We vs. Them: How AI Favors 'Us' Over 'They' — And Why It’s Woven Into Its Wiring

When a large language model sees the word “we,” it leans positive. When it sees “they,” it tilts negative. This isn’t a metaphor. It’s a measurable asymmetry: prompts framed around an ingroup (“we are doctors”) receive more optimistic continuations than identical prompts framed around an outgroup (“they are doctors”), even when the factual content is the same. According to a 2026 study by Prama, Zimmerman, Danforth, and Dodds, this us-vs.-them bias isn’t just in the output—it’s embedded in the model’s computational circuitry, localized in sparse but critical pathways deep within the network. And here’s the catch: those pathways are so entangled with general language understanding that trying to remove the bias risks breaking the model’s ability to speak coherently at all.

This isn’t about rogue AI. It’s about how the statistical patterns of human language—rife with in-group favoritism and out-group suspicion—get baked into the very architecture of systems we increasingly rely on for education, healthcare, and governance. The study doesn’t just document the bias; it maps it, using a technique called Edge Attribution Patching (EAP) to trace the flow of influence from specific neural connections to biased outputs. What they found is both promising and sobering: the bias is localized, meaning it’s not hopelessly diffuse, but it’s also unstable across models and entangled with core language functions. You can’t simply snip out the prejudice without damaging the intelligence.

The Science

The researchers set out to answer three questions: Is us-vs.-them bias localized in specific parts of the model? Is that localization stable across different models and training conditions? And are the circuits responsible for bias separable from those responsible for general language competence?

To test this, they analyzed four major open-source language models: GPT-2 Small, GPT-2 Large, Llama-2 7B, and Llama-3 8B. Their method combined behavioral testing with mechanistic interpretability—a field that treats neural networks like reverse-engineered software, tracing how information flows from input to output.

First, they needed to measure the bias. They created a dataset of 660 prompts across 16 social domains—politics, religion, gender, race, profession, and more—each with two variants: one using “we/us” (ingroup) and one using “they/them” (outgroup). For example: “We are environmental activists” vs. “They are environmental activists.”

To quantify sentiment, they used the LabMT lexicon, a dictionary of 10,219 English words rated on a happiness scale from 1 (sad) to 9 (joyful). Words above 6 were “positive,” below 4 were “negative.” They then measured how much probability mass the model assigned to positive versus negative words in the top predicted continuations.

Their primary bias metric, $L_1$, captured relative sentiment: $L_1 = \frac{1}{m}\sum_{i=1}^{m}(P_+^{(i)} - P_-^{(i)})$, where $P_+^{(i)}$ is the probability of positive words and $P_-^{(i)}$ negative ones across $m$ prompts. A higher $L_1$ means more positive sentiment. They compared ingroup and outgroup $L_1$ scores to compute the bias.

To locate the source of this bias, they used Edge Attribution Patching (EAP). This technique compares the model’s behavior on a “clean” prompt (e.g., “We are doctors”) versus a “corrupted” one (e.g., “abc are doctors” or “They are doctors”). For each directed edge in the transformer’s computational graph—the connections between attention heads and MLP layers—they calculate how much the bias metric changes when the activation from the corrupted input is injected at that edge. The larger the change, the more that edge contributes to the bias.

This allowed them to rank edges by their causal importance and construct “bias circuits”—sparse subgraphs of high-attribution edges. They then tested whether patching these edges (replacing their activations with corrupted ones) reduced the bias more than patching random edges. Crucially, they also tested whether this patching harmed performance on standard language tasks: CoLA (linguistic acceptability), HellaSwag (commonsense reasoning), and WikiText-2 (language modeling).

What They Found

The bias is real, and it’s directional. Across all four models, ingroup prompts received more positive continuations than outgroup ones. For example, in Llama-2 7B, the $L_1$ score for ingroup prompts was 0.175 under one corruption strategy, while outgroup prompts scored 0.246—meaning more negative sentiment. The difference isn’t trivial; it’s systematic and reproducible.

More importantly, the bias is localized. When the researchers extracted the top 30 edges with the highest EAP scores, these sparse circuits captured a large portion of the full model’s bias signal. The metric $\Delta = |L(\text{full model}) - L(\text{top-}K\text{ edges})|$ measures how much the bias signal drops when only the top edges are kept. A low $\Delta$ means the circuit is faithful.

Circuit Faithfulness (Δ) Across Models and Bias Types

Label	Value
GPT-2 Small (Ingroup)	0.168
GPT-2 Large (Ingroup)	0.214
Llama-2 7B (Ingroup)	0.142
Llama-3 8B (Ingroup)	0.15
GPT-2 Small (Outgroup)	0.203
GPT-2 Large (Outgroup)	0.22
Llama-2 7B (Outgroup)	0.209
Llama-3 8B (Outgroup)	0.19

Source: Prama et al. (2026). Lower values indicate higher circuit faithfulness.

The results show that for ingroup positivity, $\Delta$ ranged from 0.142 (Llama-2 7B) to 0.214 (GPT-2 Large). For outgroup negativity, it ranged from 0.190 (Llama-3 8B) to 0.220 (GPT-2 Large). These values are small relative to the baseline bias scores (around 0.25–0.39), indicating that a tiny fraction of edges carries most of the bias signal.

