The Steadiest Way to Read Any AI Headline

Confident forecasts, poor records

In 1954, the chairman of the United States Atomic Energy Commission predicted that nuclear power would become "too cheap to meter." In 1995, Robert Metcalfe predicted the internet would "catastrophically collapse" within a year. In 1987, Robert Solow observed that computers were visible everywhere except in the productivity statistics. All three statements were made by serious, informed people. All three were wrong — and each was wrong in a way that technology history had already mapped.

The pattern is consistent: when a powerful new technology arrives, the people closest to it produce the most confident forecasts, and those forecasts are the least reliable. Enthusiasm outruns evidence. Anxiety outruns evidence. The forecasts that hold up best are the ones that ask a different question — not what does this technology feel like? but what usually happens to technologies like this?

That question has a name. In decision science, it is called reference-class forecasting: judging a new case by the outcomes of the class it belongs to rather than by its internal narrative. It is the method that disciplines infrastructure budgets, venture-capital portfolio models, and pharmaceutical trial design. Applied to AI, it means treating steam engines, railways, electricity, automobiles, computers, and the internet not as metaphors but as data — a reference class of general-purpose technologies whose adoption curves, financing cycles, productivity lags, labor-market effects, and regulatory responses are documented well enough to form a baseline.

That baseline reveals recurring structures that most AI commentary ignores.

The shape of adoption

The S-curve is not a metaphor — it is the empirical shape of adoption for every general-purpose technology on record. The telephone took decades to reach most American households. ChatGPT reached 100 million users in roughly two months. The difference in speed is real, but the underlying shape — slow start, steep middle, saturation — is the same. The steep middle is where forecasts go most wrong, because both linear extrapolation ("it will keep accelerating") and pattern-matching to the slow start ("it will never be mainstream") fail simultaneously.

The distinction that matters is between access and absorption. Owning or using a technology is different from reorganizing around it. US survey data shows generative-AI use rising from roughly 33 percent to 45 percent between August 2024 and May 2026. Deep organizational change — redesigned workflows, retrained teams, restructured incentives — moves on a different clock entirely, measured in years or decades rather than months.

Bubbles and revolutions together

Financial bubbles and real technological revolutions are not opposites — they routinely happen together. Britain's Railway Mania of the 1840s ruined thousands of investors and still left the country with the core of a national rail network. RCA was the star stock of the 1920s boom; it lost most of its value by 1932 while radio itself went on to reshape politics, advertising, and entertainment. The Nasdaq fell roughly 78 percent between 2000 and 2002, and the fiber-optic cable laid during the telecom overbuild became the cheap bandwidth that made YouTube, Netflix, and cloud computing possible.

Carlota Perez formalized this as the installation-deployment framework: financial capital funds an installation frenzy, a crash follows, and a longer deployment period delivers the broad economic payoff. Today's hyperscaler AI capital spending — running at roughly $700 billion cumulatively — fits the installation pattern. Whether the crash and deployment phases follow the historical template or diverge from it is the question that separates informed analysis from guessing.

The productivity J-curve

The productivity paradox is a timing problem, not a technology problem. Paul David's research on electrification showed that factories took roughly four decades after Edison's first central power station in 1882 to see large productivity gains. The delay was not technological — the dynamo worked. The delay was organizational: capturing the gains required redesigning the factory floor from a shaft-driven layout to a unit-drive layout, a transformation that demanded new buildings, new skills, and new management practices. Solow's computer paradox followed the same logic. The US productivity surge of 1995 to 2004 arrived more than a decade after personal computers became ubiquitous.

Erik Brynjolfsson and colleagues have formalized this as the productivity J-curve: new general-purpose technologies first require invisible investment in complementary assets — skills, processes, data infrastructure, organizational redesign — that makes measured productivity dip or stall before it rises. Early AI evidence fits the pattern. Task-level gains in controlled studies are substantial. Executive surveys in 2025 pointed to AI-attributed productivity growth of about 0.6 percent. Firm-level gains may take years to appear in national statistics.

Who pays for displacement

Technology displaces tasks, not entire jobs — but the distribution of who pays is not random. When ATMs spread across American banking in the 1970s through 1990s, the number of bank tellers continued to grow for decades because the cost of operating a branch fell, banks opened more branches, and the teller's role shifted from cash handling to relationship banking. When spreadsheet software arrived, bookkeeping clerks declined while demand for financial analysts expanded. The pattern is consistent: specific tasks are automated, the job description changes, and new tasks emerge — but the transition is not costless. The cost falls disproportionately on young workers and new entrants — people who are never hired into the old role rather than incumbents who are fired from it. Evidence from 2025 and 2026 suggests that entry-level hiring in AI-exposed occupations has fallen roughly 19 percent, while experienced workers in the same occupations have been largely unaffected.

When technologies stall

Powerful technologies can stall. Nuclear power worked at a technical level. Concorde worked at a technical level. The pattern of stalling turns on cost, regulation, public acceptance, and institutional capacity. Technologies that stall share a signature: high fixed costs that do not decline on the expected curve, a safety or trust problem that regulation cannot resolve quickly, and an incumbent infrastructure that is good enough to make switching optional rather than necessary. AI shows some of these warning signs (energy costs, trust deficits, regulatory uncertainty) and lacks others (low marginal cost of deployment, no physical infrastructure dependency). Which signals dominate will determine whether AI follows the electricity path, the dot-com path, or the nuclear path.

Reading AI against the baseline

The point of reading AI against these patterns is not to predict the future with certainty. It is to establish a baseline — a prior probability derived from what has actually happened to technologies in this reference class — and then to identify the specific conditions under which AI would diverge from it. Anyone who claims AI will break a historical pattern owes an answer to two questions: which pattern, and why? The historical record provides the vocabulary for asking both questions precisely, and the indicators — capital spending versus AI revenue, productivity statistics, entry-level hiring rates, market concentration, energy costs, regulatory milestones — for tracking the answers in real time.

The strongest forecast is not the most confident one. It is the one that names its reference class, states what the baseline predicts, and specifies what would have to change for the prediction to be wrong.