, and the resulting problems might require the machines’ advice. His guess for when the scenario would arrive was 1978. He also explicitly said his expectations might be wrong.

They were, on that timetable.

In his 1965 paper, Good judged it more probable than not that an ultraintelligent machine would be built before the twentieth century ended, leading to an intelligence explosion. That event did not happen before 2000. A probabilistic forecast can land on the losing side without proving that the forecaster’s entire theory is impossible. It still belongs in the record. The date does not disappear because the idea remains interesting.

Minsky expected the problems of creating artificial intelligence to be substantially solved within a generation, writing in 1967. Hans Moravec’s 1977 calculations put human equivalence within reach in ten years as computing became cheaper. Later, Moravec anticipated multipurpose universal robots by 2010. Not merely a machine repeating one factory operation, but something broadly programmable for ordinary chores.

We have had decades of useful machines since then. We have also had decades of discovering how much was being smuggled into words like intelligence, ordinary, and chore.

I find the domestic robot particularly instructive. A demonstration can happen on a cleared table under favorable lighting. A home contains the object you forgot you left on the floor, the wet surface, the awkward handle, the thing that looks disposable and isn’t. Ordinary life has an appalling number of exceptions. We barely notice how much competence we expend getting through it, then assume a machine that can perform one impressive intellectual task must be almost ready to handle the rest.

Ray Kurzweil supplied a useful service in 2010: he published an assessment of his own earlier predictions. I disagree with plenty of his scoring, but it is better to have a scorecard to argue with than another confident interview conducted as though the last one never happened.

His 1999 book anticipated that most text would be created by speaking by 2009. Speech recognition existed. People used it. Most text was not being created that way, and he marked that prediction wrong. He also marked wrong his expectation of intelligent highways on which a car could take over long-distance driving while its occupant relaxed. His assessment of synthetic speech was qualified: extended passages still sounded artificial.

The distinction is almost painfully mundane. A capability can exist without becoming the dominant way people do something. A demonstration of voice recognition does not tell us how often people will want to dictate a private message in a crowded room. Engineering is difficult enough before human habits arrive and refuse to cooperate.

Then came the cars.

Nissan announced in 2013 that it would have multiple commercially viable autonomous-drive vehicles ready by 2020. Volvo’s 2015 announcement aimed for 100 self-driving cars in customers’ hands around Gothenburg by 2017. Ford announced in 2016 that it would mass-produce commercial autonomous vehicles by 2021, without steering wheels or pedals. BMW, Intel, and Mobileye announced a partnership aimed at highly and fully automated driving in series production by 2021.

The deadlines passed. Driver assistance improved, but assistance was not what those announcements had invited people to imagine. BMW’s production level 3 system arrived in 2024. Ford’s Argo AI investment ended with a write-down and a wind-down. Nissan was still announcing a route toward commercial autonomous mobility services for fiscal 2027 in its 2024 plans.

Lyft cofounder John Zimmer predicted in 2016 that autonomous vehicles would provide most Lyft rides within five years. He also expected private car ownership to all but end in major American cities by 2025. These are useful predictions to keep intact because they say something definite. A few autonomous rides do not constitute most rides. A person selling a car does not constitute the disappearance of private ownership across major cities.

Musk said at TED in 2017 that a Tesla should complete a fully autonomous Los Angeles to New York trip before that year ended. In 2019, he said there would be more than a million Tesla robotaxis on the road the following year. General Motors, meanwhile, was announcing commercialization of its Cruise autonomous vehicle for 2019.

None of those deadlines was met as promised. GM eventually stopped funding Cruise’s robotaxi development. That does not make autonomous driving imaginary. It makes the distinction between a working system somewhere and a promised system at a promised scale rather important, especially when somebody has already collected money on the strength of the promise.

And no, every delayed software release does not deserve to be paraded beside the great failed forecasts of scientific history. Apple’s personalized Siri belongs here as a smaller, familiar example: the 2024 presentation promised improvements over the following year; the new Siri AI began rolling out in English beta in September 2026, with some actions still described as coming soon. Late delivery. Not permanent impossibility. Not every advertised capability completely available to everyone.

Even irritation ought to be able to count.

There is another part of this history that smells worse because the future was barely involved. The claims concerned products being sold already.

