Most guests on Unscripted build AI or deploy it. Augusto Gonzalez studies a stranger question: can a large language model actually model a human being, and if so, which humans? His own research produced an uncomfortable answer.
The short version: AI can approximate how a population might respond, at least directionally, but today's models quietly flatten humanity's moral and cultural diversity toward a Western, wealthy, educated default. Making the models bigger does not fix this, because the bias is baked into the data available on the web. The useful question is not whether AI represents humans, but which humans it represents, and who is left out.
Who is Augusto Gonzalez?
Augusto Gonzalez is an economist and researcher from Argentina who uses large language models to simulate whole human populations. He builds what he calls synthetic cultural agents, AI stand-ins that let researchers run classic economic experiments on groups we have rarely been able to study at scale. He started his PhD the same week ChatGPT launched, when many serious people were dismissing these models as stochastic parrots, and bet his research on them anyway.
What are synthetic cultural agents, and why would anyone want them?
A synthetic cultural agent is an AI configured to approximate how a specific human population might think or behave when presented with a scenario. The value is in the piloting stage.
Businesses already run focus groups and pilot studies to calibrate a campaign, a product, or a policy, and even those small studies are expensive. Studying a small-scale society with no profit incentive attached is harder still. Augusto's argument is that you can use these agents to cheaply pilot a question, get a directional read, and decide where to invest in the expensive, rigorous human research. The agent is a tool for calibrating your assumptions, not a replacement for studying real people.
What did the Hadza experiment show?
In one project, Augusto's team simulated the Hadza, a small tribe in Tanzania, and ran economic experiments on the synthetic version. The AI matched the real anthropology from the one existing field study on the subject. The method was relatively simple, a retrieval system with custom search tools that pulled the ethnographic information the model needed.
The result was both amazing and unsettling. It worked, but at first they did not fully understand why, and reviewers reasonably worried the model was just regurgitating a paper from its training data. So they ran robustness checks. When they tried the base model without the methodology, the results were completely off, which told them the method was doing real work to steer the model toward the target population. That finding kicked off a broader research agenda on debiasing these models away from their default assumptions.
Why won't bigger models fix cultural bias?
Because the bias comes from the training data, not the model size. Augusto's research showed that models are good at representing the average person in wealthy, educated, industrialized societies, precisely because those populations dominate the text available on the web. They are poor at representing everyone else.
Scaling a model trained on the same skewed data does not solve the problem. Augusto points to the large labs now spending heavily to collect what they call more inclusive or representative data, which in his view proves the point: what is freely available online is not enough, because not everyone in the world is online in the first place. If you want a model to represent India or Argentina, you cannot simply prompt it and hope. You have to get your hands dirty and build a methodology.
Which humans are actually inside the model?
This is the question Augusto wishes more people were asking. Everyone worries about whether AI can represent humans. Almost no one asks which humans. There is no single type of human. The models capture a real but narrow slice of human nature, and, he notes, economics has quietly had the same problem for decades, validating theories of human behavior on data drawn from that same narrow slice.
What is the real danger: exporting values or exporting prejudice?
Augusto drew a distinction worth holding onto. The larger risk is not exporting values, it is exporting subtle prejudices dressed up as data. The blatant biases can be corrected with algorithms. The subtle ones are harder. Statements like "all Argentines drink mate" sound harmless, but they encode a false confidence that a single message or policy will fit everyone.
He gave a personal example. When he asked a mainstream model what to expect from Argentina, it described warm people who drink mate, eat asado on Sundays, and greet with a kiss on the cheek. A reasonable average, he said, but his own family does not drink mate, plenty of his friends do not greet that way, and decades of corruption have eroded trust in ways the tidy summary omits. The deeper mistake is assuming intent behind an action: seeing a cultural tradition and reading it as a personality trait. That becomes dangerous in areas like mental health, where advice that fits someone in one country may be wrong for someone in another. As he put it, most prejudice stems from the wrong belief that everyone else operates in the same framework you do.
Should AI be used to make policy decisions?
Augusto's answer was direct: not right now, and maybe not ever. He grounded it in economic history. The Washington Consensus promoted free markets that worked in some countries and failed in Argentina, largely because the designers did not put enough effort into understanding the people receiving the policy. He connected this to the Lucas critique, the idea that once you implement a policy, people adapt their expectations and the policy stops behaving as predicted. Using AI to skip the work of understanding a population repeats the same mistake. The responsible use is to run the cheap pilot, calibrate your priors, and then do the hard human work anyway.
What is the fix?
Augusto's bet is open source. He does not believe a single lab can or will represent every community, because the profit incentive is not there. His hope is that enough shared architectures and data pipelines will let each community build AI to represent itself, rather than waiting to be represented by someone else. That, in his view, is the only path that gives underrepresented populations a real voice in the systems that will shape the coming decades.
Frequently asked questions
Can AI accurately simulate human populations? It can approximate responses directionally, which is useful for piloting research cheaply. It is not accurate enough to replace real study of people, and it is least reliable for populations underrepresented in training data.
Why are AI models biased toward Western perspectives? Because most of the text used to train them comes from wealthy, educated, industrialized, English-language sources. The models reflect that data, so they represent those populations well and others poorly.
Does making AI models larger reduce cultural bias? No. The bias originates in the training data, not model size. Scaling a model trained on skewed data reproduces the same skew, which is why labs are now investing in more representative data collection.
Should AI be used for policy decisions in non-Western populations? Augusto argues no, at least not yet and possibly not ever. AI can help run a cheap pilot to calibrate assumptions, but real understanding of the affected population still requires human research.
What is the proposed solution to AI cultural bias? Open source. Shared architectures and data pipelines that let individual communities build AI to represent themselves, rather than relying on a single lab that lacks the incentive to represent everyone.
Listen to the full conversation with Augusto Gonzalez, and catch every episode, on the Unscripted with Jeff Pedowitz page.