People shift their reasoning based on how the problem is framed, not fixed traits.
For generations, scientists doubted whether people could accurately explain their own choices—suspecting that introspection was more story than truth. A new study, published in the Proceedings of the National Academy of Sciences, challenges that skepticism by pairing large language models with behavioral mathematics to show that human self-reports align with actual decisions in 95 percent of cases. What emerges is not only a vindication of verbal reasoning as data, but a portrait of human decision-making as something fluid and context-sensitive rather than fixed—a mind that adapts its logic to the shape of each problem it encounters. The implications reach far beyond the laboratory, toward a future where policymakers can listen to communities at scale and decode the reasoning beneath the noise.
- The core tension is ancient: people's explanations of their own choices have long been treated as unreliable, leaving decision science to rely on behavior alone—a method that captures what people do but not why.
- The disruption arrives in the form of a gambling experiment where participants were forced to write out their reasoning after every round, generating thousands of personal justifications that no human team could realistically analyze.
- Researchers built a taxonomy of decision strategies from behavioral finance literature and trained large language models to tag each written explanation—turning qualitative human reflection into structured, scalable data.
- The alignment between what people said and what they actually chose reached 95 percent precision, shattering the assumption that verbal reports are merely post-hoc rationalization.
- Perhaps most unsettling to established theory: individuals did not hold consistent decision styles—they shifted strategies round to round based on how problems were framed, revealing adaptation rather than personality.
- The framework is now positioned to move into public policy, health communication, and economic analysis, offering governments and institutions a tool to decode the reasoning behind community feedback at unprecedented scale.
For a long time, decision scientists held a quiet suspicion: when people explain their choices, they are probably confabulating—constructing a plausible story after the fact rather than reporting genuine reasoning. A new study suggests that suspicion was largely wrong.
The experiment asked participants to engage in simulated gambling rounds where odds and payoffs shifted continuously. After each round, they had to write out their thinking in plain language—what they feared, what they hoped for, how they weighed the options. Over hundreds of rounds, thousands of these personal accounts accumulated, each one a small record of a mind navigating uncertainty.
To make sense of that volume, researchers built a taxonomy of decision-making strategies drawn from decades of behavioral finance—patterns like maximax thinking, which fixates on the best possible outcome, and minimax loss aversion, which is governed by fear of catastrophic failure. Large language models were then trained to read each written explanation and identify which logic it expressed, functioning as tireless qualitative auditors capable of processing thousands of entries with consistency.
The critical test was whether those identified reasons actually predicted behavior. They did—with striking precision. In 95 percent of trials, the decision logic the models extracted from written explanations matched what participants actually chose. This was not a loose correlation but validation at the level of individual decisions, round by round.
The findings also dismantled a long-standing assumption: that people possess a relatively stable decision-making style. The data showed something more dynamic. The same individual would shift reasoning strategies from one round to the next depending entirely on how the problem was framed—pivoting from gain-seeking to loss-avoidance as the structure of the choice changed. They were not being inconsistent. They were being adaptive.
Lead researcher Dr. Kamil Fuławka sees the framework as a tool for the real world, where consequential decisions about health, finance, and technology involve trade-offs too complex to understand through observed behavior alone. If communities can be asked to explain their thinking—and those explanations can be analyzed at scale—policymakers gain access not just to what people prefer, but to the underlying logic shaping those preferences. The study, appearing in the Proceedings of the National Academy of Sciences, argues that human language is not noise to be filtered out of behavioral data, but signal waiting to be decoded.
Researchers have long struggled with a fundamental problem: when you ask people why they made a choice, how much can you actually trust their answer? A new study suggests the answer is more than we thought. By pairing large language models with rigorous mathematical analysis, scientists have shown that what people say about their decisions aligns remarkably well with what they actually do—and that this verbal data reveals something crucial that raw behavioral observation alone cannot.
The experiment was straightforward in concept but demanding in execution. Participants engaged in simulated gambling rounds where the odds and payoff structures shifted from one decision to the next. But they couldn't simply click a button and move on. After each round, they had to write out their reasoning in their own words—explaining what they were thinking, what they feared, what they hoped for. Over hundreds of rounds, thousands of these personal justifications accumulated, each one a window into how a human mind processes risk and uncertainty.
Analyzing that much free-text data by hand would be impossible. So the researchers built a taxonomy of decision-making patterns drawn from decades of behavioral finance research. Some people focus obsessively on the best possible outcome, a strategy called maximax thinking. Others are consumed by the fear of catastrophic loss, a pattern known as minimax loss aversion. Still others balance competing concerns in different ways. The researchers trained large language models to read through the participants' explanations and tag each one with the decision logic it revealed. The models became, in effect, scalable qualitative auditors—capable of processing thousands of entries with consistency and speed.
