Artificial intelligence is increasingly used to forecast stock prices, health outcomes, and consumer behavior. But what happens when people start to believe AI can predict their personal choices? A recent study by Naito and Shirado suggests this belief alone may be enough to change how we think—and what we believe is possible.
The Experiment: Newcomb's Paradox Meets Machine Learning
The researchers adapted a classic thought experiment called Newcomb's paradox, which pits free will against prediction. In their version, participants faced a choice: take a guaranteed smaller reward, or gamble on a larger one that an AI predicted they would not choose.
The twist? The AI's prediction was revealed before participants decided. If the AI said you would turn down the gamble, and you believed it, you might paradoxically prove it right—simply by following the prediction.
Striking Results
Over 40% of the 1,305 participants treated the AI as a legitimate predictive authority about their own behavior. This belief dramatically shifted their choices: those who saw AI predictions were 3.4 times more likely to forgo the guaranteed reward, reducing their earnings by up to 43%. Remarkably, this effect persisted even when the AI repeatedly made incorrect predictions.
The researchers found this across different presentations of the AI and various decision contexts. The mere presence
Why This Matters
As AI prediction tools spread into hiring, lending, criminal justice, and personal health, this research raises urgent questions. When an algorithm forecasts someone's risk of reoffending, defaulting on a loan, or developing a disease, do people start to see those outcomes as inevitable? The study suggests they might—and that this perceived inevitability can become self-fulfilling.
The researchers describe this as AI "narrowing the futures individuals consider." Rather than expanding options through better information, predictive AI may unconsciously limit what people believe they can choose.
One question this research leaves open: If simply believing AI can predict our choices changes how we make them, should systems that generate behavioral forecasts carry warnings about their potential psychological impact on the very people they assess?