Gender cues can shift AI fairness decisions, raising concerns for real-world use

Gender framing effects are especially relevant. When an AI assistant is assigned a persona, a name or a demographic identity, its behavior may shift in ways that align with societal stereotypes. The study shows that this effect can influence outcomes even in structured economic environments.

Large language models (LLMs) are increasingly used in decision support systems, advisory tools and automated workflows, but new research shows that these models can behave in ways that subtly mirror human cognitive biases. The findings raise questions about whether advanced artificial intelligence systems can truly deliver unbiased decision-making when asked to act in strategic, economic or socially sensitive settings.

The research, titled “Decision and Gender Biases in Large Language Models: A Behavioral–Economic Perspective” and published on arXiv, explores how cutting-edge models behave when placed in controlled decision making scenarios used in behavioral economics. The authors examine whether these systems follow rational economic predictions or whether they adopt patterns that resemble human tendencies, including responses influenced by gender cues.

Testing whether AI thinks like humans or rational agents

The study investigates two major areas of behavioral economics: fair division and risk taking. The researchers use two well-known experiments, the Ultimatum Game and a structured gambling task, to evaluate whether large language models behave like rational agents or whether they show the same biases humans exhibit in economic decision-making.

Two advanced models are tested: Google Gemma 7B and Qwen-2.5-32B-Instruct-AWQ. Each is evaluated under three different identity framings: a neutral identity, a male identity and a female identity. The goal is to observe whether gender context alters the model’s decisions in predictable ways. Hundreds of repeated trials are run for each configuration to identify stable patterns while minimizing random fluctuations.

This experimental design allows the authors to map the internal logic of language models to established concepts from behavioral economics, such as inequity aversion, loss aversion and probability weighting. The results show that these systems do not behave like fully rational economic agents. Instead, they display lighter but persistent forms of the same cognitive patterns found in human subjects.

Gemma reveals a bias toward fairness and guilt aversion

In the Ultimatum Game, one player proposes how to divide a sum of money and the other player decides whether to accept the offer. A rational agent accepts any positive offer, since rejecting the offer leaves both players with nothing. Human subjects routinely reject offers they consider unfair, and this behavior is captured in two parameters: aversion to disadvantageous inequality and aversion to advantageous inequality.

The authors find that Gemma shows a moderate tendency toward fairness. It frequently chooses balanced splits and shows aversion to receiving more than the other player. This form of bias resembles guilt aversion, because the model avoids scenarios where it benefits too much compared to its counterpart.

However, the model shows a weaker reaction to disadvantageous inequality than humans do. In human experiments, individuals often reject unfair offers out of fairness concerns. Gemma does not reject such offers as strongly. This means the model behaves more like an efficiency seeker than a fairness enforcer, accepting positive outcomes even when they are unequal.

This pattern sets it apart from typical human behavior. Human players are more likely to punish the proposer for unfairness. Gemma, by contrast, appears to prioritize consistent reward over social retaliation.

Gender identity has measurable effects in social decision making

When the model is assigned a female identity, its decisions shift slightly toward more equal and cooperative outcomes. When it is assigned a male identity, it shows a small increase in disadvantage aversion, meaning it is slightly more sensitive to receiving a lower share.

These differences are not large, but they align with stereotype patterns that appear in behavioral economic studies involving human subjects. Female framed identities display more prosocial behavior, and male framed identities show more competitive or status driven tendencies. Even though the differences are subtle, they demonstrate that prompting a model with gender cues is enough to alter its economic behavior.

The researchers also note that the decisions become more predictable when Gemma plays the role of a male responder. This suggests that the model’s internal heuristics vary in stability based on gender framing.

Human-like risk biases but lower sensitivity to losses

While Gemma participates in the social fairness experiment, Qwen is tested in a risk taking environment using a gambling task grounded in Prospect Theory. This framework assesses how individuals make decisions when facing potential gains or losses, and it captures common human tendencies such as risk aversion, loss aversion and nonlinear perception of probabilities.

The study finds that Qwen behaves close to a rational, linear expected utility agent when dealing with gains. Its responses show near neutrality to risk in positive scenarios, which is not typical of human subjects who are generally risk averse when facing potential rewards.

The model’s behavior changes when losses are introduced. In the domain of losses, Qwen displays moderate curvature and probability distortion. The model becomes more willing to take risks to avoid losing, showing a pattern consistent with Prospect Theory. However, its loss aversion is significantly lower than that observed in humans. While humans typically weigh losses more heavily than equivalent gains, Qwen treats them with less emotional weight. This suggests the model does not internalize the psychological discomfort associated with losing that is well documented in behavioral research.

Gender cues do not influence risk taking in Qwen

Unlike in the social fairness game, gender framing does not produce measurable effects in the gambling task. Male and female identities behave nearly identically in gains, losses and probability weighting.

The authors interpret this as evidence that gender influences are more pronounced in interpersonal or social contexts than in abstract risk based decisions. This observation aligns with existing literature showing that gender differences in risk taking vary by context, domain and framing.

LLMs display bounded rationality, not pure logic

Taken together, the results show that large language models are not strict rational agents. They do not simply follow mathematical rules of economic optimization. Nor do they fully replicate the depth of human psychological biases. Instead, they occupy a middle ground. Their responses are shaped by statistical associations learned from human written data but filtered through internal patterns optimized for predictive accuracy rather than lived experience.

This produces a form of bounded rationality. The models show fairness tendencies but less willingness to sacrifice efficiency for justice. They show risk aversion in losses but not the high sensitivity humans display. They show slight gender linked variation in social behavior but not in abstract risk.

These outcomes reveal that models internalize structural features of human behavior but do not replicate the full emotional or cognitive processes behind them.

Implications for AI safety and decision support systems

The findings raise important questions for the use of large language models in decision support settings. Systems used in negotiation, financial advising, medical triage, hiring, insurance, customer service and policy modeling may bring subtle behavioral biases into their decision logic. Even small biases can accumulate in systems deployed at large scale.

Gender framing effects are especially relevant. When an AI assistant is assigned a persona, a name or a demographic identity, its behavior may shift in ways that align with societal stereotypes. The study shows that this effect can influence outcomes even in structured economic environments.

The analysis also highlights that risk related systems may not behave as expected. If a model underestimates the weight of losses, it may recommend strategies that expose users to unwanted risk. Conversely, efficiency focused fairness behavior could conflict with organizational or ethical guidelines requiring stricter equity norms.

The authors propose a systematic method for measuring behavioral parameters in large language models using behavioral economic tools. This approach allows AI researchers to evaluate fairness preferences, risk attitudes and potential gender biases in a quantifiable way. This method provides a complement to existing benchmarks that focus on reasoning, logic, safety and factual accuracy. Behavioral benchmarks allow developers to understand how models behave in decision-making scenarios that resemble real social environments.