Fooling Yourself: how narratives shape beliefs

The Cost of Narrative Clutter

When we make decisions, we rarely look at raw data in isolation. Instead, we consume information through narratives—coherent stories that wrap factual evidence in layers of context. While these stories help us make sense of complex worlds, they often introduce "nondiagnostic" details. These are facts that are technically true but carry zero weight in proving a specific conclusion.

Common intuition suggests that a rational person should simply ignore these irrelevant details. However, a new study from the University of Siena and several partner institutions suggests we are much worse at this than we think. The researchers report that when irrelevant details are embedded within a story, they don't just sit there quietly. They actively erode our confidence. They pull our beliefs back toward a state of maximal uncertainty.

The failure of rational filtering

In formal logic and Bayesian inference—the mathematical framework for updating the probability of a hypothesis as new evidence emerges—an uninformative signal should leave your existing beliefs unchanged. If you are 80% sure a suspect is guilty and someone tells you a fact that is equally likely whether the suspect is guilty or innocent, your certainty should remain at 80%.

Current cognitive models often assume that humans can successfully partition "diagnostic" information (data that narrows down possibilities) from "nondiagnostic" noise. Yet, in real-world environments like legal proceedings or corporate disclosures, this partitioning fails. The authors argue that the very structure of a narrative encourages the brain to search for connections between all available elements. This happens even when those elements lack probative value (the ability to prove something). This creates a "dilution effect," where the presence of extra, useless details weakens the impact of the actual evidence.

Isolating the narrative effect

To determine if this distortion is caused by the story itself or specifically by the inclusion of irrelevant clues, the authors implemented a controlled laboratory experiment with 379 participants. They utilized a between-subjects design (assigning different participants to different groups) with three distinct conditions to isolate the variables:

1. The Urn Condition: An abstract setting where participants tracked colored balls drawn from urns. This served as a baseline for pure probabilistic updating without any storytelling.
2. The Story Condition: A fictional theft investigation where participants received both diagnostic clues (pointing to a suspect) and explicit nondiagnostic clues (details that favored neither suspect).
3. The Story No Info Condition: The same theft narrative, but when a nondiagnostic event occurred, participants were simply told "no new information was available" rather than being given a useless clue.

By comparing these groups, the researchers could separate the effect of the "story frame" from the effect of the "clutter" itself. As shown in, the experiment moved from establishing initial priors (starting beliefs) to a ten-round sequence of belief updating.

Systematic reversion to uncertainty

The results show that the combination of a narrative and explicit nondiagnostic clues creates a unique psychological phenomenon: a "reset-to-uncertainty" rule. The authors find that in the Story condition, subjects do not just make random errors. They systematically revise their beliefs toward the 0.5 midpoint—the point of maximum uncertainty.

The paper reports that in the Story condition, subjects updated their beliefs following nondiagnostic signals 76.3% of the time. This was notably higher than the 61.7% update frequency in the Story No Info condition. Crucially, the authors use a behavioral structural model to show that these aren't just noisy reports. They estimate that nondiagnostic clues are processed as if they contained actual evidence roughly 12% of the time [Table 3].

This isn't a random drift. As illustrated in, the updates following nondiagnostic signals in the Story condition concentrate around the 0.5 midline.

They do not stay on the diagonal of "no change." Even when a participant has accumulated strong evidence for one side, a single useless narrative detail can pull them back toward a coin-flip's worth of certainty. This is further evidenced in .

Here, the Story condition shows a much flatter rate of declining uncertainty as priors become more extreme compared to the other groups.

The economic price of hesitation

The authors move beyond abstract psychology to ask: how much does this actually cost? They performed a Monte Carlo simulation to model a "sequential stopping problem." This is a scenario where an agent (like an investor or a manager) waits for enough evidence before committing to a costly action.

The simulation reveals that this "reversion to uncertainty" acts as a massive brake on decision-making. Because nondiagnostic clues constantly pull beliefs back toward the middle, agents require significantly more signals to reach a confident decision threshold. The paper finds that compared to the Story No Info condition, the Story condition increased the median number of signals required to reach a decision by 41% for moderate confidence levels. For high-confidence thresholds, the delay rose by 73% [Table 5]. In environments where time is money, these narrative "clues" effectively act as a tax on productivity and speed.

The verdict: filter, don't just disclose

Is this research ready for application? The answer is a qualified yes, particularly for institutional design. The study demonstrates that simply being "transparent"—telling people that a piece of information is irrelevant—is insufficient to stop the distortion. Even when participants knew the clues were nondiagnostic, the narrative context forced a retreat into uncertainty.

The practical takeaway is that organizations, courts, and financial regulators should prioritize the active filtering of irrelevant content rather than just labeling it. If you want a decision-maker to stay focused on the evidence, you cannot just provide a disclaimer. You must remove the clutter entirely.

Code for the analysis and the simulation is reportedly available; see the paper for the canonical link at https://github.com/paolopin/narrative-belief-updating.

Figures from the paper