Can AI Really Mimic Human Behavior? What the Ultimatum Game Reveals for Market Researchers

What Happens When AI Has to Be “Fair”?

Can AI be trusted to simulate real human judgment? Not just automate decisions or crunch numbers, but actually model how people feel, react, and respond, especially when fairness is at stake?

That’s the challenge posed by the Ultimatum Game, a behavioral economics classic where human behavior often defies pure logic. It’s a game of money, morals, and social expectation, and recently it became the testing ground for an intriguing experiment: What if a large language model (LLM) played both roles in the game?

In a 2023 study by Aher, Arriaga, and Kalai* researchers used large language models to simulate thousands of Ultimatum Game interactions. (*Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies, 2023)

The results may change how market researchers think about AI. Not as a threat to methodological rigor, but as a tool that, when thoughtfully applied, can reflect the psychological patterns we’ve spent decades studying.

Let’s take a closer look at the experiment and explore its potential to shape the future of research.

The Ultimatum Game and Why Researchers Use It

The Ultimatum Game is straightforward. One player, the proposer, is given $10 and must offer a portion of it to another player, the responder. If the responder accepts the offer, both parties walk away with the proposed amounts. If the responder rejects it, neither gets anything.

In theory, any non-zero offer should be accepted. But that’s not how people behave. Offers that feel unfair, like $2 out of $10, are often rejected, even if it means losing out. The desire to punish perceived unfairness outweighs the incentive to gain something rather than nothing.

This is why the Ultimatum Game is a favorite among behavioral researchers. It surfaces real human sensitivities such as fairness, emotion, and cultural norms. It highlights the complexity behind decision-making and offers a reliable benchmark for studying how people navigate social trade-offs.

How AI Simulated Human Offers and Responses

In the Aher et al. study, a LLM was asked to play both sides of the Ultimatum Game, proposer and responder, across a wide range of trials. Rather than following strict instructions or optimizing for economic logic, the model was prompted to behave like a typical human participant.

To reflect human variability, the researchers varied the prompts to simulate different demographic backgrounds, cultural perspectives, and personality traits. The idea was to explore not just a generic average human response, but a more diverse range of behaviors rooted in social context.

The results were unmistakable. The AI made offers that looked a lot like human ones, and it responded to those offers with familiar patterns of fairness sensitivity.

What the Results Reveal About Fairness, Bias, and Behavioral Nuance

The model’s behavior echoed what researchers have observed in real-life experiments.

Offers tended to fall in the 40-50% range, consistent with what human proposers typically offer to avoid rejection. Lowball offers, like $1 or $2, were often rejected. The model’s responders appeared to have a fairness threshold, showing the same inclination to sacrifice money in order to penalize perceived greed.

What’s more, the simulated demographic context made a difference. When the model took on different identities, such as younger or older participants, or individuals from different cultural backgrounds, its responses shifted in predictable ways. It captured not just general trends but meaningful variation based on social norms and lived experience.

These weren’t scripted responses. They reflected behavioral nuance, suggesting that LLMs, when prompted carefully, can internalize and reproduce subtle psychological dynamics.

Simulating Real Market Research Contexts

While the Ultimatum Game is a powerful academic benchmark, its implications go far beyond game theory.

At The Jenny Project, we’ve used AI personas to test how consumers respond to brand messaging, pricing, and positioning. Often, what gets revealed isn’t just preference, but perception—whether something feels fair, authentic, or aligned with expectations. That’s the kind of feedback that mirrors what we saw in the Ultimatum Game: people reacting not only to what’s offered, but how it’s framed and whether it resonates with their values.

This is where AI simulation becomes especially valuable. You can surface human-like resistance to messaging, framing, or tone before launching in the real world. That gives you early signals that would otherwise take weeks or months to uncover.

If you come from a background rooted in traditional research, it’s natural to approach AI with caution. Can a model trained on text really stand in for a live participant? In one sense, no. It can’t replicate the full richness of lived human experience. But in another, more targeted sense, it can simulate behavioral patterns that are surprisingly aligned with what we see in real-world studies.

While synthetic personas are not substitutes for lived experience, they offer valuable directional insights. Especially when time or cost prevents large sample testing.

AI isn't here to replace surveys or qualitative interviews. It complements them. Just like a wind tunnel simulates airflow before a real flight, AI simulates human reactions before a real launch.

It’s not about skipping the rigor. It’s about expanding your toolkit.

Bringing Behavioral Simulation Into Practice

This kind of behavioral simulation is what The Jenny Project was built for.

Jenny is an AI insights agent designed to help researchers explore human behavior through realistic, demographically diverse simulations. Rather than predicting outcomes from a black box, Jenny generates responses based on richly constructed personas informed by behavioral science and cultural context.

She lets you test ideas quickly, surface likely objections, and explore how different people might respond to the same offer or message. And she’s transparent. You can see how the personas are built, how responses are generated, and what assumptions are in play. That makes Jenny not just a faster way to get feedback, but a more explainable one.

Curious how this works in practice? See how Jenny replicates real-world behavior and helps you test faster, smarter. Book a demo today.