🧬 biology

The adaptive nature of confirmation bias

This paper demonstrates that by modeling observations within a quantum probability framework, the phenomenon of confirmation bias emerges as an optimal, rational strategy in binary hypothesis testing that simultaneously minimizes memory requirements and exponentially reduces error probabilities.

Original authors: Dorje C. Brody, Karl J. Friston, Bernhard K. Meister, Emmanuel M. Pothos

Published 2026-06-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Dorje C. Brody, Karl J. Friston, Bernhard K. Meister, Emmanuel M. Pothos

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Idea: Is "Confirmation Bias" Actually Smart?

We usually think of confirmation bias as a flaw in human thinking. It's when we only look for information that agrees with what we already believe and ignore anything that contradicts it. We think this makes us irrational, like a stubborn mule refusing to see the open gate because it's fixated on the closed one.

This paper argues the opposite: Confirmation bias might actually be a superpower of rational thinking.

The authors suggest that if you are trying to make the best decision possible while using the least amount of mental energy and memory, your brain should naturally gravitate toward information that confirms your current hunch. In fact, doing so allows you to learn faster and remember less.

The Toolkit: Thinking Like a Quantum Computer (Without the Hardware)

To prove this, the authors use a mathematical framework usually reserved for quantum physics. They don't say the human brain is a quantum computer. Instead, they say that when we face complex choices, our minds behave as if they are navigating a "quantum" space.

The Analogy: Imagine you are trying to find a specific flavor of ice cream in a massive freezer.
- Classical Thinking: You check every single tub one by one. You have to remember every flavor you've already tasted to avoid repeating them. This takes a lot of memory and time.
- The Paper's "Quantum" Thinking: Imagine you can look at the freezer in a way that lets you check multiple flavors at once, or where the act of looking at one flavor changes the state of the others. The math they use (Hilbert spaces and matrices) models this kind of "superposition" where different pieces of information don't just sit side-by-side; they interact in complex ways.

The Two Big Surprises

When the authors ran their math on how a "rational" agent should gather evidence, they found two surprising evolutionary advantages to having a confirmation bias:

1. The "Forgetful Genius" (Minimum Memory)

Usually, we think we need to remember every piece of evidence we've ever seen to make a good decision. If you are trying to decide if it will rain, you might think you need to recall every cloud you saw this morning.

The paper claims that if you use an optimal "confirmation bias" strategy, you don't need to remember the past at all.

The Metaphor: Imagine you are playing a game where you have to guess a secret number. If you play "classically," you have to write down every number you've guessed and the result of every guess to figure out the next step.
The Paper's Finding: If you play "optimally" (with a bias toward what you currently think is true), you only need to remember the very last guess. You can forget everything else. Your brain effectively "forgets" the history because your current belief already contains all the necessary information from the past. This saves a massive amount of "mental RAM."

2. The "Exponential Speed-Up" (Faster Learning)

The second advantage is about how quickly you stop making mistakes.

The Metaphor: Imagine you are trying to find a needle in a haystack.
- Without Bias (Classical): You search randomly. To double your chances of finding the needle, you have to double the size of the haystack you search. To get 99% sure, you might need to search a huge area.
- With Optimal Bias (Quantum-style): The authors show that if you choose your evidence based on your current belief, your chance of being wrong drops exponentially.
The Result: It's like having a magic compass. Every time you take a step based on your "biased" compass, you get closer to the truth much faster than if you were just wandering randomly. You need far fewer samples (less time, less data) to be confident in your decision.

How It Works: The "Newspaper" Example

The authors use a simple example to explain this. Imagine you are trying to predict if energy prices will go up or down. You have two newspapers:

Paper A: Always says prices will go up.
Paper B: Always says prices will go down.

You have a "prior belief" (a gut feeling) that prices will go up.

The Irrational View: A rational person should read a neutral paper that says nothing, or read both equally to stay balanced.
The Paper's View: A rational person should read Paper A (the one that agrees with them).
- Why? Because reading Paper A gives you the most "information gain" relative to your current belief. It confirms your view with the highest efficiency.
- If you read a paper that disagrees with you, it creates "noise" and confusion (entropy), which wastes mental energy.
- By sticking to the paper that aligns with your belief, you sharpen your decision-making process, reduce your error rate, and don't need to remember every article you've ever read—just the last one.

The "Active Inference" Connection

The paper also links this to a concept called Active Inference. This is the idea that the brain doesn't just passively wait for the world to tell it what's happening; it actively seeks out the world to confirm its predictions.

The authors show that whether you try to:

Minimize the chance of being wrong (Error Probability), OR
Maximize the amount of new information you get (Information Gain),

...you end up with the same strategy: You naturally seek out evidence that confirms your current beliefs.

The Catch (What the Paper Does NOT Say)

It is important to note what the paper doesn't claim:

It does not say that believing fake news is good. The authors explicitly state their model assumes the information sources are "genuine" (honest signals with some noise). If the information source is designed to lie or deceive (like a troll farm), then this "rational bias" can be exploited to lead you astray.
It does not say that all human biases are rational. It only explains a specific type of active confirmation bias in the context of binary choices (Yes/No, Up/Down) where the brain is trying to be efficient.

Summary

In simple terms, this paper suggests that our brains are not broken when we seek out information that agrees with us. Instead, we might be highly efficient engineers. By focusing on what we already believe, we can:

Forget the past (saving memory).
Learn faster (reducing errors exponentially).

What looks like a stubborn refusal to change our minds might actually be our brain's way of solving a complex math problem to save energy and make the best decision possible as quickly as possible.

