Belief in AI predictions made people turn down certain cash in a series of controlled experiments, even though the system did not tell them what to do. Faced with a guaranteed extra US$1, many participants instead acted as if an undisclosed machine forecast had already fixed their choice. The finding suggests that a prediction can alter behaviour simply because people believe the predictor may know them.
That matters beyond a laboratory puzzle. Forecasts now appear in recruitment, credit, insurance, health and everyday digital services. A score does not have to issue an instruction to narrow the options somebody feels able to choose. If people treat a model’s output as evidence about their future behaviour, the forecast may start helping to produce the conduct it claims to anticipate.
How the guaranteed cash experiment worked
The research comes from Aoi Naito of Carnegie Mellon University and Hirokazu Shirado of the Institute of Science Tokyo. Their current preprint, revised in July 2026 and not yet peer reviewed, reports four preregistered online studies involving 1,305 unique participants.
The central experiment adapted a decision puzzle known as Newcomb’s problem. Participants could choose both boxes or take only Box B. Box A always contained US$1. Box B contained either nothing or US$3, depending on a prediction that had supposedly been made before the choice. Crucially, participants were not shown the prediction or the contents of Box B before deciding.
Taking both boxes always added the certain dollar. Whatever was in Box B, choosing both produced US$1 more than choosing only B. Yet participants were told that an AI system had predicted which option they would select and that its forecast had determined Box B’s contents. A comparison group was told that a random process had determined the contents instead.
The AI made no recommendation and disclosed no score. The researchers were therefore testing belief in predictive authority, not compliance with advice. That distinction separates this study from research showing that AI persuasion can change political actions after people read tailored arguments.
AI predictions changed the way people reasoned
In the first study, 41 per cent of participants in the AI condition chose only Box B, compared with 26 per cent in the random condition. The second study produced an even wider gap. About 42 to 45 per cent chose only B when an AI prediction supposedly controlled the box, while 15.3 per cent did so under either version of the random condition.
Across the first two studies, the odds of choosing only B were 3.39 times higher under AI prediction than under random determination. The participants were not merely answering an abstract survey. Their choices affected bonus payments, so rejecting Box A meant giving up money they could otherwise keep.
The paper estimates that this pattern reduced earnings by between 10.7 and 42.9 per cent relative to consistently taking both boxes, depending on the prediction regime. Those percentages do not mean an AI took money away. Participants voluntarily rejected a guaranteed addition because the predictive setup changed how they understood the decision.
That resembles a broader pattern in which AI-generated framing changes choices without physically limiting them. In another experiment covered by LiveAIWire, AI review summaries changed purchase intentions by shaping which product information stood out. Here, the forecast itself was hidden, yet the idea of being predictable still influenced behaviour.
This was not ordinary AI persuasion
Most discussions of AI influence focus on what a system says: a recommendation, explanation, ranking or emotionally targeted message. This experiment removed that content. Participants only knew that the system had made a prediction in advance. They did not know which behaviour it expected and could not follow a stated suggestion.
The effect therefore appears to come from reasoning about the predictor. A participant might think that if the model is accurate, taking only B is the route to the higher-value outcome. Yet at the moment of choice, Box B’s contents were already fixed. Taking Box A as well could not change them and could only increase the payout.
Participants nevertheless estimated the AI’s accuracy at roughly 62 per cent despite receiving no evidence about its reliability. The random mechanism was assessed at about 50 per cent, as expected. The gap suggests that some people supplied their own assumptions about machine competence and then reasoned within those assumptions.
This is also different from delegating a decision to software. Nobody handed control to the model. Research on AI access and dishonest behaviour has shown that an interface can create moral distance when people outsource actions. In the cash experiment, the machine’s mere presence as a predictor was enough to reshape an unaided human choice.
The effect extended beyond the box puzzle
A third study translated the same conflict into short scenarios involving a job interview, a mobile-data coupon and a freelance task. Participants could either preserve an option that was beneficial regardless of the forecast or sacrifice it in an attempt to align their decision with what a predictor might have anticipated.
