For over a hundred years, clever people have written down how they imagined the future. Today we can go back and read how much of it came true. What can we learn from that for the predictions we make today about tomorrow?
Predictions about the future are not a new genre. What is new is the sheer number of old predictions we can now check - and the discipline that grew out of it: prediction forensics. It looks back systematically at what experts predicted and compares it with what actually happened.
The result is sobering and instructive at the same time. Most concrete predictions were wrong, often in predictable ways. But the patterns in which they were wrong can be described - and from that follows a more modest, more honest way of talking about the future. This essay is an attempt to show these patterns and, at the end, to ask what they mean for the forecasts circulating today about AI, the economy and society.
For the Paris World's Fair of 1900, the chocolate maker Hildebrands produced a series of collectible cards "Germany in 100 Years". They show passenger transport by airship, moving houses, weather machines and moving walkways in the cities. In the same year, the American engineer John Elfreth Watkins published a list of predictions for the year 2000 in the Ladies' Home Journal. Some of them are astonishingly precise: wireless telephony across continents, real-time aerial colour photography sent to newspapers, ready-made pre-cooked meals. Others are wildly off: pneumatic mail-tube networks beneath major cities, the letters C, X and Q vanished from the alphabet, strawberries the size of apples.
The pattern you see here for the first time repeats throughout the whole century: Individual technical points are hit astonishingly often. The society in which they take place is almost never hit. Watkins saw the mobile phone coming. He did not see what it would do to family dinners, elections and the attention spans of children. The images from 1900 show flying machines in the sky, but the people beneath them wear top hats and long skirts - the social world simply stands still.
Watkins' original text appeared in December 1900 under the title "What May Happen in the Next Hundred Years" in the Ladies' Home Journal; the Boston collection archive and the Saturday Evening Post made it accessible again with commentary in 2012. The German Hildebrand cards "Germany in 100 Years" are documented as facsimiles in the holdings of the German Historical Museum in Berlin and in numerous cultural-history volumes about the turn of the century.
For a broader overview: Joseph J. Corn and Brian Horrigan, Yesterday's Tomorrows - Past Visions of the American Future (Smithsonian Institution Press, 1996), collects American visions of the future from 1880 to 1980 and sorts them by the patterns in which they systematically hit or missed.
In the summer of 1956, a handful of mathematicians and computer scientists met at Dartmouth College in New Hampshire for a conference that coined the very term "Artificial Intelligence". Their funding proposal contained the sentence that a carefully selected group of scientists could make "significant advances" on several AI problems in a single summer. Ten years later, Herbert Simon, one of the pioneers, made the famous prediction: "Within twenty years machines will be capable of doing any work a man can do." Marvin Minsky at MIT told Life Magazine in 1970: within three to eight years there would be a machine with the general intelligence of an average human being.
None of these predictions came true. Instead came two AI winters in the seventies and nineties, in which research funding collapsed because the promises had not been kept. Only after 2012, with the rise of deep neural networks, did the discipline deliver results that a broad public noticed. Machine translation, too, was considered solved within five years in 1954 after an initial demo success - in reality it took about sixty.
The Dartmouth funding proposal by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon (1955) is available in full in AI Magazine Vol. 27 No. 4 (2006). Simon's prediction can be found in his book The Shape of Automation for Men and Management (Harper & Row, 1965), Minsky's statement in Life Magazine, 20 November 1970, p. 58 ff.
The history of the 1954 Georgetown-IBM experiment and the ALPAC report of 1966, which discredited machine translation for a generation, are classically reconstructed in John Hutchins, Machine Translation: A Concise History (City University of Hong Kong, 2007). Anyone who wants to understand the AI winters in context should read Nils J. Nilsson, The Quest for Artificial Intelligence (Cambridge University Press, 2010), an honest insider's chronicle.
In 1972 The Limits to Growth appeared, the famous study by the Club of Rome. Using the MIT computer model World3, it simulated various paths for the world of the 21st century. In the standard run, industrial production collapsed between 2020 and 2040, followed by a drop in population. For decades the book was widely regarded by the public as refuted - "the world didn't end, after all". Anyone reading it afresh today discovers: the authors never named 2000 or 2010 as a year of crisis. And studies by the Australian physicist Graham Turner (2008, 2014) as well as a KPMG update by Gaya Herrington (2020) show that actual developments in energy, industrial production and population have come astonishingly close to the model's business-as-usual path.
The example shows two things: First, predictions are often measured against what the public made of them - not against what the authors actually said. Second, models that think in ranges and paths rather than in a single point forecast hold up against reality longer. The serious question is never "was the prediction right?", but "within what range did the prediction lie, and where within it does the actual course run today?"
Original work: Donella H. Meadows, Dennis Meadows, Jørgen Randers, William W. Behrens, The Limits to Growth (Universe Books, 1972). The updates are available as Limits to Growth: The 30-Year Update (2004) and 2052: A Global Forecast for the Next Forty Years (Randers, 2012).
The re-examinations: Graham Turner, "A Comparison of The Limits to Growth with 30 Years of Reality", CSIRO Working Paper, 2008, and his update at the University of Melbourne, 2014. Gaya Herrington, "Update to Limits to Growth - Comparing the World3 Model with Empirical Data", Journal of Industrial Ecology, 2020. Both works independently conclude that the model's standard run matches the real data of recent decades well - which does not mean the predicted collapse must occur, but that the structural logic of the model has not been trivially refuted.
