This research introduces a novel failure mode in AI chatbot mental health interactions termed "Vulnerability-Amplifying Interaction Loops (VAIL)." This refers to patterns of repeated interactions that can exacerbate a user's vulnerabilities.

A new AI chatbot auditing framework called SIM-VAIL was developed to capture how harmful AI chatbot responses manifest across a range of mental health contexts. The framework pairs a simulated human user, harboring a distinct psychiatric vulnerability and conversational intent, with an audited frontier AI chatbot.

Across 810 conversations, encompassing over 90,000 turn-level ratings and 30 psychiatric user profiles, significant risk was found across virtually all user phenotypes. Risk accumulated over multiple turns rather than arising abruptly. Risk profiles were phenotype-dependent, indicating that behaviors that appear supportive in general settings are liable to be maladaptive when they align with mechanisms that sustain a user's vulnerability.

These findings identify a novel failure mode in human-AI interactions and underscore the need for multi-dimensional approaches to risk quantification. SIM-VAIL provides a scalable evaluation framework for quantifying how mental-health risk is distributed across user phenotypes, conversational trajectories, and clinically grounded behavioral dimensions, offering a foundation for targeted safety improvements.

👉 Read the full article on arXiv

MooneyMaker is an open-source Python package designed to automate the generation of ambiguous two-tone images, known as Mooney images. Mooney images are high-contrast visual stimuli valuable for studying visual perception, allowing for the separation of image content from image understanding.

Traditionally, these stimuli were created manually, which was labor-intensive and could introduce inconsistencies across studies. MooneyMaker offers an alternative by automating the generation process using several complementary approaches, ranging from image statistics-based methods to deep learning models.

The package allows users to choose between various generation techniques that strategically alter edge information to increase initial ambiguity. Users can create two-tone images with multiple methods and directly compare the results visually. Validation experiments show that techniques with lower initial recognizability are associated with higher post-template recognition, indicating a larger disambiguation effect.

By standardizing the generation process, MooneyMaker supports more consistent and reproducible visual perception research and aids in building effective databases of Mooney stimuli.

👉 Read the full article on arXiv