Every society we have ever excavated built the same machine. Not the wheel, not the plough — those came later, and unevenly. The first machine was for telling people what to do next.
In Shang dynasty China, it was a heated ox bone, cracked open and read for its fractures. In Rome, it was the liver of a sacrificed animal, or the flight path of birds crossing the sky. In Delphi, it was a woman seated over a fissure in the rock, breathing something the priests would translate into verse. Mesopotamia had its own version four thousand years ago. None of these cultures were in contact with each other when they invented it. They arrived at the same idea independently, which is usually how you know an idea isn’t cultural. It’s structural. It’s about us.
The structure is this: uncertainty is unbearable, and it is cheaper to outsource the unbearable than to sit inside it.
Psychologists have a dry name for this. Arie Kruglanski calls it the need for cognitive closure — the desire for a fast, firm, final answer to a question, any answer, so long as the not-knowing stops. It comes with two settings. The first is urgency: grab the nearest available answer as quickly as possible. The second is permanence: once grabbed, defend it, keep it, stop looking. Put those two settings together and you don’t get a rational actor weighing evidence. You get a person standing in front of an oracle, asking a question they have already decided they want answered.
This is the part worth sitting with before we talk about AI at all. The oracle was never really in the business of accuracy. It was in the business of ending the discomfort of not knowing. Whether the Pythia’s mutterings had any relationship to the future was almost beside the point; the point was that the fog of war, or the uncertainty of a harvest, became bearable the moment someone official pronounced on it. Certainty was the product. Truth was, at best, a rumour about the product.
So: is AI the oracle we were looking for?
Functionally, yes, disturbingly well. It has no temple hours. It doesn’t require a sacrifice, a pilgrimage, or a priesthood to translate its output into something usable — it just answers, instantly, in your own language, with the specific confidence of something that has never once said “I’m not sure” unless asked to perform humility. And the research on how we respond to it is not subtle. Give people an answer with no source attached and tell them it came from an AI, and they will follow it even when it contradicts what they already know to be true — in one study, accuracy on a reasoning task actually fell once AI advice was introduced, from 27 percent correct to 9 percent. People consistently rate AI answers as more confident than identical human ones. And using it seems to make us worse, not better, at admitting the words “I don’t know” — the exact opposite of what a good advisor should train into us.
None of that required the AI to be right. It only required it to sound like an oracle. We built the delivery mechanism our psychology was already waiting for, and it turned out to be extremely good at delivering.
Individually, the cost shows up as a kind of quiet atrophy. Cognitive offloading isn’t new — we’ve offloaded memory to writing, arithmetic to calculators, navigation to satellites, without civilisation collapsing. But early studies on AI-assisted reasoning are finding something sharper: heavier use correlates with measurably lower scores on critical thinking, and the effect runs specifically through offloading, not through some unrelated third factor. We are not just outsourcing the answer. We appear to be outsourcing the muscle that used to check the answer.
Socially, the change is stranger than anything the old world had to contend with. Ancient societies had many oracles, disagreeing with each other by region, by school, by priesthood. A augur in Rome and a bone-reader in Anyang were never at risk of synchronising the entire species onto one interpretation of reality. Today’s oracle is closer to singular — a small number of models, trained on overlapping data, consulted by billions of people who each believe they are getting a private, personal answer. When everyone asks the same oracle the same kind of question, disagreement itself starts to look like a bug rather than a feature of a healthy society. We may be building, for the first time, a single point of failure for human judgement.
I don’t know whether that makes AI a good oracle or a uniquely dangerous one. I suspect the honest answer is that it makes it exactly as good as every oracle before it: convincing, comforting, occasionally right, and never quite as authoritative as the silence it was hired to fill.
Delphi never closed. It just got an API.
Sources: Kruglanski, “Need for Cognitive Closure” and related research (Kruglanski Arie research archive; PMC, “Factor Structure and Internal Consistency on a Reduced Version of the Revised Test of Need for Cognitive Closure,” 2022); Human Relations Area Files, “Coping with Uncertainty: Oracles, Divination and Decision-Making”; Current Anthropology, “Divination and Power: A Multiregional View of the Development of Oracle Bone Divination in Early China”; ScienceDirect/Computers in Human Behavior, “Trust and reliance on AI — An experimental study on the extent and costs of overreliance on AI” (2024-26); phys.org, “People overestimate how confident AI systems are in their responses” (May 2026); The Register, “Using AI makes people less likely to admit they don’t know something” (July 2026); Springer Cognitive Processing, “Cognitive offloading, critical thinking and attitudes towards artificial intelligence in the era of ChatGPT” (2026).
Leave a Reply