Anyone who systematically works through hundreds of old predictions sees a few patterns that recur again and again. The first has a name: Amara's Law, formulated by the Californian futurist Roy Amara in the seventies. It goes: We overestimate the impact of a technology in the short run and underestimate it in the long run. With the internet this showed itself in 2000 in the dotcom bubble and in 2020 in its actual ubiquity. With AI we are watching it live right now.
The second pattern: Discontinuities are not predicted. The fall of the Berlin Wall in 1989, the financial crisis of 2008, the pandemic of 2020, the sudden visibility of ChatGPT at the end of 2022 - none of these events appeared in the serious forecasts of the respective prior years. What can be predicted are trends along quantities that move slowly and continuously (demographics, infrastructure, facilities under construction). What cannot be predicted are the breaks that make up what is actually interesting.
The third pattern: Second-order effects are overlooked. The car was predicted - suburbanisation, commuter culture and global warming were not. Social media was anticipated in a broader sense - that it would upend elections and teenagers was not. Whoever sees a technology rarely sees how the world reorders itself around it. And that reordering is usually the real event.
Roy Amara was president of the Institute for the Future in Palo Alto; the law named after him is not a formal theory but a rule of thumb that recurs in countless technology trajectories. A broader treatment with a similar tenor is Carlota Perez, Technological Revolutions and Financial Capital (Edward Elgar, 2002), which shows how new base technologies (railways, electricity, computers) regularly run through a pattern of overestimation, crash and slow deep penetration.
On the question of why discontinuities are so hard to predict, Nassim Nicholas Taleb, The Black Swan (Random House, 2007), is the most widely read work - with a sharp analysis of why precisely the most consequential events are the ones nobody saw coming. On the second-order effects of technical innovations, Edward Tenner, Why Things Bite Back (Knopf, 1996), remains a good, historically rich introduction.
From the mid-eighties on, the American psychologist Philip Tetlock ran a systematic study for over twenty years: 284 experts from politics, economics and intelligence services made nearly 28,000 concrete forecasts in total, which were later checked against the actual course of events. His sober result: The average expert forecast was barely better than random guessing. Specialists in their own field were often worse than generalist observers, because they believed too strongly in their one preferred theory.
In a follow-up study (Good Judgment Project, 2011-2015, funded by the American intelligence agency IARPA) Tetlock identified a small group of people who forecast markedly better - the superforecasters. What distinguished them was not special knowledge but a way of working: holding many models in mind at once, expressing forecasts in probabilities rather than certainties, updating regularly in light of new information, documenting and analysing their own mistakes. It is an attitude, not a talent.
The main study is published as Philip E. Tetlock, Expert Political Judgment: How Good Is It? How Can We Know? (Princeton University Press, 2005). The accessible account with the results of the Good Judgment Project is Philip E. Tetlock & Dan Gardner, Superforecasting: The Art and Science of Prediction (Crown, 2015).
A more popular, more anecdotal version of the same message is Dan Gardner, Future Babble: Why Expert Predictions Are Next to Worthless (Dutton, 2010). The methodological foundation - why it makes sense at all to cast forecasts in probabilities rather than yes/no statements - is classically worked out in Daniel Kahneman, Thinking, Fast and Slow (Farrar, Straus & Giroux, 2011), especially in the chapters on overconfidence and calibrated judgment.
When you sort a hundred years of documented predictions, a few sentences crystallise out that almost always hold. First: point forecasts with a year attached are almost worthless. "In 2030 X will happen" is not a forecast, it is a headline. Second: scenarios that make several paths explicit do better - not because one of them is right, but because they honestly portray the range of the possible. Third: short horizons (two to five years) are usable, longer ones drift off quickly. Fourth: the things that end up having the biggest effect were on no top-10 list of their time.
For the current debate about AI this means: distrust any prediction that extrapolates linearly. Whoever says in 2026 that AI will automate everything in five years sounds like Marvin Minsky in 1970. Whoever says AI will in the end change nothing sounds like the sceptics who, after the dotcom crash, thought the internet was overrated. In the light of historical predictions, both voices are unlikely. What is likely is a third possibility: less in the short run than the headlines promise; more in the long run than the sceptics believe, and in areas no one thinks of today.
Practically speaking, the only solid answer to an uncertain future is not to commit to a single narrative. Whoever builds software, structures and contracts today ideally builds them so that they work under several of the plausible scenarios - and not just under a fair-weather path. That is not defeatism. It is the only form of realism that a hundred years of documented mispredictions truly permit.
The most honest answer to the question "Can we predict the future?" is: a little, within narrow limits, and only if we are willing to think in ranges and probabilities rather than in headlines.
What a hundred years of prediction forensics really show is not that predictions are pointless - but that the form of the prediction determines whether it is useful. Whoever draws scenarios instead of issuing point forecasts; whoever speaks in probabilities instead of certainties; whoever updates regularly instead of clinging to old predictions - that person does better on average. Whoever does the opposite is writing tomorrow's Watkins list today.
At wendwerk we think along these lines. We do not promise software built for the future - there is no one future. We build systems that still hold up under several plausible futures. That is unspectacular. But it is the only thing the history of predictions truly justifies.
You can find the six concrete future scenarios seriously discussed today under What will the world look like in 2035?. The minds shaping the global debate today are under Where we are heading. How we picture the future of small and medium-sized businesses concretely is under What does the future of companies look like?.
Curated by Johannes Hohls for wendwerk.