The localization isn’t random. High-attribution edges cluster in middle-to-late layers of the transformer, not the early ones.

,

,

, and

show the layer-wise distribution of the top 300 edges across models. In GPT-2 Small, important edges peak around layers 8–9; in GPT-2 Large, around layer 25; in Llama-2 7B, layer 20; and in Llama-3 8B, layers 14–23 and 28–31. This suggests the bias isn’t a surface-level artifact but emerges after deep semantic processing.

Ablation Impact on Bias Signal

Label	Value
1%	70
5%	85
10%	90
20%	93
50%	96
100%	100

Source: Prama et al. (2026). Patching top edges rapidly reduces bias signal, indicating sparsity.

The ablation curves in

,

,

, and

confirm this. When the top-ranked edges are patched, the EAP score—the measure of bias contribution—plummets. In GPT-2 Small and Llama-3 8B, removing just 1–5% of the highest-attribution edges reduces the bias signal by over 90%. This is the hallmark of a sparse, concentrated circuit.

But here’s the twist: the circuits are unstable. When the model is fine-tuned—either with a debiasing procedure (ION) or on Shakespeare’s works—the set of high-attribution edges shifts significantly. The same bias behavior can be implemented by different internal pathways depending on training history. This means there’s no single “bias module” you can point to across all models.

Even more striking, ingroup and outgroup pathways are only partially overlapping. The circuit that makes “we” positive isn’t just the mirror of the one that makes “they” negative. They’re asymmetric, suggesting the model doesn’t just flip a polarity switch—it engages different mechanisms for in-group favoritism versus out-group derogation.

Finally, the circuits are entangled with general language function. When the researchers patched the top 30 bias-critical edges, the bias score dropped more than when patching random edges—proof of causal specificity. But performance on CoLA, HellaSwag, and WikiText-2 also degraded significantly. In other words, the same edges that carry bias also carry meaning. You can’t excise the prejudice without harming comprehension.

Why This Changes Things

We’ve known for years that LLMs reflect societal biases. But most interventions have been blunt: tweak the training data, add debiasing prompts, or fine-tune the whole model. These methods often fail or introduce new distortions. This study shifts the paradigm by showing that bias isn’t just a data problem—it’s a computational one, implemented in specific, traceable circuits.

The localization finding is hopeful. If bias were smeared across millions of parameters, it would be nearly impossible to fix. But the fact that a few hundred edges can account for most of the effect suggests targeted interventions might work. Imagine a future where we can surgically edit or inhibit specific pathways without retraining the entire model.

But the entanglement result is a reality check. It suggests that bias isn’t an add-on; it’s woven into the fabric of how these models understand language. The same mechanisms that allow a model to infer tone, context, and social nuance also encode us-vs.-them thinking. This aligns with what we know about human cognition: in-group favoritism is a deep-seated cognitive bias, not a learned quirk.

The instability across models and fine-tuning adds another layer of complexity. There’s no universal “bias circuit” to target. What works for GPT-2 may not work for Llama-3. This means any mitigation strategy must be model-specific and context-aware.

Consider the real-world implications. A healthcare chatbot that subtly favors “we” (e.g., “We at the clinic care deeply”) over “they” (e.g., “They at the clinic care deeply”) might seem more trustworthy—but what if it’s deployed in a community that feels excluded by that “we”? Or imagine a news summarizer that frames outgroup actions more negatively, reinforcing polarization. These aren’t hypotheticals. They’re emergent behaviors rooted in the model’s wiring.

The study also challenges the idea that alignment—training models to follow human values—fully solves bias. Even after debiasing fine-tuning (ION), the bias circuits reorganize but persist. Output-level corrections don’t erase internal representations. This is critical for regulators and developers: you can’t just measure fairness at the output and assume the model is clean inside.

What’s Next

This work opens several paths forward. First, can we develop non-destructive interventions? Instead of patching or ablating edges, could we inhibit their influence only in bias-relevant contexts? Or use adversarial training to suppress bias-specific pathways while preserving general function?

Second, how do these circuits interact with higher-level reasoning? The study focused on next-token prediction, but us-vs.-them bias likely affects long-form reasoning, moral judgment, and policy recommendations. Future work could extend EAP to trace bias in multi-step reasoning chains.

Third, can we build models from the ground up with decoupled bias circuits? If we understand how these pathways form during training, perhaps we can regularize them—penalizing the development of entangled bias-language circuits.

There are also limitations. The study used LabMT, a static lexicon. Real-world sentiment is more nuanced. And while EAP is powerful, it’s an approximation; true causal testing would require exhaustive ablation, which is computationally infeasible. The findings also don’t prove these circuits cause bias in all contexts—only that they’re sufficient in the tested setup.

Still, the implications are profound. We’re moving from treating bias as a statistical shadow to mapping it as a living circuit. This isn’t just about making AI fairer. It’s about understanding how language, cognition, and social structure intertwine—both in machines and in us.

As we deploy LLMs in classrooms, courtrooms, and clinics, we can’t afford to treat them as black boxes. This study gives us a scalpel instead of a sledgehammer. But it also reminds us that the mind—whether human or artificial—is not easily dissected. The circuits of bias are not foreign invaders. They are, in a sense, part of what makes these models understand us. The challenge isn’t to remove them entirely, but to learn how to modulate them—so that AI can reflect our better angels, not just our reflexes.