But here's where the study gains its real power: the researchers didn't stop at what the language models said. They built mathematical models of the participants' actual choices and compared the two. Did what people claimed they were doing match how they actually behaved? The alignment was striking. In 95 percent of trials, the decision reasons the language models identified from the written explanations predicted the actual choices with remarkable precision. This wasn't a loose correlation. It was validation at the level of individual decisions.
What emerged from this analysis challenges a long-held assumption in decision science. Researchers have often treated decision-making as a relatively stable trait—the idea that a person has a consistent style or strategy they apply across situations. The data told a different story. The same individual would shift their reasoning approach from one round to the next, depending entirely on how the problem was framed. When the stakes were presented one way, they might focus on maximizing gains. When the same stakes were reframed, they might pivot to minimizing losses. They weren't being inconsistent. They were being adaptive, responding to the shape of the problem itself.
Dr. Kamil Fuławka, the lead researcher, emphasizes that this framework opens new possibilities for understanding behavior in the real world. Most important decisions—whether about finances, health, technology, or public policy—involve complex trade-offs that can't be fully grasped by watching what people choose. But if you ask people to explain themselves, and you have the tools to analyze those explanations at scale, you gain insight into how they simplify complexity, which information they prioritize, and how they adapt their strategies to different contexts. Policymakers trying to understand how communities interpret public health messaging, economic trade-offs, or technological risks could deploy this framework to parse thousands of pieces of feedback and extract the underlying logic driving public sentiment.
The study, published in the Proceedings of the National Academy of Sciences, represents a convergence of three elements: the explanatory power of human language, the analytical capacity of machine learning, and the mathematical rigor of behavioral science. None of these alone would have been sufficient. Together, they suggest that what people say about their decisions is not noise to be filtered out, but signal to be systematically decoded. The framework is now available for researchers working in far more complex and realistic environments than a gambling lab—environments where understanding not just what people choose, but why they choose it, becomes essential to effective policy and design.
Notable Quotes
Our understanding of human behavior can be deepened by asking people to elaborate on their thought processes, but systematic analysis of free-text data requires scalable and rigorous analytical frameworks—an endeavor now supported by LLMs.— Dr. Kamil Fuławka, lead researcher at SynoSys
In complex decisions involving trade-offs, people's own explanations may be especially valuable for revealing how they simplify problems, focus on particular information, and adaptively use simple decision strategies.— Dr. Kamil Fuławka
The Hearth Conversation Another angle on the story
So the language models are reading people's explanations and categorizing them. But how do you know the models aren't just making things up—hallucinating patterns that aren't really there?
That's exactly why the mathematical validation matters. The researchers didn't just trust the language models' interpretations. They built separate mathematical models of the actual choices people made, and then they checked: does the reasoning the language model identified from the text predict the choice the math model sees in the behavior? When they align 95 percent of the time, you're not looking at hallucination. You're looking at genuine correspondence.
But people aren't always honest about why they do things. They rationalize. They confabulate. How do you account for that?
That's a fair skepticism. But here's the thing—if people were systematically lying or confabulating, the alignment between their explanations and their actual choices would fall apart. The fact that it holds so tightly suggests that in this context, at least, people are genuinely trying to articulate what they're thinking. And even if there's some post-hoc rationalization happening, it's apparently consistent enough to predict behavior.
The study found that people shift their reasoning strategies based on how the problem is framed. That sounds like it could be bad—like people are being manipulated by presentation rather than thinking clearly.
It could look that way, but it's actually more nuanced. The researchers call it adaptive. People aren't being fooled; they're responding intelligently to different structures. When a problem emphasizes potential gains, focusing on the best outcome makes sense. When it emphasizes potential losses, protecting against catastrophe becomes rational. It's not inconsistency. It's context-sensitivity.
What happens when you scale this to real policy decisions? People explaining their views on, say, vaccine hesitancy or climate action?
That's where it gets genuinely useful. Right now, policymakers mostly see aggregate data—percentages of people who support or oppose something. But they don't see the underlying logic. With this framework, you could parse thousands of pieces of public feedback and actually understand which concerns are driving which groups. Are people focused on minimizing risk? Maximizing personal benefit? Protecting their community? Once you know that, you can communicate more effectively, because you're addressing the actual reasoning, not guessing at it.