The SEC found that Delphia, which advertised AI and machine learning incorporating customer data into investment decisions, did not have the capabilities it claimed. It also found misleading AI claims at Global Predictions. These were settled enforcement actions. They do not establish that machine learning cannot be useful in finance. They establish that putting AI in the sales pitch does not establish that the advertised AI is there.

Presto’s drive-through ordering claims were more bodily. The company said its proprietary system eliminated the need for human order taking. According to the SEC’s settlement findings, people abroad processed the vast majority of orders through those units. A reported rate of orders completed without intervention referred to restaurant staff, leaving substantial offshore human involvement outside the reassuring number.

The human had not become unnecessary. The human had become less visible in the description.

That bothers me more than another delayed robot. Somewhere in the account of labor vanishing, there is labor, actual people doing the work while the achievement is attributed to the apparatus in front of them. It is a fairly old indignity with a fashionable interface.

DoNotPay’s claims that its service could substitute for human legal expertise brought FTC action and restrictions on unsupported advertising. Evolv’s weapons-screening claims brought allegations that the system missed weapons while flagging harmless objects. Neither case needs embellishment. An unproved professional-replacement claim and an imperfect safety system are serious enough without turning them into proof that every application of AI is a fraud.

But where are the failed predictions of doom? There are fewer clean examples than the symmetry of my original request suggested.

Moravec offered one in a 1995 interview. He put European unemployment at roughly 15 percent and said it would essentially never reverse as society moved toward people being subsidized by productive machines. The irreversibility claim did not hold up. Eurostat puts EU unemployment at 6 percent in 2025. His loose description of Europe and today’s EU measure are not an identical statistical series, and I am not going to pretend they are. The broader prediction of an irreversible unemployment condition still fares badly against the subsequent record.

Other supposed examples begin falling apart when I inspect them. The famous estimate that 47 percent of American employment was susceptible to computerization measured exposure under a model; it was not a dated promise that 47 percent of Americans would lose their jobs. A warning that advanced AI could cause human extinction does not become false because a chatbot has been released and I am still here complaining about it.

I would enjoy a stronger argument. This is the one the evidence permits.

Vernor Vinge is another case where the quotation needs its surroundings. His 1993 paper’s opening anticipated superhuman intelligence within thirty years, followed by the end of the human era. But the body also said he would be surprised if the event arrived before 2005 or after 2030. That wider window has not closed as I write in October 2026. The language describes a posthuman transition, not simply a prediction that everybody would be murdered by a computer in 2023. Calling the whole thing conclusively falsified would be convenient. It would also involve quietly removing the sentence that inconveniences me.

There are genuine risks here. Fraud does not need general intelligence. Neither does a bad decision made at scale, a vulnerable system granted access to records, or an employer using automation to squeeze people who were already being squeezed. A grand catastrophe remaining uncertain does not make these smaller injuries imaginary, and the fact that earlier deadlines failed does not grant the next system immunity from doing damage.

I use these tools. I have reasons to want them to improve. That is part of why the certainty irritates me. I can get something useful from a system and still want to know what failed, where it needs supervision, and whether the impressive result survives an ordinary untidy task. I don’t need to first announce my allegiance to either the glorious future or the approaching extinction.

I keep coming back to the dates because a date gives the claim somewhere to stand where we can see it. What can the machine do? Under which conditions? How much of the work is being done by people outside the frame? What would count as failure? After the deadline passes, I would like to hear the answer from the same mouth that supplied the confidence.

Instead, another date arrives.

I am getting older in the interval. The future remains remarkably fresh-faced.

PS: Sources, dates, and qualifications

Checked against the sources below on October 23, 2026. This is a selected history, not an exhaustive census. A missed deadline does not establish permanent impossibility; an unrealized probabilistic forecast is not a logically disproved probability. Enforcement findings and allegations are identified separately from forecasts. References are collected here to leave the essay uninterrupted.

