In a breakthrough that blurs the line between science fiction and reality, researchers have developed an artificial intelligence (AI) system capable of predicting human decisions with uncanny precision.

Dubbed Centaur, this AI doesn't just guess whether a user will click an ad. It anticipates how humans will navigate complex moral dilemmas, learn new skills or even strategize in unfamiliar scenarios. The implications are staggering – from revolutionizing marketing and education to raising urgent ethical questions about privacy and free will.

Centaur, detailed in a study published July 2 in Nature, was trained on a staggering dataset — 60,000 people making over 10 million decisions across 160 psychological experiments. Unlike traditional models that specialize in narrow tasks like predicting stock trades or gambling habits, Centaur operates as a general predictor of human behavior. It outperforms decades-old cognitive models, suggesting AI may soon understand us better than we understand ourselves.

The system was built by fine-tuning Meta Platforms' Llama 3.1 language model – the same technology behind ChatGPT – using a technique that modifies only a fraction of its programming. Remarkably, the training took just five days on a high-end processor – a testament to the accelerating power of machine learning.

Centaur didn't just match existing psychological models; it demolished them. In head-to-head tests, it predicted human choices more accurately than 14 specialized cognitive and statistical models in 31 out of 32 tasks. Even more striking, it adapted to new scenarios it had never encountered such as altered versions of memory games or logic puzzles.

This adaptability suggests something profound. Human decision-making, for all its complexity, follows underlying patterns that AI can decode.

As one researcher noted, the human mind is "remarkably general" – capable of both mundane choices (picking breakfast cereal) and monumental ones (curing diseases). Centaur's success implies that our behavior may be more predictable than we'd like to admit.