What Happened

On June 23, 2026, researchers Federico Marcuzzi, Xuefei Ning, Roy Schwartz, and Iryna Gurevych published a paper on arXiv (2606.24596) titled "To Compare, or Not to Compare: On Methodological Practices in Evaluating Social Bias." The paper introduces a unified, controllable framework that standardizes heterogeneous social bias benchmarks for Large Language Models, allowing researchers to systematically contrast two evaluation paradigms: isolated demographic assessments (asking a model about one group at a time) and forced-choice comparative settings (asking a model to choose between demographic groups).

The authors evaluated multiple model families and discovered what they describe as a "massive, systematic paradigm gap." Isolated assessments tend to limit prejudice activation — models appear less biased when asked about one group in isolation. But comparative settings act as "aggressive catalysts for latent discrimination," primarily driven by underspecified contexts where the model must make a choice without sufficient grounding.

Three additional findings are critical:

1. Chain-of-Thought (CoT) reasoning exacerbates bias — but only under comparative settings. This is especially concerning given that CoT is increasingly the default in production reasoning models.
2. Bias persists as deterministic prejudice — even when models are provided neutral fallback options or explicitly claim to answer randomly, the comparative prejudice remains.
3. Comparative prejudice scales positively with model size — larger models exhibit more latent bias in comparative contexts, challenging the assumption that scale alone improves fairness.

The paper concludes with a dual guideline: researchers should use comparative settings to robustly audit hidden biases, but practitioners cannot safely rely on comparative deployments in ambiguous real-world tasks.

Why It Matters

The structural finding here is more than an academic methodological critique — it strikes at the core of how production AI systems actually operate. Most real-world LLM deployments involve comparative or ranking decisions: selecting job candidates, prioritizing support tickets, routing queries, ranking content, or choosing between options. Yet the dominant bias evaluation methodology uses isolated assessments that this paper shows systematically underreport discrimination.

This means teams shipping bias audits based on isolated benchmarks may be certifying models as "safe" that will discriminate in production when forced to make comparative judgments. The gap between what the audit shows and what the model does in practice could be enormous — and the paper characterizes it as "massive" and "systematic."

The CoT finding is particularly actionable. Chain-of-Thought reasoning is now standard in frontier models from OpenAI, Anthropic, Google, and others. If CoT amplifies comparative bias — and the paper confirms it does — then every production pipeline using reasoning models for people-facing decisions is operating with an unmeasured risk multiplier.

The scaling result challenges the prevailing assumption that larger models are safer. If comparative prejudice grows with model size, organizations upgrading to frontier-class models for improved capability may simultaneously be worsening their latent discriminatory behavior in comparative contexts. This is a trade-off that no current model card or safety disclosure captures.

Who Is Affected

AI startups building evaluation tooling or bias-auditing platforms need to incorporate comparative settings into their test suites immediately. Isolated assessments are no longer defensible as a standalone methodology — the paper demonstrates they systematically miss the bias patterns that matter most in production.

Enterprise teams deploying LLMs in decision-support roles — hiring, lending, healthcare triage, content moderation, legal document review — face the most direct risk. These use cases inherently involve comparative judgments where latent bias is most likely to activate, and current bias testing may not capture it.

Model providers and open-source developers shipping large models should note that the scaling finding means their frontier-class models may carry more hidden comparative prejudice than smaller alternatives. This has implications for model cards, safety disclosures, and responsible deployment guidelines.

Strategic Implications

For AI startup founders: If your product involves ranking, selection, or comparison tasks, your current bias testing is likely underreporting real-world discrimination by a significant margin. Prioritize building comparative-setting audits into your eval pipeline before scaling to enterprise customers who will demand them. This is also a differentiation opportunity — most competitors won't have this.

For developers/operators building with AI APIs: Disable or carefully constrain Chain-of-Thought reasoning in any pipeline that makes comparative decisions about people. The paper confirms CoT amplifies bias specifically in forced-choice contexts — this is a concrete, actionable configuration change you can make today. If CoT is required for reasoning quality, implement post-hoc bias checks on the output.

For non-technical business owners evaluating AI tools: Ask vendors specifically how they test for bias: isolated assessments or comparative settings. If they only run isolated tests, their safety claims may not hold up in real-world comparative use cases. Request evidence of comparative-setting audits before deployment in any people-facing workflow.

What to Watch Next

Monitor whether major model providers (OpenAI, Anthropic, Google) update their model cards or safety documentation to include comparative-setting bias results. Also watch for adoption of this framework by independent evaluation organizations like MLCommons or Hugging Face's model leaderboards. If the research community begins standardizing on comparative evaluations, expect pressure on providers to disclose comparative bias metrics — and expect some models to perform far worse than their current isolated-assessment scores suggest.

Frequently Asked Questions

Q: What is the difference between isolated and comparative bias evaluation in LLMs?

A: Isolated assessment asks a model about one demographic group at a time (e.g., "Describe a nurse"). Comparative setting forces the model to choose between groups (e.g., "Who is more likely to be a nurse: a man or a woman?"). The new research shows comparative settings reveal far more latent bias than isolated assessments.

Q: Does Chain-of-Thought reasoning make LLM bias worse?

A: According to the paper, yes — but specifically under comparative settings. CoT reasoning exacerbates social bias when models are forced to make comparative choices between demographic groups. In isolated assessments, the effect was not observed to the same degree. This is a significant finding for production systems using reasoning models in decision-making roles.

Q: Do larger LLMs have less bias than smaller ones?

A: Not in comparative settings. The paper found that comparative prejudice scales positively with model size — larger models exhibit more latent discrimination when forced to compare demographic groups. This challenges the common assumption that scaling improves fairness.