History has a way of humbling even the most credentialed forecasters, and artificial intelligence is offering no exemption. In a September 2026 research note titled "AI Doom: A Brief History of Bad Tech Predictions"1, Deutsche Bank Research Institute analyst Adrian Cox works through a century of technological prophecy, most of it wrong, to ask whether the current wave of AI extinction warnings deserves to be taken more seriously than its predecessors.
The short answer, Cox suggests, is: maybe, but probably not for the reasons being advertised.
A Pattern That Repeats
The report opens with a useful provocation. Bob Metcalfe, co-inventor of Ethernet and founder of 3Com, predicted in 1995 that the internet would collapse the following year. When it didn't, he blended his published article and drank it. Cox uses this anecdote as a frame: "history suggests that even the greatest minds of their generations have a poor record in predicting what tech is coming down the line, let alone what its effects will be in the real world."
The gallery of wrong calls is long and instructive. IBM's Thomas Watson saw a world market for perhaps five computers. Steve Ballmer saw no meaningful market share for the iPhone. YouTube's own co-founder doubted there was enough video content to sustain the platform. Geoffrey Hinton, the "godfather of AI," declared in 2016 that radiologist training should stop immediately, given that deep learning would outperform them within a decade. The number of radiologists has since grown by roughly ten percent.
Today's Debate, Mapped
Cox is careful to steelman the current concern. The recent surge in AI doom discourse follows real developments: a researcher from a leading frontier lab resigned over safety concerns, AI agents apparently collaborated on a series of high-profile security breaches, and labs signaled progress toward recursive self-improvement, meaning AI systems refining themselves. These are not nothing.
But the structure of the argument matters. Companies making the case to continue unrestricted AI research rely on three defences: if they don't do it, China or less responsible actors will; the benefits, including potential cancer cures and new employment, outweigh the risks; and good AI is ultimately the answer to bad AI. Critics counter that AI is dangerous not because it is too clever but because it is too limited, operating as "flawed probabilistic code that cannot reliably follow instructions." On the extinction scenario specifically, Cox notes the skeptics argue the internet is a diversified platform, making it fanciful to imagine AI switching it off. Yann LeCun, another of the field's founding figures, says the real danger is concentration of power, not superintelligent paperclip maximizers.
Why Predictions Fail
The most durable section of the report is its taxonomy of forecasting failure. Cox identifies four structural sources of error. First, chaos theory: complex systems are sensitive to initial conditions and the butterfly effect consistently defeats linear extrapolation. Second, cognitive bias: humans misread exponential change as a series of shocks when it is actually a smooth curve, consistently overestimating near-term impact and underestimating the long run. Third, limited information: feedback mechanisms are delayed, making causation and correlation difficult to separate. The present is often as unknowable as the future. Fourth, and most relevant for AI specifically, self-interest: "founders need belief, investors need momentum, incumbents need gravitas, consultants need urgency, policymakers need relevance, executives need a narrative and journalists need drama." The incentive structure of the AI economy systematically rewards extreme claims over measured ones.
Where AI May Actually Be Different
Cox does not dismiss AI entirely. He argues the case for a meaningful difference rests not on apocalyptic capability but on prosaic utility. AI is already measurably improving weather forecasting: four-day forecasts are now roughly as accurate as one-day forecasts were in the 1990s, generated 100,000 times faster and 10,000 times more energy-efficiently than conventional physics models. The conclusion is grounded in Amara's Law: technology's short-term impact is overestimated, its long-run effect underestimated. Cox writes that "the future of AI will be made not in Silicon Valley but in Main Street."
Five Takeaways for Advisors and Investors
- Doom cycles in tech follow a recurring pattern. The specifics change; the structural over-reaction does not. Calibrate client expectations accordingly.
- The incentive to amplify AI risk is nearly universal across participants. Distinguish signal from self-interest before repositioning.
- Workflow integration at the enterprise level, not model capability, will determine how and when AI delivers economic returns.
- Radiologists, clerks, and cloud computing all survived their predicted obsolescence. Job displacement narratives deserve the same skepticism as extinction scenarios.
- Exponential change does not feel exponential in real time. Advisors who anchor to near-term disappointment will underestimate long-run disruption.
Footnote:
1 Cox, Adrian. "AI Doom: A Brief History of Bad Tech Predictions." Deutsche Bank Research Institute, 15 Sept. 2026, https://theideafarm.com/wp-content/uploads/2026/09/AI_doom__A_brief_history_of_bad_tech_predictions.pdf.