Technical Summary: The Adaptive Nature of Confirmation Bias

Problem Statement
Confirmation bias, the tendency to favor information consistent with prior beliefs, is traditionally viewed as an inferential error that precludes rationality, particularly within the framework of Bayesian inference. While existing literature attributes this bias to bounded rationality, working memory limitations, or reward structures, these explanations often fail to account for the potential rational status of confirmation bias independent of specific reward contexts. Furthermore, standard Bayesian models struggle to explain why such biases might offer evolutionary advantages, such as reduced memory requirements or accelerated error reduction. This paper addresses the question: Can confirmation bias be formulated as an optimal, adaptive strategy for information seeking and hypothesis testing?

Methodology
The authors reformulate the problem of binary hypothesis testing and evidence selection within the mathematical framework of quantum probability, specifically utilizing the space of square-root probabilities (a Hilbert space).

Representation: Unlike classical probability where observations are modeled as random variables on a probability space, observations here are represented by matrices (operators) acting on a Hilbert space.
Incompatible Observations: The framework allows for a family of mutually inequivalent information sources (observations) that do not admit a joint probability distribution. This is achieved by representing different evidence sources as non-commuting matrices.
Rational Updating: Belief updating is defined by the criterion of biological efficiency, equating survival with the minimization of energy consumption (and thus information processing cost). A rational update is defined as the transformation of a prior belief to a posterior belief that minimizes the "distance" (specifically the Bhattacharyya distance or angular separation in Hilbert space) between the prior and posterior, subject to consistency with the evidence. This leads to the von Neumann-Lüders projection postulate, which generalizes the classical Bayes rule.
Active Inference: The study extends to active inference, where a decision-maker actively selects the evidence source (parameterized by an angle $\theta$ $θ$ ) to optimize a specific criterion. Two criteria are compared:
1. Minimizing the expected probability of error in hypothesis selection.
2. Maximizing the expected mutual information (information gain).

Key Contributions and Results

Rational Origin of Confirmation Bias:
By minimizing the expected error probability in a binary hypothesis testing scenario, the authors derive an optimal evidence selection strategy. The results show that the optimal choice of evidence ( $\theta^*$ ) is not neutral but shifts toward the hypothesis with the higher prior probability. As the prior probability $p$ approaches 1, the optimal evidence choice converges to the parameter associated with that hypothesis. This mathematically demonstrates that an active confirmation bias is a rational consequence of minimizing error probability when dealing with non-commuting (incompatible) information sources.
Exponential Reduction in Error Probability:
In sequential evidence gathering, the paper establishes that an optimal level of confirmation bias leads to an exponential reduction in the error probability as a function of sample size $n$ .
- Classical/Unbiased Case: Without the ability to select evidence based on priors (or with fixed, unbiased sources), error probability decays slowly.
- Optimal/Confirmation Bias Case: When evidence is selected optimally (biased toward the current posterior), the error probability $P^{(n)}$ scales asymptotically as $\exp(n \ln(\cos^2(\delta/2)))$ , where $\delta$ is the separation between hypotheses. This represents a significant acceleration in learning compared to classical counterparts.
Minimal Memory Capacity:
A remarkable finding is that the optimal sequential strategy with confirmation bias reduces the memory requirement for inference.
- In a standard sequential process without active selection, the posterior depends on the total count of outcomes, requiring memory that grows linearly with $n$ (or exponentially in terms of history states $2^n$ ).
- Under the optimal confirmation bias strategy, the posterior probability after $n$ steps depends only on the outcome of the final trial. The decision-maker effectively "forgets" the specific history of previous trials, needing only the current posterior and the new observation. This reduces the memory complexity from exponential ( $2^n$ ) to a constant (requiring only the state of the most recent event).
Equivalence of Criteria:
The paper demonstrates that in this Hilbert space framework, the strategy that minimizes the expected error probability is identical to the strategy that maximizes the mutual information between the observation and the hypothesis. Thus, confirmation bias is rational under both "error minimization" and "information maximization" (active inference) criteria, provided the inference is conducted over non-commuting observations.
Quantum Supremacy in Cognition:
The authors argue that the exponential reduction in error probability constitutes a form of "quantum supremacy" in cognitive processing. Crucially, this does not require physical quantum resources (e.g., quantum particles) but rather the behavioral property of incompatibility—the inability to simultaneously assess two propositions with a joint probability distribution. This suggests the human brain may exploit the mathematical advantages of quantum probability structures for efficient inference.

Significance and Claims
The paper claims to provide the first operational framework for "active quantum inference," offering a protocol where optimal evidence selection is sought over the space of matrices.

Evolutionary Advantage: The findings suggest that confirmation bias is not merely a cognitive flaw but an adaptive mechanism that offers significant evolutionary advantages: it minimizes the metabolic cost of memory (by requiring only the smallest capacity) and maximizes the speed of learning (by exponentially reducing error rates).
Reframing Rationality: The work challenges the prevailing view that confirmation bias is inherently irrational. Instead, it posits that under conditions of incompatible information sources, seeking evidence that aligns with prior beliefs is the mathematically optimal strategy for a biological system aiming to minimize uncertainty and energy consumption.
Limitations and Scope: The authors note that this rationality holds under the assumption that information sources are genuine (signal plus noise) and not designed to be deceptive. In social contexts where sources may be manipulative, the adaptive benefits of confirmation bias could be subverted. The paper concludes that while the framework explains the mechanism of rational confirmation bias, it does not validate the bias in all societal contexts, particularly where "fake news" or deceptive signals are present.

In summary, the paper formulates confirmation bias as a rational, adaptive strategy arising from the minimization of error probability and energy consumption within a quantum probability framework, revealing that such bias leads to exponential learning efficiency and minimal memory requirements.