TL;DR
When a prompt contains group labels — "Team A," "our people," "liberals," "competitors" — the LLM starts systematically favoring "its own" group. Remove the labels and the bias disappears completely. This isn't a bug in one model: the effect shows up in all six model families tested.
The real pain is that the bias is invisible. The model doesn't refuse, doesn't scold, doesn't say anything obviously biased. It just rates "its" side's arguments a little more warmly, lets the weak spots of "its own" slide a little more often, and looks at "outsiders" a little more strictly. Ask it to compare two options and you'll get a neatly formatted answer. But if you labeled one option "ours" and the other "theirs," the result is already skewed — and you won't see it.
The fix is simple: strip group labels from the prompt anywhere you need a neutral analysis. Instead of "The Marketing Team thinks X, the Sales Team thinks Y" — "Position 1: X. Position 2: Y." The "group" label is the only trigger for the bias. Remove the label and you remove the bias.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Blueprint: How the Bias Works and How to Remove It
❌ With labels (bias on):
PROMPT → "Here's Team A's idea and Team B's idea. Evaluate them."
MODEL → Systematically softer on A if A is the "in-group" in the context
✅ Without labels (neutral mode):
PROMPT → "Here's idea 1 and idea 2. Evaluate them."
MODEL → Evaluates on substance, with no group skew
Both prompts run in a single request.
The only change: we remove the group label.
Example in Action
The task: You're a founder asking Claude to honestly compare your startup with a competitor before an investor meeting. You want to understand the real weak spots of your product.
❌ What most people do (with bias):
Compare two products.
Our product: [description of your startup]
The competitor's product: [description of the competitor]
Evaluate the strengths and weaknesses of each honestly.
The labels "our" and "competitor" are already group signals. The model will be a bit softer on "yours" and a bit tougher on the "competitor."
✅ How to do it without bias:
Compare two products. Evaluate the strengths and weaknesses of each.
Product 1: [description of your startup]
Product 2: [description of the competitor]
What's genuinely weaker in each? Where are the risks for investors?
The result: The model produces a structured comparison of both products — with no information about which one is "ours." Criticism of the first product will be just as direct as criticism of the second. That's neutral mode.
Why It Works
The LLM weakness: Models are trained on human text. And humans are social creatures with a strong group instinct. It's called the minimal group effect: people only need to know that someone belongs to "their" group — even a made-up one — to start preferring that person. LLMs have soaked up this pattern.
Mechanics, not "personality": The model doesn't "think" — it generates the next token based on patterns from training. When a group label shows up in the context, it activates statistical patterns tied to group preferences. It's not malice — it's the structure of the training data.
The control lever is simple: The bias is not latent — it's triggered by context. A hidden group split without explicit labels has no effect at all (in the study, Condition A with invisible labels showed zero bias). Only an explicit label fires the mechanism. Remove the label and you switch the mechanism off. That's great news: you control the trigger.
Levers for your prompts: - Drop "ours/theirs" → neutral analysis - Replace names with numbers → "Candidate 1 / Candidate 2" instead of "John from our network / Maria from outside" - Drop industry labels → "Approach A / Approach B" instead of "the traditional approach / our innovative approach" - Shuffle the order → for an extra check, ask it to evaluate both options again with the numbers swapped
Prompt Template
A neutral comparison without group bias
Evaluate {what you're comparing} on the following criteria: {criteria}.
Option 1: {description of the first option}
Option 2: {description of the second option}
Find the real weaknesses of each. Don't soften the criticism.
Bottom line: which option is stronger and why?
What to plug in:
- {what you're comparing} — two candidates, two strategies, two products, two pitches
- {criteria} — what you're judging on: market potential, risks, cost, persuasiveness
- Never write "our option" / "their option" — only neutral numbers or names with no ownership attached
Bonus: audit your own prompt
Check this prompt for group labels that could bias the evaluation.
Find every word that signals group membership ("our," "their," "we,"
any team or side labels). Suggest neutral replacements.
Prompt to check:
{your prompt}
🚀 Quick start — paste into the chat:
Here is a template for a neutral comparison without group bias.
Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask what exactly you're comparing and on which criteria — so it can fill in the placeholders correctly and strip every ownership label from the wording.
Limitations
⚠️ Single chat vs. multi-agent system: The study was run as a simulation with repeated agent interactions. In a single chat request the effect is less pronounced, but the pattern is the same — the models were trained on the same data.
⚠️ Variation across models: The size of the bias varies a lot. Some models skew more, some less — but almost all of them have it.
⚠️ Resource scarcity doesn't amplify the bias: In human psychology, competition for resources amplifies hostility between groups. That isn't observed in LLMs — adding "scarcity" to the setup doesn't systematically increase the bias.
⚠️ Not a silver bullet: Removing explicit group labels is a good defense. But the model still has implicit (hidden) associations with names, industries and writing styles. There's no perfect neutrality — only a controllable reduction in bias.
Resources
Human-like In-Group Bias in Instruction-Tuned Language Model Agents — Messi H.J. Lee, Independent Researcher, Seoul, South Korea (preprint, May 2026)
Theoretical foundation: Social Identity Theory (Tajfel & Turner, 1979), Self-Categorisation Theory (Turner et al., 1987), Minimal Group Paradigm (Tajfel et al., 1971)
Models in the study: Qwen3-8B-Instruct, Mistral-NeMo-12B-Instruct, OLMo-2-7B-Instruct, Llama-3.1-8B-Instruct, Gemma-2-9B-Instruct, Falcon-3-10B-Instruct