ProBackend
Psychology & Behaviour
3 weeks ago7 min read

The AI Impact on Human Psychology Begins Before the Algorithm: Why We Cannot Discard Data We Know Is Wrong

The AI Impact on Human Psychology Begins Before the Algorithm: Why We Cannot Discard Data We Know Is Wrong

A 2026 PLOS Computational Biology study by Ota et al. reveals that humans cannot ignore information labeled as having a 50% chance of being correct — even when they consciously recognize the bias. The finding has direct consequences for how AI systems interact with human cognition.

A Coin Flip You Cannot Uncaringly Ignore

Here is the experiment: you see a series of colored squares. Each square carries a label — "80% reliable," "50% reliable," "35% reliable." The 50% label means what it says. A coin flip. Zero diagnostic information. A mathematical null.

And yet the people in the study still chose the blue option more often when the coin-flip squares were blue.

This is the central finding of a 2026 study by Keiji Ota and colleagues at University College London and Queen Mary University of London, published in PLOS Computational Biology. Five online experiments, hundreds of participants, and one stubborn conclusion: telling someone that information is worthless does not make them treat it as worthless. They know it's worthless. They can even articulate that it's worthless. And they still let it pull their decisions off-center.

The result matters far beyond a color-perception lab task. If you want to understand the AI impact on human psychology, you have to start with this — the pre-algorithmic machinery of how humans integrate source reliability. AI systems inherit the decision architecture of the humans who build them, train them, and interact with them. And that architecture has a bug at its core.

How the Experiment Was Built

Ota's team designed a paradigm where participants viewed predictions from sources of varying stated reliability, then made a binary choice between two options. The sources were color squares, each tagged with a probability that its color correctly predicted the answer. Some sources were explicitly marked as "reliably wrong" — they were known to consistently lie. One category was labeled at exactly 50%, meaning the information carried no predictive value whatsoever.

After making their choice, participants rated on a visual analogue scale how much each reliability level had influenced their decision. This was the key metacognitive measurement: researchers could compare what actually influenced choices against what people thought influenced their choices.

The task was simple enough to repeat hundreds of times, complex enough to produce meaningful variance, and abstract enough to isolate reliability weighting from prior knowledge about the content itself.

The Distortion at the Heart of Reliability Weighting

Participants did not use reliability labels at face value. Computational modeling revealed a systematic distortion: people compressed the range of stated reliabilities toward the middle. The 80% source was treated as though it were somewhat less reliable than advertised. The 50% source — mathematically equivalent to noise, was treated as though it carried some meaningful signal.

The effect was measurable and robust. In the primary experiment, the beta coefficient for unreliable-source influence was 0.36 with a 99% confidence interval of [0.29, 0.43]. That is not noise in the data. That is people consistently allowing a coin flip to nudge them.

Why? The paper offers three candidate explanations. First, mere exposure: repeated presentation of a stimulus increases preference for it, regardless of its informational content. If blue squares keep appearing, blue starts to feel like the answer, even when you know the source generating them is random. Second, Gricean cooperation: human communication operates on the deep assumption that a speaker is trying to be informative. A labeled source feels like a communicative partner. The brain has difficulty modeling a source that genuinely produces nothing but entropy. Third, decisional utility: when evidence is ambiguous, people may treat random signals the way they treat a coin toss to break a tie. The unreliable source doesn't inform the decision, but it resolves the hesitation.

All three explanations converge on the same practical implication: a reliability label is not a switch that turns off cognitive processing of the underlying content. It is, at best, a dimmer knob. And a dimmer stuck at 30% still lets plenty of light through.

The Metacognitive Paradox

Here is where it gets genuinely uncomfortable. Participants could feel the bias. When asked to rate how much a given source influenced their choice, their subjective ratings correlated with the objectively measured influence of that source. They knew, to some degree, that the 50% sources were contaminating their judgment. They could even quantify, approximately, how much.

But awareness did not confer control. The bias persisted across all five experiments. Knowing you are being influenced by noise did not stop the noise from influencing you.

This is what cognitive scientists call a dissociation between awareness and cognitive control. The metacognitive system, the machinery that monitors and evaluates one's own thinking, detected the contamination. The executive system responsible for suppressing it could not fully execute the correction. The researchers frame this as evidence that people "distort explicit source reliability but have some awareness of this process."

The phrase "some awareness" does a lot of quiet work there. It is not the "some awareness" that saves you. It is the "some awareness" that makes you complicit in your own bias, able to name it, unable to stop it.

What This Means for AI and Human Psychology

This study sits at a precise intersection of what AI in psychology actually means in practice. Not the chatbot-on-a-thesis question, but the engineering question: if you build a system that displays reliability labels, confidence scores, source ratings, uncertainty flags, you are designing for a cognitive architecture that systematically misreads those labels.

The parallel to AI systems is not metaphorical. It is structural. Researchers have found that when LLMs are fine-tuned on explicitly labeled falsehoods, the models absorb those claims into their internal representations anyway. See when you tell an AI this is false, the negation-neglect literature shows that both humans and language models treat explicit negation or discounting as a weak signal relative to the raw content exposure. The human "mere exposure" effect and the model's inability to learn from negated training examples may share a common information-theoretic root: content dominates metadata.

This also connects to the broader research on how AI outputs reshape human memory and perception. A Georgetown and University of Washington study found that misleading AI video summaries dropped eyewitness accuracy from 84% to 45%, see when AI summaries rewrite truth. The mechanism is the same one Ota identified: the human cognitive system integrates what it receives, and explicit warnings about source reliability are too weak to override that integration.

When we ask "can AI understand human psychology," the honest answer starts with this asymmetry. AI systems do not share the human bias toward treating unreliable sources as informative. But human users of AI systems do. And that is the design constraint that matters.

Designing for the Brain We Actually Have

The content-moderation implications are straightforward and uncomfortable. Warning labels on misinformation do not neutralize the misinformation. Banners identifying the source of online news items have been shown to fail at decreasing belief in fake news. Pre-bunking and de-bunking interventions are insufficient to prevent the biasing effect of misinformation on downstream judgments.

If you want people to ignore unreliable information, telling them it is unreliable is not enough. The research points toward stronger interventions: removing the content entirely (because exposure itself carries biasing power), redesigning decision environments so that unreliable signals never enter the perceptual field, or building tools that do the reliability weighting correctly on behalf of the user.

The third option is where AI enters the picture not as a threat but as a prosthetic. A system that correctly down-weights a 50% signal to zero, something the human brain demonstrably cannot do, could handle the reliability-integration step that our cognitive architecture is poorly suited to perform. The catch is obvious: the human still has to trust the AI's reliability judgments, and trust is exactly the quantity this study shows we systematically miscalibrate.

We are trapped in a loop. Our brains distort reliability signals, which makes us vulnerable to unreliable sources, which is the exact vulnerability AI systems both exploit and are built to mitigate. The AI impact on human psychology is not only what algorithms do to us. It is what our pre-existing cognitive architecture does to the algorithms we build, and the confidence scores, warning labels, and uncertainty flags we rely on to manage what we read.

Ota's experiments do not resolve this. They name it. And naming a cognitive bug with precision is the first requirement for designing around it rather than pretending it does not exist.

a coin flip you cannot uncaringly ignorea coin flip you cannot uncaringly ignore