Baptiste Caramiaux

As Vision-Language Models are increasingly deployed in safety-critical applications, the trustworthiness of their explanations becomes crucial. Explainable AI (XAI) methods for Vision-Language Models often suffer from semantic hallucination, where attribution maps highlight prominent image regions even when prompted with incorrect text descriptions (e.g., highlighting a dog when prompted “cat”). Although this problem is widespread, a formal mathematical analysis of XAI methods and CLIP embeddings is largely missing in the literature. We demonstrate that this phenomenon is not specific to a single architecture but is a fundamental consequence of Linear Semantic Leakage in high-dimensional embedding spaces. We propose a unified theoretical framework, Linear Semantic Attribution (LSA), which generalizes across discriminative methods. We introduce OSP, a geometric intervention that utilizes the residual property of OMP to disentangle unique semantic signals from shared concepts. We prove theoretically and demonstrate empirically that OSP minimizes hallucination by orthogonalizing the query vector against distractor concepts, rendering the attribution model blind to shared features while preserving fidelity for correct prompts. Our code is available at: https://github.com/emirhanbilgic/Orthogonal-Semantic-Projection

Non-experts increasingly engage in user-driven algorithm auditing, interacting directly with AI systems to probe, document, and reflect on biased behavior. Yet, auditing remains challenging due to model opacity and limited support for navigating and interpreting outputs. This paper explores the design and evaluation of interfaces grounded in the sensemaking framework to support non-experts in auditing gender bias in image captioning. In a between-subjects study, 60 participants audited an image captioning model using one of three interface conditions: a Baseline interface, a Masking Tool for image manipulation, or a Filtering Tool for organizing captions. Our findings show that interface design shaped what participants noticed, how they interpreted model behavior, and supported their hypotheses. The Image Masking Tool enabled fine-grained testing of visual cues and context, while the Text Filtering Tool revealed broader asymmetries in gendered language. We argue that incorporating sensemaking into auditing practices can advance accountability and transparency in machine learning systems.

In the mid-2010s, media artists began developing practices using machine learning (ML) as an artistic medium. Since 2022, the rise of large generative models, the mainstreaming of AI as consumer products, and intensifying ethical disputes have reconfigured the conditions of their artistic practice. This paper aims to understand how artists working with ML over the past decade respond to these shifts, shedding light on how practices, tools, and culture co-evolve. We address this question through thematic analysis of semi-structured interviews with 30 artists active before 2020. Our findings show how artists experience narrowing aesthetics and reduced malleability of post-2020 ML systems, have diverging views on where to locate moral responsibility with large AI models, and face shifting cultural reception that challenges the legibility of their work. We map how artists envision their practice going forward and discuss those orientations with respect to HCI conversations on design and creativity.