Lessons from History

Seventy years of AI have produced booms, busts, failed forecasts and a handful of hard-won rules of thumb. Here are the most useful ones, what they really say, and how to use them when you hear the next big claim.

In 1965 Herbert Simon, one of the founders of AI and later a Nobel laureate, repeated in print his forecast that “machines will be capable, within twenty years, of doing any work that a man can do.” Twenty years later the field was entering its second winter. In 2016 Geoffrey Hinton, a pioneer of deep learning and another future Nobel laureate, suggested that deep learning would outperform radiologists within about five years and that people should stop training new ones; a decade on, radiologists are still in demand. Yet the same decade saw machines master Go and hold fluent conversations long before most experts expected. The smartest people in the field have been wrong about AI in both directions. What can the rest of us learn from that?

Booms, busts and the hype cycle

AI’s history has a rhythm. A breakthrough arrives; enthusiasm and funding pour in; promises outrun results; money is withdrawn; work continues quietly; and eventually another breakthrough arrives. The two great slumps are known as AI winters, roughly 1974–1980 and 1987–1993.

The most famous picture of this rhythm is the hype cycle, introduced by the analyst Jackie Fenn at the consultancy Gartner in 1995. Step round it below.

HYPEOne turn

This guide’s framing Treat the hype cycle as a useful cartoon, not a law of nature. It is a consultancy’s model, and researchers who have checked it against Gartner’s own data, such as Ozgur Dedehayir and Martin Steinert in 2016, found inconsistencies: many technologies never follow the full curve, and some simply vanish. Its value is as a reminder that excitement and usefulness move on different timetables.

A sharper version of the same idea is Amara’s law, attributed to the American futurist Roy Amara of the Institute for the Future (the date is uncertain): “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.” The internet in 2000 fits it well. So, arguably, does AI in the 1960s, whose dreams of machine translation and speech recognition came true only decades later.

What caused the winters

Winters are not random. Each followed a specific mismatch between promise and delivery.

Overpromising: confident forecasts in the 1950s and 1960s set expectations no one could meet. The combinatorial explosion: methods that worked on toy problems collapsed when the number of possibilities grew, a point at the heart of Sir James Lighthill’s 1973 report to the UK Science Research Council. Failed flagships: the 1966 ALPAC report ended most US funding for machine translation, and Minsky and Papert’s Perceptrons (1969) cooled interest in neural networks. Politics: in the US the Mansfield Amendment (1969) pushed military research funding towards projects with direct military use. Hardware: computers simply lacked the memory and speed that later methods needed. The First AI Winter →

Consensus In both cases the underlying ideas did not die. Neural networks, machine translation and knowledge representation all returned, usually once data and computing power caught up.

Five rules of thumb

Over the decades, observers have named a few recurring patterns. Tap each tile.

Notice how these interact. The AI effect makes real progress feel smaller than it is. Moravec’s paradox explains why an AI that writes an essay still cannot reliably fold your laundry. The ELIZA effect makes systems seem more capable than they are. The Bitter Lesson warns researchers against betting on cleverness over scale, though critics note it is a reading of history, not a guarantee about the future.

Why predictions fail in both directions

The record of AI forecasting is humbling. In 1958 Simon and Newell predicted that within ten years a computer would be world chess champion; it took until 1997. Yet many experts thought a computer would not beat a top Go professional until the mid-2020s, and AlphaGo did it in 2016. Few researchers in 2015 expected a system that could write fluent essays on any subject within a decade. In a 2023 survey of 2,778 AI researchers by Katja Grace and colleagues, the median forecast for when machines could outperform humans at every task moved 13 years earlier than in the same survey a year before.

Why so unreliable? A few reasons recur:

Is this time different?

Debated Every boom has believed it was the real one. Is today’s generative-AI boom another peak before a crash, or something more durable? Choose a position to hear its case and its weak point.

The question

Is this time different?

Choose one to hear its case.

Other views
  • Different in kind: some researchers argue that general-purpose systems improving across many tasks at once break the old pattern of narrow successes, so historical analogies mislead.
  • Different in risk: others, including Hinton and Bengio, say the key difference is not economic but that capable systems may become hard to control, so the cost of being wrong is higher.
  • Not different at all: sceptics such as Gary Marcus see the familiar overreach, and predict that reliability limits will bring expectations down to earth.

Speculative No one knows which view will prove right. History cannot answer the question for us; it can only make us suspicious of anyone who claims to be certain As of 2026.

Quick check

Moravec’s paradox says that…

Abstract reasoning turned out easier to mechanise than perception and movement.

What is the best description of Gartner’s hype cycle?

Jackie Fenn introduced it at Gartner in 1995. Many technologies do not follow the full curve.

The “AI effect” describes…

That last option is the ELIZA effect. The AI effect keeps moving the goalposts.

// Pause and reflect

  1. Think of an AI claim you have read this month. Which of the five rules of thumb helps you judge it?
  2. Which do you think is more costly for society: overestimating AI in the short run or underestimating it in the long run?
  3. If a third AI winter came, what do you think would cause it? If it never comes, why not?

Your notes are saved privately in this browser.

What happened next

That is the end of the story so far. To see the whole sweep again, revisit The Complete AI Timeline or AI History at a Glance. To understand how today’s systems actually work, the Novice track starts with What Is AI, Really?

Key terms on this page

Sources & further reading