Why We Trust Unreliable Sources: The Brain Distorts Stated Reliability When Deciding

Every day, people navigate a torrent of advice, predictions, and recommendations from sources of wildly varying quality. Weather apps, financial tipsters, online reviews, and social media accounts all claim to offer useful guidance, yet some are far more dependable than others. A new study published in PLOS Computational Biology by Keiji Ota, Anthony Ciston, Patrick Haggard, Thibault Gajdos Preuss, and Lucie Charles tackles a deceptively simple question: when a source explicitly tells you how reliable it is, do you actually use that information rationally when making a decision? The answer, it turns out, is a resounding no, and the details of how we go wrong are more surprising than anyone might have expected.

The research team built a novel laboratory paradigm designed to isolate exactly how explicit reliability markers shape human choices. Participants were asked to make a simple choice between two options, but before committing to a decision they saw predictions from several advisory sources. Crucially, each source came with an explicit, probabilistic label stating the likelihood that it would provide correct information. Some sources were flagged as highly trustworthy, others as moderately reliable, and some were even labelled as outright liars, meaning their predictions were known to point toward the wrong answer with high probability. This design allowed the researchers to ask whether participants took these stated probabilities at face value, weighting each source’s advice in proportion to its advertised trustworthiness, as an ideal Bayesian observer would.

The results revealed a systematic distortion. Participants did not use the stated probabilities of source correctness as advertised. Instead of treating a source labelled as unreliable as nearly worthless, they appeared to assign it a moderate degree of trust, as if the label had been softened or discounted. Computational modelling made this pattern precise: the researchers estimated the effective reliability that each participant implicitly assigned to every source, and these estimated values diverged substantially from the explicit markers. Sources described as trustworthy were not rewarded with the full weight their labels justified, while sources described as unreliable were not dismissed as thoroughly as they should have been. In effect, the entire reliability spectrum was compressed, with extreme labels pulled toward the middle.

This compression has a striking consequence: even sources explicitly flagged as unreliable biased participants’ choices. The advice of a known bad source nudged decisions in the direction it pointed, rather than being inverted or ignored. An ideal decision maker who knows a source lies most of the time should either disregard it entirely or, better still, treat its predictions as evidence for the opposite conclusion. Participants did neither. Their behaviour suggested that a source labelled as a liar was processed as if it were merely mediocre, a source of noisy but not systematically misleading information. The explicit warning, in other words, failed to translate into the appropriate inferential strategy.

The picture became even more intriguing when the researchers examined sources known to reliably predict the incorrect answer. The presence of these lying sources did not simply add noise; it impaired performance in a specific and measurable way. Using a drift diffusion framework, in which decisions are modelled as the gradual accumulation of evidence toward one option or the other, the team found that lying sources increased the leakiness of evidence accumulation. In a leaky accumulator, evidence decays over time rather than being preserved, so the decision process becomes less efficient and less able to build a clean, decisive signal. The lying sources did not just push choices in the wrong direction; they degraded the very machinery of the decision, making the integration of all available information sloppier.

Why might the human mind distort explicit reliability in this way? The authors suggest that participants may not take probabilistic labels at face value because the mapping between a stated number and the appropriate weight in a decision is not intuitive. A label saying a source is correct seventy percent of the time must be converted into a likelihood ratio, and that conversion may be systematically miscalibrated. Alternatively, people may harbour a general scepticism about the labels themselves, treating extreme claims, whether of high competence or of outright dishonesty, as exaggerated. Whatever the underlying cause, the result is a decision system that behaves as though the world is flatter and more moderate than the evidence declares it to be.

Yet the study is not simply a story of human irrationality. Alongside each choice, participants rated how much they felt a given source had influenced their decision, and these metacognitive judgements revealed genuine insight. Participants were aware of the pull that unreliable sources exerted on their choices, and they could evaluate how much a given source had increased the likelihood of their response. In other words, even though the integration of reliability information was biased at the level of the choice itself, people retained some access to the process that produced their decisions. They knew, at least approximately, which sources had swayed them and by how much.

This dissociation between biased performance and partially accurate metacognition is one of the most compelling aspects of the findings. It suggests that the distortion occurs early or implicitly in the decision pipeline, while a separate monitoring process tracks the outcome with reasonable fidelity. Metacognitive accuracy of this kind matters because it is the foundation of self-correction. A decision maker who knows that a dubious source nudged her choice is in a position to discount that influence on the next trial, or to seek better information before acting. The study implies that this corrective capacity exists in humans, even when the automatic integration of advice goes astray.

The implications reach far beyond the laboratory. Modern information environments are saturated with explicit and implicit reliability signals: verified badges, fact-check labels, star ratings, community notes, and reputation scores. If people systematically compress the reliability spectrum, treating flagged misinformation as merely mediocre rather than as inverted evidence, then the practical value of these labels may be far lower than designers assume. A warning label that shifts a source from perceived liar to perceived mediocrity still leaves that source exerting a misleading pull on belief and choice. The findings suggest that interventions may need to do more than attach probabilities to sources; they may need to teach people how to convert those probabilities into appropriate inferential weight, including the counterintuitive step of reading a known liar’s claims backwards.

By combining a carefully controlled paradigm, computational modelling of choice, and trial-by-trial metacognitive reports, the study offers a rare, quantified window into how humans consume advice in an uncertain world. The message is double-edged. Our decisions are demonstrably distorted by explicit reliability information, in ways that let even confessed liars bias our choices and degrade our evidence accumulation. But we are not blind to our own biases. Some part of the mind keeps watch, registering which voices pulled us and how hard, and that quiet awareness may be the raw material from which better information habits can be built.

Subject of Research: How explicit reliability information about advice sources is integrated into decision-making and metacognition

Article Title: How does explicit reliability guide choices? Effect of trustworthiness on choice and metacognition

Article References: Ota, K., Ciston, A., Haggard, P., Preuss, T. G., & Charles, L. (2026). How does explicit reliability guide choices? Effect of trustworthiness on choice and metacognition. PLOS Computational Biology, 22(10), e1014818. https://doi.org/10.1371/journal.pcbi.1014818

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014818

Keywords: decision-making, metacognition, source reliability, computational modelling, drift diffusion, trustworthiness, Bayesian inference, evidence accumulation, misinformation, cognitive bias, PLOS Computational Biology, psychology