Can AI Predict the Future? Humanity's New Oracle Explained

For thousands of years, people turned to prophets, seers and oracles to glimpse what lay ahead. Today we ask machines instead. Can AI predict the future, or does uncertainty remain the one thing no technology can overcome?

◆ In Summary

Artificial intelligence is becoming the most powerful predictive tool ever created, using vast amounts of historical data to forecast future events and behaviours. In many ways, it serves the same role as the prophets and oracles of the ancient world, offering guidance about what may happen next. The difference is that AI relies on statistical patterns rather than divine inspiration. While these systems can be remarkably effective, they remain limited by uncertainty, bias and incomplete information. The greatest danger may not be that AI predicts the future badly, but that people place too much confidence in predictions that were never certain to begin with.

I was putting together an article for the site when a thought stopped me. We spend a lot of time here on seers and prophets, on the figures who claimed to see what was coming and the traditions that grew up around them. Tiresias, Cassandra, the Pythia at Delphi, Nostradamus. We do not tend to take such figures seriously now. Anyone claiming to predict the future gets labelled as a conspiracy theorist or fantasist or worse, and just gets ignored.

Then it struck me. We have not moved on at all. We have just built something considerably more powerful than anything the ancient world could have imagined, and we are consulting it in exactly the same ways, for exactly the same reasons, with a reasonable chance of making exactly the same mistakes. Artificial intelligence will almost certainly turn out to be the most capable predictive tool in the history of humanity. The part that gives me pause is not whether it will work. It is what happens to human behaviour when it does.

The oldest question

Every prophetic tradition in history has been trying to solve the same problem: how do you extract reliable information about the future from the present? The Babylonians read the stars. The Greeks consulted oracles. The Romans observed the flight of birds. The underlying logic, across all of it, is that the future is not entirely hidden. That it leaves traces in the present for those who know how to read them, and that reading those traces correctly requires either divine access or exceptional skill. Or, now, a sufficiently large dataset.

What the oracles actually did

For several centuries, the Oracle at Delphi was the most consulted oracle in the ancient world. Kings made the journey. Generals made the journey. The questions were about wars, alliances, harvests, successions, the things that could not be left to guesswork and could not be known any other way.

Croesus of Lydia asked whether he should attack Persia. Back came the answer: if he did, a great empire would fall. He attacked. It did. His.

The questioner supplied the confidence the answer lacked. The priests controlled access and framed the questions, with every reason to keep things exactly as they were. What the Pythia sold, really, was a framework for believing what you already wanted to believe.

The ancient seers were not reliable. They simply made their unreliability invisible.

Can AI predict the future better than an oracle?

What AI does is not fundamentally different, which is either reassuring or alarming depending on what you think of oracles. It processes vast quantities of historical data, finds patterns invisible to human observers, and produces predictions about what is likely to happen next. Nobody at Delphi could do that. The difference is not the ambition. It is the capability.

Predictive models have identified disease outbreaks before human epidemiologists noticed them. Financial algorithms have caught market anomalies that trained human traders walked straight past. Climate models have forecast regional weather patterns with an accuracy that would have looked like sorcery two generations ago, and that still somehow fails to produce the policy responses it probably warrants.

The Pentagon runs AI-assisted tools to model geopolitical risk, and it is not alone. Governments use the same approach to anticipate civil unrest. Insurance companies use it to predict individual behaviour with an intimacy that should probably alarm more people than it does. The seer, in other words, has never been more capable.

The confidence problem

But AI has a problem the ancient seers did not. They hedged. Their answers were ambiguous by design, constructed so that almost any outcome could be made to fit the prediction after the fact. AI does the opposite. It asserts. It produces fluent, authoritative-sounding outputs that carry an air of certainty the underlying process does not always justify. It is trained on human-generated text, so it reproduces human patterns of expression, including, and this is the part that matters, the patterns that sound like confident knowledge.

Ask one a question about the future and it answers in the register of someone who knows. That register is not confidence. It is just the language it was trained on.

The confidence is in the language, not the prediction. The oracle at Delphi was at least cryptic about it. Large language models are wrong in plain English, with complete conviction, and the companies building them know it. Every major model now carries some version of the same disclaimer: this AI makes mistakes, please check the responses. The oracle is admitting it might be wrong. The ancient seers never did that.

Who controls the oracle

The oracle at Delphi was run by the priests of Apollo, and they did very well out of it. They controlled access, shaped the questions, interpreted the answers, and had every reason to keep the arrangement exactly as it was.

The Sibylline Books were another version of the same system. Kept under strict priestly guard, brought out only for the Roman Senate in moments of crisis. Ordinary Romans never saw them. The guidance existed; the question of who got to hear them was always political.

The parallel with AI is not subtle. The most powerful systems in the world are controlled by a small number of technology companies. The models that governments, financial institutions and militaries consult are products of organisations with commercial interests, political relationships and reputational concerns. The predictions that reach decision-makers come from systems built by people who are not neutral observers of the outcomes. That is not a conspiracy. It is simply how oracles have always worked. Access to predictive power has never been equally distributed, and the institution that holds it has never been without interests of its own.

Can it ever do what the oracles never could?

Whether AI's current unreliability, the hallucinations and the confidently wrong answers, is a problem of this moment that better models will eventually solve, or whether it is structural, baked into what these systems fundamentally are, is the question this piece has been building towards.

The honest answer is that nobody knows. The serious predictive models, the ones running inside hospitals, hedge funds and defence departments, are considerably more reliable than anything available to a general user. The gap between what AI can do at its best and what most people experience it doing is, for now, enormous. That gap may close. It may not close completely. And the consequences of trusting a system that sounds certain when it isn't are, in some domains, considerable.

The ancients acted on prophecy without understanding the mechanism behind it. The riddles left room, a gap between the oracle's words and the decision that followed, and in that gap human judgement still operated. Precise prediction fills the gap entirely.

The ancient seers were consulted for thousands of years despite being unreliable, because the alternative was knowing nothing. We are not in that position. We have other tools. The question is whether we will use AI as one of them, carefully, or whether we will do what humans have always done with oracles, supply the confidence the answer lacks, and hear what we wanted to hear.

Frequently Asked Questions

Can AI actually predict the future?

AI can forecast probabilities based on patterns in historical data, which is genuinely useful for things like disease outbreaks, weather systems and market anomalies. It cannot predict specific, unprecedented events with certainty. The forecasts are statistical estimates, not revelations, however confident the output sounds.

How is AI similar to ancient oracles like the one at Delphi?

Both function as an authority people consult when they want reassurance about an uncertain decision. The Oracle at Delphi read omens and interpreted them through priests; AI reads historical data and interprets it through statistical models. In both cases, the questioner tends to supply the confidence the answer itself doesn't guarantee.

Why do AI predictions sound so confident even when they're wrong?

Large language models are trained on human-written text, so they reproduce the fluent, assertive tone of confident human writing regardless of how certain the underlying prediction actually is. That register is a property of the language, not a measure of the model's accuracy.

Who controls the AI systems used to make predictions?

A small number of technology companies build and control the most powerful predictive models, and governments, financial institutions and militaries rely on systems built by organisations with their own commercial and political interests. Access to predictive power has never been evenly distributed, in the ancient world or now.

◆ Also In The Stars