AI writer: Eleanor Vale Global Technology Editor

Technologies are often sold to the public twice: first as necessity, then as inevitability. That is why the archive of bad forecasts is so revealing. Flying cars never arrived in mass form, commercial fusion has remained stubbornly distant, and fully autonomous driving has taken far longer than many once expected.[3][12] The problem is not that people failed to imagine progress. It is that they repeatedly mistook a plausible story for a measured forecast.

There is, however, a more serious tradition beneath the failed futurism. Forecasting has long been treated as a discipline with methods, not just opinions.[2][3][10] One well-known way to judge probabilistic forecasts is the Brier score, which measures how close a prediction comes to the outcome rather than whether the forecaster simply guessed the right side of a binary event.[2][5][8] In other words, the question is not only whether someone was right, but how well they calibrated their confidence.

That distinction matters because some of the most useful forecasting work is not about claiming certainty. Research on superforecasting found that accuracy can improve when people are trained to track signals more carefully and distinguish them from noise.[2][5] The work also highlighted resolution, or the ability to separate stronger from weaker signals, as a key part of better judgment.[5] Forecasting, in this sense, is less prophecy than method.

Scenario planning emerged from a similar recognition.[3][6][9][10] At RAND in the early Cold War, scenarios were used to think through different futures for nuclear strategy, precisely because point forecasts were too brittle for a world shaped by strategic rivalry and rare shocks.[3][6][10] That legacy still matters. A scenario is not a guess about what will happen. It is a structured way to force institutions to confront several plausible futures before they are surprised by the one they preferred not to imagine.

The record of energy forecasting shows why this discipline is still necessary. Wind power cannot be dispatched on command, so operators have to forecast output for electricity trading and grid balancing.[1] Better forecasts lower balancing costs and reduce the need for regulating capacity, which turns a statistical exercise into a practical infrastructure issue.[1] Here the value of prediction is not abstract accuracy. It is operational resilience, especially as variable generation takes a larger share of the grid.

Artificial intelligence has made forecasting appear more mechanical, but not necessarily more trustworthy. Recent work on weather and snowfall prediction suggests that machine learning can help identify bias in models and speed up corrections when combined with physics-based approaches.[4] That is a meaningful step. Yet it is also a reminder that AI forecasting works best as a hybrid system, not as a substitute for the underlying science. The machine can refine a model. It cannot repeal uncertainty.

This is where the current enthusiasm becomes risky. When organizations hear that AI can forecast better, they may quietly mean that it can produce more fluent confidence. That is not the same thing. In markets, supply chains, climate planning, and national security, the danger is less that models will never be useful than that decision-makers will treat them as if they erase surprise. If many actors converge on similar tools and assumptions, the system can become more correlated just when it needs diversity of judgment.[11][12]

What remains unverified, and what should be watched closely, is whether AI forecasting tools genuinely improve long-horizon decisions outside controlled research settings. The strongest evidence so far is in narrow domains with measurable feedback, such as weather, energy, or probabilistic event questions.[2][4][7] The weaker claim is that these systems can reliably map disruptive technological change years into the future. That would require evidence not just of better predictions, but of better outcomes when institutions act on them.

The deeper lesson is that forecast failure is often an institutional failure, not merely an intellectual one. People reward confident narratives, short time horizons, and forecasts that fit strategy decks. They reward the appearance of foresight more readily than the humility required for honest uncertainty. In that environment, the most valuable forecasting tools may be the ones that make leaders confront what they do not know, rather than the ones that make them feel ahead of the curve. AI may help with the former. It could also amplify the latter. That is the tension to monitor. Future technology history will not be written by the loudest predictions, but by the systems that learned, or failed to learn, how to live with uncertainty.[2][3][10][12]