How Occam's razor guides human decision-making

In the messy reality of daily life, we are constantly forced to choose between competing stories to explain what we see. When a car won't start, we might blame a dead battery, a faulty alternator, or a ghost in the engine. Our brains must sift through these possibilities, often with very little information, to decide which explanation is most likely true. For centuries, philosophers and scientists have proposed a guiding rule for this process, known as Occam's razor. The principle suggests that when two explanations fit the facts equally well, the simpler one is usually the better choice. It is a mental shortcut that favors the story with fewer moving parts, fewer assumptions, and less room for error. But while this idea has long been a favorite of thinkers, it has remained unclear whether human brains actually use it in the precise, mathematical way that theory predicts, or if we simply have a vague preference for things that feel less complicated.

A team of researchers has now moved this question from the realm of philosophy into the laboratory, using a series of carefully designed experiments to measure exactly how people weigh simplicity against accuracy. They found that human decision-making does indeed follow a specific, quantifiable form of Occam's razor. When faced with noisy, uncertain data, people naturally prefer the simpler explanation, not just because it feels right, but because their brains appear to be performing a complex statistical calculation that averages over all possible versions of a story to avoid being fooled by random chance. This preference for simplicity is so deeply ingrained that it persists even when people are explicitly told to ignore it and focus only on the single best fit. In a striking contrast, computer programs trained to solve the same problems only adopted this simplicity bias when they were specifically programmed to do so, suggesting that the human tendency to favor simple explanations is a fundamental feature of our cognition, not just a learned strategy.

The researchers set out to test this by creating a visual game where participants had to act as detectives. On a computer screen, a cloud of ten red dots appeared, scattered somewhat randomly around a central point. These dots were generated by one of two hidden shapes, either a straight line or a curved arc, which served as the "source" of the data. The participants' job was to guess which shape produced the dots. The shapes were drawn in black, and the dots were red, creating a clear visual puzzle. To make the task scientifically rigorous, the researchers designed four different versions of the game. In each version, they pitted two shapes against each other in a way that isolated a specific mathematical feature of complexity. One version compared a single point to a long line, testing how people handle the number of variables. Another compared two lines of different lengths, testing the size of the possible area the shape could cover. A third version looked at how the shape was positioned relative to the data, and the fourth examined how the shape curved away from the data.

In these experiments, the researchers could calculate the mathematically perfect answer for every single trial. A perfect observer, following the strict rules of Bayesian statistics, would not just look for the shape that was closest to the dots. Instead, this ideal observer would consider every possible position the shape could have taken and average them out. This process naturally penalizes shapes that are too flexible or too large, because they spread their "explanatory power" too thin across the screen, making them less likely to be the true source of a specific cluster of dots. The researchers predicted that if humans were using this kind of statistical reasoning, their choices would shift away from the complex shapes and toward the simple ones, even when the complex shapes were technically closer to the dots.

The results confirmed that humans do exactly this. When the data was ambiguous, meaning the dots were roughly equidistant from both shapes, participants consistently chose the simpler option. But the study went further than just showing a general preference; it measured the strength of that preference with high precision. By analyzing thousands of choices from over 200 participants, the team found that people were sensitive to four distinct geometric features of the shapes: their dimensionality (how many parameters they had), their boundaries (where they ended), their volume (how much space they covered), and their robustness (how they curved). The participants' choices aligned quantitatively with the predictions of the mathematical model. They were not just vaguely preferring "simple" things; they were weighting these specific geometric factors in a way that closely matched the theoretical ideal of integrating over all possible causes.

To understand if this was a learned behavior or something more fundamental, the researchers compared the human participants to artificial neural networks, a type of computer program designed to mimic the brain. They trained these networks to solve the same dot-and-shape puzzles. When the networks were trained to find the mathematically perfect "generative" solution—which requires using the simplicity bias—they learned to do so and performed better than the humans. However, when the researchers changed the instructions and told the networks to find the "maximum likelihood" solution—meaning they should simply pick the shape closest to the dots, ignoring complexity—the networks adapted instantly. They stopped showing any preference for simplicity and performed perfectly on the new task.

The human participants, however, did not change so easily. Even when they were given the same instructions to ignore complexity and just pick the closest shape, they continued to favor the simpler options. Their brains seemed unable to turn off the mechanism that integrates over all possibilities. This suggests that the human preference for simplicity is not a flexible strategy we adopt when we are told to be smart, but a core part of how we process information. We seem to naturally average over the many ways a story could be true, which protects us from overfitting to noise, whereas machines only do this if we explicitly build that capability into them.

The study also revealed that while humans are remarkably good at this, they are not perfect. There was a wide range of individual differences in how strongly people applied these simplicity rules. Some participants were very close to the mathematical ideal, while others were less sensitive to certain features, like the size of the shape or its curvature. Interestingly, those who were further from the ideal tended to make more mistakes, suggesting that tuning this internal "simplicity filter" is crucial for making accurate decisions in a noisy world. The researchers noted that people seemed to weigh the different features differently, perhaps because some are easier to compute than others. For instance, counting the number of variables in a shape might be simpler for the brain than calculating the total volume of space it covers, leading to a natural imbalance in how we apply the rule.

Ultimately, this work provides a rare, quantitative glimpse into the hidden machinery of human thought. It shows that our brains are not just pattern-matching machines that look for the closest fit. Instead, they act like sophisticated statisticians that constantly ask, "How many other ways could this have happened?" and use the answer to guide our choices. By integrating over all possible explanations, we avoid the trap of believing a complex, flexible story just because it happens to fit the current data by chance. This study confirms that Occam's razor is not just a philosophical suggestion for how we should think, but a description of how we actually do think, a built-in feature of human cognition that helps us navigate a world full of uncertainty and noise.