1. Simon and Newell, 1957/1958: Heuristic Problem Solving: The Next Advance in Operations Research, the original November 14, 1957 address published in 1958. The ten-year chess forecast appears on printed page 7. IBM Research’s account of Deep Blue records the 1997 match victory.
2. Simon’s technical-capability forecast: The New Science of Management Decision, original 1960 book reproduced online, page 38. The twenty-year forecast therefore points to 1980. It explicitly distinguishes technological capability from economic comparative advantage. Quote Investigator’s bibliographical check explains the common misdating to 1965.
3. Good’s 1978 guess: The Social Implications of Artificial Intelligence, 1962, printed pages 192–195. The scenario and its explicit uncertainty belong together.
4. Good’s twentieth-century forecast: Speculations Concerning the First Ultraintelligent Machine, published in 1965; the conclusion on printed page 78 assigns more-than-even odds to the event before the century ends. The filename of this archived copy uses 1966.
5. Minsky: Computation: Finite and Infinite Machines, 1967, original book reproduction, page 2. His generation-length expectation is separate from the disputed three-to-eight-year quotation attributed to him in a 1970 magazine article; that disputed quotation is not used here.
6. Moravec’s optimistic timelines: Intelligent Machines, September 3, 1977 and the preface to his later robotics book. Computing-capacity extrapolation alone did not deliver human-equivalent intelligence or broadly competent universal robots on the anticipated schedule.
7. Kurzweil’s own audit: How My Predictions Are Faring, October 2010. See printed page 38 for speech-created text, page 41 for synthetic speech, and page 82 for intelligent highways. These were expectations for 2009 from The Age of Spiritual Machines.
8. Nissan: Original August 27, 2013 announcement and February 2024 commercialization plan targeting fiscal 2027. Driver assistance should not be substituted for the earlier autonomous-drive promise.
9. Volvo: February 19, 2015 Drive Me announcement and December 2017 update. The latter described the first participating families and a progression toward more advanced autonomy.
10. Ford: Mark Fields’s August 16, 2016 company announcement. Ford’s third-quarter 2022 SEC filing, note 10, records the additional time and capital needed for level 4 commercialization and the Argo wind-down.
11. BMW, Intel, and Mobileye: July 1, 2016 joint announcement. BMW’s current account dates production level 3 to spring 2024. Level 3 conditional automation is not unrestricted driverless operation.
12. Lyft: John Zimmer’s September 18, 2016 essay. Its separately numbered predictions concern majority-autonomous Lyft rides within five years and private ownership almost ending in major US cities by 2025.
13. Tesla: TED’s original 2017 event account records the cross-country demonstration deadline. Tesla’s April 22, 2019 Autonomy Day recording contains the million-robotaxi forecast; an archived transcript makes the wording searchable.
14. General Motors: June 2018 company announcement specifies Cruise commercialization in 2019. GM’s 2024 annual filing records the subsequent decision to stop funding robotaxi development.
15. Apple’s delayed Siri: WWDC 2024 keynote and September 14, 2026 release announcement. The latter describes an English beta rollout and further capabilities still to come, not complete worldwide delivery of every advertised function.
16. Delphia and Global Predictions: SEC announcement and settled findings, March 18, 2024. Both firms settled without admitting or denying the findings. These concern false or misleading representations about existing products, not forecasts of future technical limits.
17. Presto: SEC settlement order, January 14, 2025, especially its summary of proprietary-system claims and offshore human involvement. Presto settled without admitting or denying the findings.
18. DoNotPay: FTC case record and final-order timeline. The action addressed unsupported claims of substituting for professional legal expertise; it does not establish that every legal use of AI is ineffective.
19. Evolv: FTC complaint and proposed-settlement announcement, November 2024. The failures discussed here are identified as allegations in that enforcement case, not a trial verdict.
20. Moravec’s unemployment claim: Original 1995 interview and Eurostat’s 2025 EU unemployment figures. His rough Europe-wide estimate and the present EU series should not be treated as a precisely matched before-and-after comparison.
21. Automation exposure: Frey and Osborne, The Future of Employment, original 2013 working paper. Susceptibility to computerization is not an observed layoff count or a guarantee of job elimination.
22. Vinge’s actual window: The Coming Technological Singularity, original 1993 NASA conference paper. Read the thirty-year opening alongside the 2005–2030 window on the following page. The wider window remains open at this essay’s publication date.

The thumbnail is an AI-generated editorial illustration, not a historical photograph or a technical diagram.