Investigating Associational Biases in Inter-Model Communication of Large Generative Models

TL;DR

Associational biases can propagate and amplify in inter-model AI pipelines, particularly in human activity and affective perception. The authors introduce an iterative inter-model loop (text-to-image and image-to-text) and an explainability pipeline (token-conditioned Grad-CAM with region-level attribution) to quantify demographic drift and its grounding cues using PHASE and RAF-DB data. They find consistent drifts toward younger and more female-presenting representations, with region-based analyses revealing reliance on spurious cues like background and hair. The work also links these drifts to downstream prediction differences and outlines data-, training-, and deployment-time mitigation strategies, emphasizing safeguards and ongoing bias monitoring in interconnected AI systems.

Abstract

Social bias in generative AI can manifest not only as performance disparities but also as associational bias, whereby models learn and reproduce stereotypical associations between concepts and demographic groups, even in the absence of explicit demographic information (e.g., associating doctors with men). These associations can persist, propagate, and potentially amplify across repeated exchanges in inter-model communication pipelines, where one generative model's output becomes another's input. This is especially salient for human-centred perception tasks, such as human activity recognition and affect prediction, where inferences about behaviour and internal states can lead to errors or stereotypical associations that propagate into unequal treatment. In this work, focusing on human activity and affective expression, we study how such associations evolve within an inter-model communication pipeline that alternates between image generation and image description. Using the RAF-DB and PHASE datasets, we quantify demographic distribution drift induced by model-to-model information exchange and assess whether these drifts are systematic using an explainability pipeline. Our results reveal demographic drifts toward younger representations for both actions and emotions, as well as toward more female-presenting representations, primarily for emotions. We further find evidence that some predictions are supported by spurious visual regions (e.g., background or hair) rather than concept-relevant cues (e.g., body or face). We also examine whether these demographic drifts translate into measurable differences in downstream behaviour, i.e., while predicting activity and emotion labels. Finally, we outline mitigation strategies spanning data-centric, training and deployment interventions, and emphasise the need for careful safeguards when deploying interconnected models in human-centred AI systems.

Figures (8)

• Figure 1: An illustrative example demonstrating how associational biases can arise through the inter-model communication of large generative models.
• Figure 2: Overview of inter-model communication and explainability pipelines. (a) The inter-model communication pipeline runs through image generation/image description loops. (b) The explanability pipeline uses image descriptions obtained from the LLaVa model and leverages Grad-CAM to output regional activation.
• Figure 3: Original (O) and collected (C) distributions for PHASE activities before and after the loop.
• Figure 5: Original (O) and collected (C) distributions for PHASE emotions before and after the loop.
• Figure 7: Original (O) and collected (C) distributions for RAF-DB before and after.