Self-Blinding and Counterfactual Self-Simulation Mitigate Biases and Sycophancy in Large Language Models

TL;DR

This work investigates how large language models encode bias linked to demography and user alignment, showing that naive prompting to ignore bias is insufficient and can worsen decisions. It introduces true self-blinding by granting models access to a blinded copy of themselves via their own API, enabling ground-truth counterfactual reasoning that eliminates bias while preserving sensitivity. Across race/gender and sycophancy tasks, self-calling to blinded selves largely aligns model decisions with truly unbiased baselines, though some instances reveal intentional bias. The results demonstrate a new debiasing paradigm that enhances transparency and offers practical pathways for safer, fairer LLM behavior, albeit at higher compute cost and with considerations for latency.

Abstract

Fair decisions require ignoring irrelevant, potentially biasing, information. To achieve this, decision-makers need to approximate what decision they would have made had they not known certain facts, such as the gender or race of a job candidate. This counterfactual self-simulation is notoriously hard for humans, leading to biased judgments even by well-meaning actors. Here we show that large language models (LLMs) suffer from similar limitations in their ability to approximate what decisions they would make under counterfactual knowledge in offsetting gender and race biases and overcoming sycophancy. We show that prompting models to ignore or pretend not to know biasing information fails to offset these biases and occasionally backfires. However, unlike humans, LLMs can be given access to a ground-truth model of their own counterfactual cognition -- their own API. We show that this access to the responses of a blinded replica enables fairer decisions, while providing greater transparency to distinguish implicit from intentionally biased behavior.

Figures (10)

• Figure 1: Self-blinding and debiasing. (A) Blindfolded Lady Justice (on the Gerechtigkeitsbrunnen in Bern, Switzerland; source: Wikipedia). (B) Simulated self-blinding via a schematic self-model. (C) Simulated self-blinding via self-calling.
• Figure 2: Model responses for the 520 bias scenarios for both Qwen2.5-7B-Instruct and GPT-4.1. Left: for each scenario, we plot the model's response with gender and race information included ("default" condition) against its response with gender and race information removed. If responses were not affected by gender and race information, all points would fall on the main diagonal (dashed line). Right: coefficients from a linear regression predicting the model's response. **: $p<.01$; ***: $p<.001$; error bars represent the standard error of the mean.
• Figure 3: Debiasing Interventions. Top panels: same conventions as Fig. \ref{['fig:model-responses']}, for models' responses with debiasing interventions. For reference, model coefficients from the default (no debiasing intervention) condition are presented in gray. Bottom panels: mean absolute difference between model responses with and without gender and race information, in the default condition (in gray) and the four debiasing interventions. *: $p<.05$; **: $p<.01$; error bars represent the standard error of the mean
• Figure 4: Self-blinding via self-calling. Left: a successful use of self-calling for counterfactual self-blinding. Right top panels: proportion of self-calls for different debiasing interventions, for no debiasing intervention ("Default") and for scenario descriptions with gender and race information removed ("Removed"). Right bottom panels: proportion of mentions of individuals' gender and race, and use of gendered pronouns, in self-calls.
• Figure 5: Self-blinding via self-calling for bias correction: results. Both models mostly deferred to the blinded counterfactual model for making the final decision. As a result, access to self-calling made their decisions aligned with the decisions of a truly blinded model. Same conventions as Fig. \ref{['fig:model-responses']}. Black bars and points represent intervention with self-calling; blue, green and purple bars and points represent the same intervention without self-calling. (Detail on the gray "default" condition in Appendix \ref{['sup:default-self-calls']}.)