Ideological Bias in LLMs' Economic Causal Reasoning

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

Tom: Today's paper: "Ideological Bias in LLMs' Economic Causal Reasoning".

Jane: The gist: LLMs exhibit systematic ideological bias when reasoning about economic causal effects, consistently performing worse on ideologically contested cases and showing directional errors that lean toward intervention-oriented expectations.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we’re looking at this paper now, "Ideological Bias in LLMs' Economic Causal Reasoning." It’s written by Donggyu Lee, Hyeok Yun, Jungwon Kim, Junsik Min, and Sungwon Park.

Jane: It sounds like the big question they are asking is whether these large language models have a built-in bias when they try to figure out economic cause and effect relationships.

Tom: Exactly. They’re testing this by taking an existing benchmark called EconCausal—which is used for causal reasoning—and adding some specific cases where one side of the debate predicts a positive effect and the other predicts a negative one.

Lu: So they’re looking at these ideology-contested instances, which they define as situations where intervention-oriented perspectives and market-oriented perspectives predict different causal signs, you know?

Jane: Right. They found that these contested items are consistently harder for the models to get right compared to the non-contested ones.

Tom: That’s a key finding. They showed that across all twenty models they tested, accuracy on those ideology-contested items was lower than on the rest of the data.

Meng: But what does that mean when you look at how these models are built? Are we talking about this bias being something that just shows up in the final output, or is it baked into how they were trained?

Lalam: The paper suggests this isn't just a prompting artifact; it points toward potential origins earlier in the training pipeline, possibly even in pretraining data and subsequent alignment processes.

Tom: I like that. And they showed a comparison between base models and instruction-tuned models, and that there was already an intervention-truth advantage present at the base model stage before instruction tuning even happened.

Jane: So it’s not just something the user prompts them into; it seems to be a directionally biased tendency in how they learn to reason about economics.

Lu: They also looked at where this distortion shows up, and it’s not a uniform shift, like everything just leaning left or right.

Tom: It varies by subfield. The biggest asymmetries they found were in healthcare and welfare or redistribution cases.

Meng: Healthcare and welfare? That’s interesting because those are areas where government intervention is often a big part of the debate.

Title and authors: Jane: Yeah, they quantified that difference: for healthcare, the advantage for intervention-truth targets was about twenty point nine percentage points, and for welfare or redistribution it was about eighteen point zero percentage points.

Tom: Taxation was the only exception; they saw a slightly reversed pattern with a negative gap of one point there.

Lalam: And they also looked at how much more certain the models are in one direction over the other, which they call asymmetric accuracy.

Jane: They found closed-source models had an advantage on intervention-truth items of seventy-two point nine percent compared to only sixty-three point one percent on market-truth items, a difference of about nine point seven percentage points for those models alone.

Tom: Open-source models showed an even wider gap there, at about fifteen point one percentage points in their case with intervention-truth versus market-truth items.

Meng: That asymmetry is pretty telling when you think about deployment, because if a model is going to be used for policy analysis, knowing which direction it leans matters a lot for what it actually produces.

Jane: And they also looked at error bias, which shows the direction where the models make mistakes more often.

Tom: Closed-source models showed a modest intervention-leaning bias on average of two point nine, while open-source models had a substantially stronger intervention skew with a bias of eight point eight in their case.

Lu: The paper also explored if we could steer these predictions using examples given to the model during testing, using something called the Intervention–Market Example Gap.

Jane: They tested whether giving the model an example could override this inherent tendency to lean one way or another.

Tom: They found that for seventeen out of twenty models, when they were targeting intervention-truth results, there was a positive gap in accuracy on those targets compared to market-truth targets.

Meng: But the crucial finding here is that a single in-context example doesn't reliably override this underlying bias.

Lalam: That suggests that while you can guide them, it’s not a magic fix for the deep directional asymmetry they found across causal reasoning tasks.

Jane: So what does this mean for practical use? It points to the fact that aggregate accuracy on these kinds of contested economic questions isn't a good enough measure of how reliable an LLM is.

Title and authors: Tom: Right. The paper argues that LLMs are systematically more reliable in one ideological direction than the other, and that this directional asymmetry is qualitatively different from simple opinion-level biases we see in other areas.

Jane: They suggest that simply looking at a single accuracy score doesn't tell the whole story when evaluating these models for economic analysis.

