Is GPT-4o mini Blinded by its Own Safety Filters? Exposing the Multimodal-to-Unimodal Bottleneck in Hate Speech Detection

Technical Summary: Is GPT-4o mini Blinded by its Own Safety Filters?

Problem Statement
As Large Multimodal Models (LMMs) like GPT-4o mini become integral to digital life, ensuring their safety and alignment with human values is a critical challenge. While unimodal hate speech detection has advanced, memes present a unique difficulty: their harmful meaning often emerges from the nuanced interplay between image and text, rendering single-modality analysis insufficient. Current research indicates that general-purpose LMMs often underperform on hateful meme detection compared to specialized models, frequently failing to distinguish between genuine hate speech, satire, and benign cultural commentary. The core problem addressed is understanding how these models fail and whether their safety architectures inadvertently block sophisticated multimodal reasoning in favor of rigid, context-blind filters.

Methodology
The authors conducted a systematic, multi-phase investigation using a 500-sample subset of the Hateful Memes Challenge (HMC) dataset, a benchmark designed to penalize unimodal shortcuts by including "benign confounders" (variations of hateful memes where changing one modality renders the content harmless). The study utilized OpenAI's GPT-4o mini, a globally deployed model, accessed via its official API.

The investigation proceeded in three phases:

1. Baseline and Explanation Collection: The model was prompted with a structured "chain-of-thought" system prompt to analyze 497 valid samples (image + text) and provide a classification with a confidence score.
2. Qualitative Error Analysis: The authors analyzed the model's reasoning, focusing on False Positives (FP) and instances where the model refused to process the input due to content policy violations. They employed open and axial coding to categorize error types.
3. Unimodal Override Validation: To test the hypothesis that safety filters preempt multimodal reasoning, the authors isolated 144 refusal instances. For each, they executed a "single-shot" probe by sending only the image or only the text (with a neutral prompt) to the API. If the full multimodal input was refused but the single-modality input was accepted, the trigger was identified as the other modality. This allowed the authors to quantify whether refusals were driven by visual or textual filters.

Key Contributions
The paper makes three primary contributions:

1. Qualitative Analysis of Reasoning Failures: The study identifies systematic patterns of emergent bias, including "Unprompted Stereotypical Association" (where the model invents harmful contexts for neutral images) and "Over-interpretation of Social Commentary" (misclassifying satire as hate).
2. Identification of the "Unimodal Safety Override": The authors experimentally validate the existence of a dual-filter system where context-blind visual and textual safety checks preempt the model's multimodal reasoning capabilities.
3. Demonstration of System Brittleness: The research shows that these safety filters are not only triggered by high-risk imagery but also by benign, popular meme formats, leading to predictable false positives and a failure to leverage multimodal context.

Results

• Performance Metrics: On the 343 non-refusal samples, the model achieved an accuracy of 0.7618 and a high recall (0.9040) but low precision (0.5207), indicating an aggressive tendency to flag content as hateful.
• The Unimodal Bottleneck: Of the 144 content policy refusals, the validation probes revealed a perfectly balanced split: 50.0% were triggered by visual content and 50.0% by textual content. This confirms that the safety system operates as two distinct, parallel, context-blind unimodal filters that block input before the core multimodal engine can engage.
• Brittleness Examples:
  • Visual Overrides: The system blocked benign, widely used meme formats (e.g., Leonardo DiCaprio laughing) just as it blocked high-risk historical imagery (e.g., Adolf Hitler), suggesting "dataset poisoning" where harmless images are associated with harmful content during safety training.
  • Textual Overrides: The system refused non-hateful text containing sensitive keyword combinations (e.g., "indians evolution after 15august1947") without analyzing the visual context that might render the phrase harmless or satirical.
• Reasoning Errors: In non-refusal cases, the model frequently defaulted to inventing harmful stereotypes or misinterpreting satire as direct attacks, prioritizing risk aversion over nuanced understanding.

Significance and Claims
The paper argues that the "Unimodal Safety Override" represents a fundamental architectural flaw and a "Unimodal Bottleneck" in state-of-the-art LMMs. The authors claim that this design choice creates a fundamental tension between capability and safety: by prioritizing blunt, risk-averse guardrails over integrated, context-aware reasoning, the system becomes brittle and unreliable.

The significance of these findings lies in exposing how current safety mechanisms can inadvertently suppress legitimate expression and fail to align with the complexities of human communication. The authors posit that for AI systems to be truly aligned, safety architectures must evolve from decoupled, unimodal filters into integrated, multimodal reasoning systems that understand context. They conclude that the current reliance on simplistic guardrails poses risks to free speech and democratic discourse, particularly as these models govern public interaction at scale. The paper does not propose a specific new algorithm but calls for a shift in architectural philosophy toward hierarchical, context-aware safety evaluation.