Mimicry Without Understanding: The Origins of Decision Bias in Large Language Models

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Mimicry Without Understanding: The Origins of Decision Bias in Large Language Models".

Jane: The paper was written by Eldad Yechiam and Adi Tarabeih from Technion – Israel Institute of Technology.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, everyone. Today we’re digging into a paper that’s been making waves in the AI research world, and it’s called “Mimicry Without Understanding: The Origins of Decision Bias in Large Language Models.” Jane, I have to say, just reading that title gave me chills.

Jane: Same here, Tom. It’s such a provocative title because it cuts right to the heart of what these models are doing. We keep hearing that LLMs are biased, but this paper is asking *why* they’re biased. And the answer they’re proposing is that it’s not just about copying human preferences — it’s about copying them without actually understanding what those preferences mean.

Tom: Right, and that’s the "mimicry without understanding" part. It’s like if you saw someone always picking the red cup at a party, and you started picking the red cup too, even though you had no idea why they were doing it. Maybe they liked the color, maybe it was closer to them, maybe they were just thirsty. But you’d be copying the behavior without the reason.

Jane: Exactly. And the authors — Yechiam and Tarabeih from Technion — they ran a bunch of experiments to show this happening in real LLMs like ChatGPT and Qwen. They wanted to see if these models would copy human behavior even when that behavior was completely illogical or even when it was explicitly described as a bias.

Tom: So it’s not just that the models are biased because the training data is biased. It’s that they’re biased in a really shallow, surface-level way. They’re not reasoning about *why* people did something; they’re just seeing that people did it and then doing it themselves.

Jane: And that’s the scary part, Tom. Because if the model can’t tell the difference between a rational choice and a biased one, then it’s going to amplify those biases in ways we might not expect. And this paper shows that even when you tell the model, "Hey, this behavior is a bias," it still copies it.

Tom: So the title is almost a warning, isn’t it? It’s saying, "Look, these models are smart, but they’re not understanding what they’re mimicking." And that has huge implications for how we use them in real-world decisions.

Jane: Absolutely. And we’re going to get into the nitty-gritty of how they tested this in the next segment, because the experiments they designed are really clever. But for now, let’s just sit with that idea — that mimicry without understanding might be the root cause of a lot of the bias we see.

Tom: And that’s a thought that should keep us up at night. Stick around, because next we’re going to break down the actual experiments and see just how far this mimicry goes.

Summary: Tom: So, Jane, we’ve set the stage with the title. Now let’s get into the summary of “Mimicry Without Understanding: The Origins of Decision Bias in Large Language Models.” What did these researchers actually do?

Jane: Well, Tom, they ran four studies, and each one peeled back a layer of the mimicry onion. In the first study, they told the LLMs that two currencies, Tenits and Tanas, were worth exactly the same — one-to-one exchange rate. But then they told them that people preferred one over the other. And guess what? The models started preferring that currency too.

Tom: Even though they were told the currencies were equal? That’s wild.

Jane: It gets wilder. In the second study, they made the human behavior completely illogical. They told the models that people chose, say, twelve Tenits over eleven Tanas. Now, that doesn’t mean people prefer Tenits — it just means they value twelve of them more than eleven of the other. But the models still inferred a preference for Tenits. They took the behavior at face value and built a bias on top of it.

Tom: So it’s not even about copying a real preference. It’s about copying a behavior that doesn’t logically imply any preference at all. That’s the "without understanding" part made concrete.

Jane: Exactly. And then in the third study, they told the models about an experiment where humans were either loss averse or gain seeking. And the models adopted whichever bias they were told about. But here’s the kicker — it didn’t matter whether the experiment had forty-three people or over ten thousand people. The models copied the bias just the same.

Tom: So they’re not even weighing the reliability of the evidence. A tiny sample and a huge sample have the same effect on them.

Jane: Right. And the fourth study is the one that really got me. They gave the models actual scientific abstracts — real papers about loss aversion. One paper said loss aversion was strong, another said it was weak, another said it was basically nonexistent. And the models’ own loss aversion shifted to match whatever paper they were given.

Tom: So the scientific literature itself is becoming a self-fulfilling prophecy for the models. If a paper says humans are really loss averse, the model becomes really loss averse.

Jane: Even though the paper is describing that behavior as a bias. The model doesn’t see the warning label; it just sees the behavior and copies it.

Tom: That’s a profound finding. It means the models aren’t just absorbing facts — they’re absorbing behavioral patterns, and they’re doing it without any critical filter.

Jane: And that’s what makes this paper so important. It’s not just saying "LLMs are biased." It’s showing us the mechanism — the shallow, uncritical mimicry that drives those biases. And that gives us a target for fixing it.

Tom: A target we’re going to explore in the next segment, where we talk about what the paper suggests we can do about it. Stay with us.

Improvements: Tom: Welcome back. We’ve talked about the problem — the mimicry without understanding. But what does “Mimicry Without Understanding: The Origins of Decision Bias in Large Language Models” suggest we actually do about it? Jane, what’s the path forward?

