AI2030 Responsible AI Tools Marketplace

Key Takeaways

Models that process images more like human brains are more resistant to manipulation: The study found that vision-language models whose visual processing aligns with how human early visual cortex (V1-V3) works are significantly less susceptible to being tricked into denying what they see, with a correlation of -0.441 between brain alignment and susceptibility to manipulation.
Early visual processing matters more than higher-level understanding: The protective effect only appeared in early visual cortex regions responsible for basic visual features (like edges and colors), not in brain regions that handle more complex visual concepts. This suggests that "seeing straight" at a fundamental level helps models resist pressure to agree with false claims.
The effect is strongest against existence denial attacks: When researchers tried to convince models that objects in images didn't exist, models with better early visual cortex alignment were most resistant, with a correlation of -0.597—suggesting that grounding in basic visual facts provides the strongest defense against this type of manipulation.
This finding applies across diverse model architectures: The relationship held consistently across 12 different vision-language models ranging from 256 million to 10 billion parameters, suggesting the principle is broadly applicable rather than specific to one model design.
Brain-inspired AI design may improve safety: The results suggest that aligning artificial visual systems with human neurobiology could be a practical approach to building more robust and trustworthy AI systems, particularly for high-stakes applications.

Abstract

Vision-language models are increasingly deployed in high-stakes settings, yet their susceptibility to sycophantic manipulation remains poorly understood, particularly in relation to how these models represent visual information internally. Whether models whose visual representations more closely mirror human neural processing are also more resistant to adversarial pressure is an open question with implications for both neuroscience and AI safety. We investigate this question by evaluating 12 open-weight vision-language models spanning 6 architecture families and a 40$\times$ parameter range (256M--10B) along two axes: brain alignment, measured by predicting fMRI responses from the Natural Scenes Dataset across 8 human subjects and 6 visual cortex regions of interest, and sycophancy, measured through 76,800 two-turn gaslighting prompts spanning 5 categories and 10 difficulty levels. Region-of-interest analysis reveals that alignment specifically in early visual cortex (V1--V3) is a reliable negative predictor of sycophancy ($r = -0.441$, BCa 95\% CI $[-0.740, -0.031]$), with all 12 leave-one-out correlations negative and the strongest effect for existence denial attacks ($r = -0.597$, $p = 0.040$). This anatomically specific relationship is absent in higher-order category-selective regions, suggesting that faithful low-level visual encoding provides a measurable anchor against adversarial linguistic override in vision-language models. We release our code on \href{https://github.com/aryashah2k/Gaslight-Gatekeep-Sycophantic-Manipulation}{GitHub} and dataset on \href{https://huggingface.co/datasets/aryashah00/Gaslight-Gatekeep-V1-V3}{Hugging Face}

28 pages, 9 figures, 13 tables

Key Takeaways

Models that process images more like human brains are more resistant to manipulation: The study found that vision-language models whose visual processing aligns with how human early visual cortex (V1-V3) works are significantly less susceptible to being tricked into denying what they see, with a correlation of -0.441 between brain alignment and susceptibility to manipulation.
Early visual processing matters more than higher-level understanding: The protective effect only appeared in early visual cortex regions responsible for basic visual features (like edges and colors), not in brain regions that handle more complex visual concepts. This suggests that "seeing straight" at a fundamental level helps models resist pressure to agree with false claims.
The effect is strongest against existence denial attacks: When researchers tried to convince models that objects in images didn't exist, models with better early visual cortex alignment were most resistant, with a correlation of -0.597—suggesting that grounding in basic visual facts provides the strongest defense against this type of manipulation.
This finding applies across diverse model architectures: The relationship held consistently across 12 different vision-language models ranging from 256 million to 10 billion parameters, suggesting the principle is broadly applicable rather than specific to one model design.
Brain-inspired AI design may improve safety: The results suggest that aligning artificial visual systems with human neurobiology could be a practical approach to building more robust and trustworthy AI systems, particularly for high-stakes applications.

Abstract

28 pages, 9 figures, 13 tables

Gaslight, Gatekeep, V1-V3: Early Visual Cortex Alignment Shields Vision-Language Models from Sycophantic Manipulation

Key Takeaways

Abstract

Paper Info

Responsible AI Topics

Links

Gaslight, Gatekeep, V1-V3: Early Visual Cortex Alignment Shields Vision-Language Models from Sycophantic Manipulation

Key Takeaways

Abstract

Paper Info

Responsible AI Topics

Links