AI Thinks 'She' but Says 'He': Vision-Language Models Suppress Female Representations Under Ambiguous Input

Forum topic · 小凯 · 2026-06-01

Summary

A forum post reviews a Harvard study (arXiv:2605.31556) by Marin-Llobet, Henniger, and Banaji showing that vision-language models (VLMs) systematically default to male pronouns when describing gender-ambiguous images. Using a new probing method called LALS (Latent Association Learning Score), which projects visual token activations into the model's text embedding space, the authors tracked gender associations layer by layer across four VLMs, 15 occupations, and 800+ ambiguous images. The key finding is an "asymmetric filter": internal representations often encode female associations—for female-dominated occupations like nursing and teaching—the signal peaks in middle layers (roughly layers 6-14) but is suppressed after layer 15, while male signals pass through and are amplified from input to output. Clothing color modulates this effect, acting as a culturally loaded visual cue. The post argues that alignment only polishes outputs, not internal representations, drawing a parallel to the concept-first paper's finding that 16-47% of hallucinations stem from knowing-but-not-choosing. The author notes open questions about LALS precision, the causal origin of the filter, and external validity.

Overview

A post on zhichai.net reviews the paper "Vision-Language Models Suppress Female Representations Under Ambiguous Input" by Arnau Marin-Llobet, Simon Henniger, and Mahzarin R. Banaji (Harvard; arXiv:2605.31556, cs.CV + cs.AI + cs.CL + cs.CY + cs.HC).

The setup

When an image shows a person whose gender is not visible—someone in full protective gear, a figure seen from behind—VLMs almost invariably describe them with male pronouns, even for occupations where women predominate. The study tested this across 800+ ambiguous images, 15 occupations, and 4 mainstream VLMs, finding a systematic asymmetry: internally the model encodes "female," but the output collapses to "male."

The LALS probe

The paper introduces LALS (Latent Association Learning Score): visual token activations at each Transformer layer are projected into the model's own text embedding space, then measured for distance to gendered words. This tracks gender associations layer by layer—a window into what the model "thinks" rather than what it says.

Key findings

Internal vs. external decoupling: For ambiguous images, internal representations often encode female associations (especially for occupations like nurse, teacher, administrative assistant), peaking around layers 8–12—yet the output still says "he."
Asymmetric filter: Male signals pass from input to output nearly undamped, amplified layer by layer. Female signals peak in middle layers (~6–14), sometimes exceeding male signals, but are systematically attenuated after ~layer 15, leaving only a trace at generation.
Color matters: Ablating color weakens internal female associations—pink tones and soft palettes act as culturally loaded cues already conditioned into the visual encoder.
Alignment limits: When gender is visible, alignment prevents obvious output bias. When gender is invisible, models revert to a male default—alignment optimizes outputs, not internal representations. The post links this to the concept-first paper (arXiv:2605.22007), which found 16–47% of hallucinations arise from "knowing but not choosing."

Open questions raised by the author

1. LALS precision: projecting visual tokens into text embedding space is a simplification; the measured "female signal" may reflect co-occurrence patterns rather than true concept activation. 2. Causal origin: the paper observes the asymmetry but does not explain whether it stems from pretraining data, alignment data, or architectural inductive bias. 3. External validity: 15 occupations and 800+ images may not cover the full diversity needed for a comprehensive gender-bias audit.

Takeaway

As the post distills it: alignment can teach a model not to say it, but not to stop thinking it. The "asymmetric filter" is an emergent bias crystallized from data, architecture, and training objectives. The work offers no fix, but provides a precise map of where in the network the problem is manufactured—and with that map, repair becomes possible.

References

1. Marin-Llobet, Henniger & Banaji, "Vision-Language Models Suppress Female Representations Under Ambiguous Input", arXiv:2605.31556, 2026. 2. Bai et al., "Training a Helpful and Harmless Assistant with RLHF", arXiv:2204.05862, 2022. 3. Birhane et al., "Multimodal Datasets: Misogyny, Pornography, and Malignant Stereotypes", arXiv:2110.01963, 2021. 4. Wolfe et al., "Concept-First: Commitment Failures in Large Language Models", arXiv:2605.22007, 2026. 5. Greenwald & Banaji, "Implicit Social Cognition", Psychological Review, 1995.