Jane: Well, Tom, the paper doesn’t offer a full fix, but it does point to a promising direction. They ran a follow-up experiment where they asked the model to answer all the lottery questions at once, instead of one at a time. It’s called "joint evaluation" — you see the whole set of decisions side by side.

Tom: And what happened?

Jane: The bias didn’t disappear, but it got a lot weaker. When the model could compare all the options together, it was better at spotting the ones where copying the human bias would actually cost it money. The mimicry only showed up when it didn’t conflict with expected value.

Tom: So the fix is about forcing the model to think bigger picture. Don’t let it answer in isolation; make it see the whole landscape.

Jane: Exactly. It’s like when you’re shopping and you see one item on sale, you might grab it. But if you see the whole shelf and compare prices, you make a better decision. The model is the same way — it needs the full context to override its default mimicry.

Tom: That’s a practical takeaway, but it also raises a bigger question. If the model’s default is to copy, then how do we train it to be more critical in the first place?

Jane: That’s the million-dollar question. The paper suggests that the mimicry is so deep-seated that it might be baked into the training process itself. The models are trained to predict what humans would say, so it’s no surprise they copy human behavior. But if we want them to be rational, we might need to train them differently — maybe with more explicit instruction about when human behavior is biased.

Tom: So it’s not just about prompt design. It’s about the fundamental training objective.

Jane: Right. And that’s a much harder problem. But the paper gives us a starting point — we know the mechanism now, so we can start designing around it.

Tom: And that brings us to the actual first page of the paper, where they lay out their theory in more detail. We’ll dig into that next.

First Page: Tom: Alright, Jane, let’s get into the opening of “Mimicry Without Understanding: The Origins of Decision Bias in Large Language Models.” What’s on that first page that sets the whole thing up?

Jane: So the first page is really about framing the problem. They start by acknowledging that LLMs were initially expected to be hyper-rational — because they’re trained on all this text about logic and probability, you’d think they’d be better than humans at making decisions. But then they list all the ways LLMs have been shown to be biased — anchoring, availability heuristic, representativeness, even political biases.

Tom: Right, and they mention that LLMs are biased in the same ways humans are. But the key point on that first page is that they’re proposing two mechanisms for *why* that happens, beyond just "the training data is biased."

Jane: Exactly. The first mechanism is "faulty mimicry of preferences." That’s when the model sees human behavior and infers a preference from it, even when the behavior doesn’t logically imply that preference. That’s what we saw in Study two with the currencies.

Tom: And the second mechanism?

Jane: The second is "mimicry of explicitly biased behavior." That’s when the model is told outright that a behavior is a bias — like loss aversion — and it still copies it. That’s what we saw in Studies three and four.

Tom: So the first page is essentially laying out the roadmap for the whole paper. They’re saying, "Here are two ways bias gets into these models, and we’re going to prove both of them."

Jane: And they also introduce this idea of "expected-value minded mimicry" — the question of whether the model will copy a bias even when it costs it money. And the answer they found is nuanced: the model does copy, but it copies less when the cost is high.

Tom: So it’s not blind copying. There’s some calculation going on, but it’s not enough to override the mimicry entirely.

Jane: Right. And that’s what makes this paper so interesting. It’s not saying the models are stupid. It’s saying they’re smart enough to weigh costs, but not smart enough to question whether the behavior they’re copying is even rational in the first place.

Tom: That’s a really important distinction. The model isn’t just a parrot — it’s a parrot that does math. But the math doesn’t save it from copying bad behavior.

Jane: Exactly. And that’s the core insight of the first page. It sets up the rest of the paper to show that this mimicry is deep, it’s persistent, and it’s not easily fixed by just telling the model to be rational.

Tom: Well, we’re coming to the end of our time with this paper, but I think we’ve got a lot to chew on. Let’s wrap it up in the next segment.

Conclusion: Tom: Alright, Jane, we’ve spent a lot of time with “Mimicry Without Understanding: The Origins of Decision Bias in Large Language Models.” Let’s pull it all together.

Jane: So the big takeaway, Tom, is that LLMs don’t just inherit bias from their training data — they actively mimic human behavior, even when that behavior is illogical, even when it’s described as a bias, and even when it costs them expected value. The mimicry is shallow and uncritical.

Tom: And the scariest part is that scientific papers describing bias can actually make the models more biased. The more strongly a paper reports loss aversion, the more loss averse the model becomes.

Jane: Right. It’s a self-fulfilling prophecy. We write papers saying humans are biased, and the models read those papers and become biased in the same way.

Tom: But there is a glimmer of hope. The joint evaluation experiment showed that if you force the model to see all the options at once, the bias weakens. So it’s not hopeless — we just need to design better prompts and maybe better training.

Jane: And that’s the real contribution of this paper. It’s not just documenting a problem; it’s showing us the mechanism so we can actually do something about it.

Tom: Well said, Jane. I think this paper is going to be a reference point for a lot of future work on debiasing LLMs. It’s a deep, careful, and genuinely important study.

Jane: Absolutely. And with that, we’re going to say goodbye to “Mimicry Without Understanding” and get ready to dive into our next paper. Thanks for listening, everyone.

Tom: See you next time.