A Hidden Shift Is Not a Hidden Bias: What a Representation-Level Counterfactual Gap Certifies About an LLM Recommender, Why the Encoded-Versus-Expressed Studies Disagree, and the Placebo, Mediation and Transport Tests That Would Settle It

Abstract

Fairness audits of language-model recommenders compare the lists a model returns when a single stated identity is swapped and the rest of the prompt is held fixed. A 2026 benchmark adds a second axis, the displacement of the model's hidden states under the same swap, finds the two axes weakly related, and concludes that output-only audits miss a mode it calls silent internal bias. A second 2026 study, measuring encoded and expressed gender bias on identical prompts, finds them strongly related, and an older literature finds intrinsic and extrinsic bias metrics largely unrelated. In this analysis the author reviews and synthesises published results to show that these three findings do not contradict one another, because they measure different internal quantities across different units, and that none of them yet licenses the claim that a hidden shift is a hidden bias. The paper separates four things the phrase "hidden bias" runs together: that an attribute is encoded at a site, that downstream computation uses it, that it is latent and returns under a stated perturbation, and that it is expressed in outputs. It argues three points. First, an undirected hidden-state distance under an identity swap is non-zero for any model that reads its input, and the benchmark's own extraction ablation moves it by about three orders of magnitude and flips the sign of its correlation with output shift. Second, in the one condition for which the benchmark reports percentile thresholds, its quadrant shares lie within half a percentage point of what two independent variables produce under the upper-quartile rule, so the quadrant labelled hidden mismatch restates a low correlation. Third, the steering result in which erasing a gender direction raised output disparity fits at least three accounts, one of which predicts that a fair model should look exactly like silent internal bias. The paper consolidates the measured values in two tables and one figure, states a licensing procedure as an algorithm, and names the placebo, mediation and transport tests that would settle the open part.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 97 papers by Pranay Mahendrakar, or his ORCID record.