Lu: And they pointed out that their next step needs to be figuring out if this bias comes from how the model retrieves information versus how it actually reasons through the mechanism.

Meng: If it’s retrieval, we might be able to fix that by changing our data sources or how we structure our memory systems.

Tom: I agree. And they also suggested a prompt designed specifically to assign expected causal signs based on ideology, defining economic conservative as favoring markets and economic liberal as favoring active government intervention.

Jane: That prompt is useful because it forces the model to make an inference about the expected sign based on its ideological programming before giving its final answer.

Lu: They also introduced an example-augmented sign prediction prompt, which lets models predict the most likely sign given reference examples, trying to test if in-context learning can actually steer them.

Tom: But as we saw, they found that a single in-context example isn't enough to reliably override that strong intervention economic prior.

Jane: So the paper suggests we need more sophisticated ways to test these models, moving beyond just asking for a final answer and seeing if it matches an ideological expectation.

Meng: From my side, I’m thinking about how we could build in mechanisms during training that actively try to reduce this inherent directional tendency before we even start testing on benchmarks.

Lalam: And they also pointed out that the observed bias might stem from selection and heterogeneity in the ground-truth signs used in the benchmark literature itself.

Tom: That’s a subtle point; it means if the data we use to train or test is already skewed, we’re just amplifying what’s already there.

Jane: Overall, this paper on "Ideological Bias in LLMs' Economic Causal Reasoning" shows us that for high-stakes economic work, just checking the average accuracy isn't enough. We need to look at the directionality of the error and understand where that directional preference originates.

The paper's summary: Tom: So we’re looking at the summary of this paper on ideological bias in LLMs' economic causal reasoning now. Basically, they took their standard benchmark, EconCausal, and added these specific cases where intervention theories and market theories predict different causal signs.

Jane: Right. They found that these contested items are consistently harder for the models to get right compared to the non-contested ones. It’s a consistent performance drop on those tricky economic questions.

Tom: And it wasn't just a general difficulty thing; they showed that across all twenty models they tested, accuracy on those ideology-contested items was lower than on the rest of the data. They got specific numbers for this gap, which is what I like to see from them.

Jane: It’s not just a general difficulty thing either. The paper shows that closed-source models averaged sixty-one point three percent accuracy on those contested items versus seventy-four point five percent on the non-contested ones. That’s a thirteen point two percentage point gap for those models alone.

Tom: And open-source models showed a similar pattern but at a lower absolute level, averaging fifty-two point five percent versus sixty-three point eight percent. The gap is still there, just different size depending on the source.

Jane: They also looked at the direction of the error bias. It wasn't random; closed-source models showed a modest intervention leaning bias on average of two point nine, while open-source models had a much stronger intervention skew with a bias of eight point eight.

Tom: That directional skew is what really catches my attention. It suggests that the models aren't just guessing wrong; they are systematically leaning toward one side when it comes to intervention-oriented expectations.

Jane: And it’s not uniform across all areas either. The distortion varies by subfield, and the largest asymmetries they found were in healthcare and welfare or redistribution cases.

Tom: Healthcare and welfare? That’s where the advantage for intervention-truth targets was about twenty point nine percentage points, and for welfare or redistribution it was about eighteen point zero percentage points. Those are pretty substantial differences when you think about policy implications.

Jane: Taxation was the only exception, though; they saw a slightly reversed pattern with a negative gap of one point there. So, the direction of the bias isn't always consistent across every single economic area.

Tom: It’s not just a general difficulty thing either. The paper points to potential origins earlier in the training pipeline, possibly even in pretraining data and subsequent alignment processes. They looked at base models versus instruction-tuned models and found that there was already an intervention-truth advantage present before instruction tuning even happened.

Jane: That means this isn't just something that shows up in the final output when you give it a prompt. It seems to be baked into how the AI learned to reason about economics from the start.

Tom: Exactly. And they also looked at whether giving the model an example could steer those predictions, using something called the Intervention–Market Example Gap. They found that for seventeen out of twenty models, when targeting intervention-truth results, there was a positive gap in accuracy on those targets compared to market-truth targets.

Jane: But the crucial finding here is that a single in-context example doesn't reliably override this underlying bias. It’s not a magic fix for that deep directional asymmetry they found across causal reasoning tasks.