Across those scenarios, 26.7 per cent made the one-box-type choice when told an AI had predicted them, compared with 10.6 per cent when no prediction was involved. A human expert had a still larger effect, with 36.6 per cent choosing the restrictive option. Predictive authority was therefore not uniquely mechanical. People also narrowed their choices when they believed a knowledgeable person had forecast their behaviour.
That result is an important restraint on the headline. The study does not show that people invariably trust AI more than humans. It shows that belief in an informed prediction can alter decisions, and that AI systems can occupy that influential role even without demonstrating accuracy.
Repeated failure did not erase the effect
The fourth study asked participants to play five rounds under a fixed prediction policy. Its design was registered in advance in an accessible preregistration. The predictor did not learn from participants during the experiment, allowing the researchers to examine whether behaviour changed while the underlying system stayed the same.
Even after five consecutive prediction failures in one condition, 30.6 per cent made the one-box choice in the final round. That was twice the 15.3 per cent benchmark observed in the earlier random condition. Failure weakened the premise that the AI knew what a person would do, but it did not remove the behavioural pull.
The researchers also calculated that behavioural adjustment could raise the predictor’s apparent accuracy from 50.7 to 59.2 per cent without any improvement to the model. People were changing to fit the prediction process. A system could therefore look more capable partly because users adapted themselves around it.
What this means for you
A forecast should be treated as information about uncertainty, not as a command or a description of destiny. When a hiring platform estimates that an applicant may leave, a lender predicts default risk or an app anticipates a purchase, the output may influence the person being assessed as well as the institution making the decision.
The practical question is whether the forecast changes the available outcomes or merely changes your beliefs about them. In the box task, the result was already fixed, so taking the guaranteed money could not make the other box worse. In real services, the relationship may be more complicated because institutions can respond to a score. Separating those mechanisms is essential.
Users should also ask what evidence supports a claimed accuracy rate. The participants inferred above-chance AI competence without receiving reliability data. Interfaces that display predictions without calibration, uncertainty or a clear explanation of consequences may encourage stronger conclusions than the evidence warrants.
Why the study does not prove AI knows your future
The experiments examined responses to a stated prediction setup. They did not demonstrate that the AI could infer a person’s choice from rich behavioural data, nor did participants interact with a deployed commercial predictor. The fourth study deliberately used a fixed policy, making it especially clear that apparent success could emerge from human adaptation rather than machine learning.
The samples were recruited online through Prolific, the financial stakes were modest and three of the studies used stylised decisions. The third study improved realism with everyday scenarios, but vignettes still cannot reproduce the pressure, information imbalance or long-term consequences of employment, finance and insurance decisions.
The paper is also a preprint. Its methods and conclusions may change through peer review, and the authors say data and materials will be made available upon publication. The full study record should therefore be read as emerging evidence rather than a settled rule about human behaviour.
Even with those limits, the result identifies a useful risk. Systems designed to forecast people can become part of the environment that shapes them. Work on AI behavioural twins often asks whether a model can reproduce human choices. This study asks the reverse question: what happens when people believe a model has already reproduced theirs?
Predictive authority can become self-reinforcing
The most significant lesson is not that humans are irrational whenever AI appears. It is that predictions can acquire causal force. A forecast may influence the behaviour used to judge its own quality, making prediction and intervention difficult to separate.
For organisations, that creates a measurement problem. If a risk score changes how staff treat a customer, and the customer then changes behaviour, later outcomes cannot be attributed to the model alone. Evaluation must account for feedback between the prediction, the decision-maker and the person being predicted.
For individuals, the safeguard is simpler: identify what the prediction actually changes. If a benefit remains available regardless of the forecast, surrendering it may only make the prediction feel more powerful. AI does not need to order people around to narrow their futures. Sometimes, belief that it already knows the future is enough.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.