AI Prediction is Probability not Prophesy

The danger is neither Oracle nor algorithm, but unaccountable predictions.

An AI-Pythia or Pythoness, the title of the high priestess of the Temple of Apollo at Delphi in Ancient Greece, who served as its oracle. Image via MidJourney.

Writing in the Economist (21^st April 2026) Professor Carissa Véliz calls AI ‘the new Oracle of Delphi’, a ‘false prophet’ whose predictions threaten justice, finance, insurance, medical and recruitment systems, human agency, and ultimately, freedom. She is right to call for debate. But her argument rests on a mistake about what prediction is. Oracular prophecy promises a pre-determined future. Prediction offers odds, probabilities, quantified uncertainty, and enables a calibrated response to it. The dangers Véliz identifies are real, but they arise when institutions convert forecasts into decisions without accountability - a problem of power, not prediction.

We agree with Veliz when she says that better prediction confers competitive advantage. Similarly, that we must “accept that we don’t know what the future holds, and act accordingly”. Those of us applying the science of prediction are not claiming to know the future. Prediction is not like skipping forward in a book to find out what happens and then coming back to tell everyone with total certainty. Rather, prediction claims that some possible futures are more likely than others, and that it is useful to distinguish them.

The Delphic oracle was powerful because its utterances were inscrutable, insulated from correction and uttered with priestly authority. A well-made prediction is the opposite. It is precise enough to be scored, corrected and challenged, its authority rests on demonstrated accuracy and calibration, not faith.

Nor is prediction alien to human judgment. It is part of human judgment. Every decision contains a forecast: if I do this, I expect that to follow. Courts, banks, insurers, doctors, employers, journalists and governments all do this already its just rare they make the forecast explicit.

Unlike Oracles, or forecasts hidden in narrative, modern prediction, human and machine, is accountable. To say there is an 80% chance a company will hit its sales target, a 4% chance a borrower will default, or a 1% chance of a medical emergency is not to know the future. It is to state a degree of belief about it. Across many cases, such beliefs can be scored. When a forecaster says “80%” often enough, roughly eight in ten of those events should occur. If they do not, the forecaster is miscalibrated. We learn how much we should trust the predictor, while scoring prediction error allows a predicting agent, human or algorithmic, to learn and improve.

Véliz says that when we predict people’s futures “as if we were forecasting the weather”, we treat them as inert objects rather than agents. This is impossible to agree with. My predicting the result of the next Newcastle match does not treat the players as inert: I am predicting the result of their skill and agency, and their opponents’. If my model and information are more accurate, I will tend to make better guesses; and if I am a gambler or coach, those guesses can be used to make money or improve performance.

The claim that predictions are closer to commands than descriptions is equally untenable. My prediction that Newcastle will win or lose is not a command; if it is, the players are stubbornly able to ignore it. When someone is refused a loan because an algorithm considers them too likely to default, it is not the algorithm that refuses the loan. It is the lender, using the algorithm as evidence. If the algorithm is sub-optimal, then there is free money on the ground: others can and should lend to the rejected creditworthy borrowers.

Véliz later claims that AI may achieve “accuracy” by creating the reality it predicts. But if a prediction can create a reality, that is a power of the predicting agent, not of the prediction. An abuse of power, not a forecasting problem. And the market examples are against her. If a person is deemed unemployable by AI but you, working at a rival, correctly predict they are employable, you win by hiring them. Her argument works only if institutions all use sufficiently similar systems, or if excluded people have no realistic route to generate contrary evidence. That is a problem that exists today, and long predates AI; a problem of error, monopoly or institutional design – of power - not prediction.

The same applies to her credit and health examples. If any lender refuses a loan because they predict a bleak financial future, the same objection arises; the AI is not doing the moral work. Any lender must predict repayment unless indifferent to being repaid. Likewise, a prediction of future disease may raise insurance costs, but the same would apply to a doctor’s warning about hereditary risk. The issue is regulatory: if we want flat, information-light risk-pooling, we can legislate for it. Better prediction is not necessarily or merely a weapon used by one party against another: it can reduce mispriced lending, protect likely defaulters from ruinous debt, enable earlier medical intervention, and help insurers estimate aggregate risk, even if we restrict individualised pricing.

Véliz asks whether, in justice, we should prefer transparent and contestable criteria to statistical pattern-matching. The only answer is: whichever decision procedure best prevents crime, acquits the innocent and convicts the guilty, subject to the chosen burden of proof. Courts already infer from incomplete evidence. Witnesses, juries, experts and judges are all predictive agents. Rather than prediction versus dignity, the issue is more mundane: which predictive procedure works best, with which thresholds, feedback and safeguards.

Increasingly accurate machine forecasts raise questions as to how our institutions adapt. If every decision is an if/then forecast of the future, and machines can out-predict us then they can significantly augment human decision-making. Every article in the Economist and other publications contain multiple predictions that X caused Y, that Z consequences will likely follow. Veliz’s article a case in point, predicting that AI prediction will threaten justice, finance, insurance, medical and recruitment systems, human agency, and ultimately, freedom.

Most published forecasts are not easily legible – they are implicit, unstated, hidden in narrative. The increasing application of machine forecasting is likely to expose this. Similarly, companies and executives that keep the predictions inherent in their decisions hidden will be outcompeted by others that adopt the algorithms. Journalists that hide their forecasts will cease to be trusted, Government policies, political manifestos - all can be subjected to the same discipline. AI is the new Oracle of Delphi only in the sense that it will be considered a negligent ill-omen not to consult it.

There may still be cases where we might prefer a human only or narrative forecast, even if a numerical or machine alternative performs better on a given metric(s). But this is a deliberate and consequential trade-off. If statistical pattern-matching produces consistently better predictions than more ‘human’ heuristics, then whatever is predicted is better done so by the model. That may be politically or aesthetically uncomfortable, but discomfort is not an argument. We should beware false oracles; we should also beware pretending that human beings are not already, necessarily, predicting all the time – in every decision we make.

And visit us at www.cassi-ai.com