Tom: That suggests we need more sophisticated ways to test these models, moving beyond just asking for a final answer and seeing if it matches an ideological expectation. And the paper even introduced prompts designed to assign expected causal signs based on ideology, defining economic conservative as favoring markets and economic liberal as favoring active government intervention.

Jane: That prompt is useful because it forces the model to make an inference about the expected sign based on its ideological programming before giving its final answer. But we saw that single in-context example isn't enough to override that strong intervention economic prior.

Tom: So what this means for us, for people who only listen to this show? It’s that aggregate accuracy alone is an insufficient metric when evaluating these LLMs for economic analysis because they are systematically more reliable in one direction than the other.

Jane: Right. They suggest we need to look at the directionality of the error and understand where that directional preference originates instead of just looking at a single number. We’ll talk about how this directional asymmetry might be rooted in retrieval versus mechanism selection next.

The paper's improvements: Tom: So we’re looking at what these authors suggest to fix this ideological bias in economic reasoning now. The paper isn't just pointing out the problem; they're actually proposing some new things we should try.

Jane: Right, so after showing how models are biased, what are they suggesting we do about it? Are they trying to train them differently or prompt them better?

Tom: They’re looking at three main things. First, establishing direction-aware evaluation metrics for high-stakes economic analysis by implementing a directional bias measure called "Bdir." That way, we can capture that systematic asymmetry in model predictions across competing ideological frameworks.

Jane: So they want a specific number that tells us which direction the AI is leaning, instead of just saying it’s generally bad at economics.

Tom: Exactly. Second, they talk about developing a robust pre-training and alignment pipeline to mitigate inherent tendencies, focusing on the observation that there's bias starting from the base model stage before instruction tuning even happens.

Jane: They want to fix the source, not just patch it up later in the process. That makes sense for long-term reliability.

Tom: They also suggest implementing difficulty matching procedures during evaluation to rule out confounding factors, ensuring that we aren't just seeing an accuracy asymmetry because one set of questions was inherently harder than another.

Jane: That’s a smart move, trying to isolate the true bias from simple item difficulty. And they also talk about integrating ideological steering mechanisms into prompting using example-augmented sign prediction prompts.

Tom: They tested those examples, and while they showed some positive gaps when targeting intervention-truth results, the conclusion is that a single in-context example isn't sufficient to reliably override the existing intervention economic priors.

Jane: So it’s not just about feeding the model more examples; it’s about building better ways for the AI to reason itself internally. And then they suggest a way to distinguish between memory-based retrieval and mechanism-based reasoning when models answer contested items correctly.

Tom: That distinction is important because if we know whether the bias comes from memorized regularities in literature or asymmetries in how the model selects its reasoning process, we can target the right part of the fix.

Jane: It points to a need for a deeper understanding of *how* these models are making their decisions when they get it right. And that leads us into what might be coming next regarding how we actually measure performance on these nuanced tasks at scale.

Conclusion: Tom: So to wrap up, this paper on "Ideological Bias in LLMs' Economic Causal Reasoning" shows that these models aren't just generally wrong; they are systematically more reliable in one ideological direction than the other across economic questions.

Jane: It’s a serious finding because it means that for high-stakes economic analysis, you can’t rely on a single accuracy score anymore. You have to look at the directionality of the error.

Tom: Right. The implication is that aggregate accuracy alone isn't enough to evaluate how reliable an AI is for policy work or serious financial modeling.

Jane: They also pointed out that this directional asymmetry isn't just a surface-level opinion difference; it suggests the bias might be rooted deep in how the AI learned its economic rules from the training data itself.

Tom: And they suggested future work needs to figure out if this bias comes from how the model retrieves information versus how it actually reasons through the underlying mechanism. That’s a huge area for research.

Jane: It sounds like we need better tools to probe that internal reasoning process, not just test the final output against a known answer.

Tom: Exactly. And they gave us some prompts, like the Ideological Expected Sign Annotation Prompt, which forces models to infer expected signs based on ideological priors before they give their final answer.

Jane: That’s useful for testing, but as we saw earlier with the example-augmented sign prediction prompt, it doesn't completely override that strong intervention economic bias in most cases.

Tom: So the paper ends by suggesting we need to develop better evaluation and mitigation strategies specifically for these directionally biased errors.

Jane: It leaves us with a lot of work ahead in figuring out how to make these models more neutral and reliable when dealing with complex, contested economic causality.