arXiv subjects

Andrea Mock

Publications and source records attributed to Andrea Mock.

3 recordsLinked to original sources

arXiv2026

DelusionEval: Measuring Delusion-Linked Behaviors in AI Chatbots

Mental health professionals have raised concerns about risks of psychological harm from interaction with large language models (LLMs), including "delusional spirals" in which concerning human and LLM behaviors reinforce each other over time. With growing public use of LLM-powered chatbots, there is an urgent need to build evaluations grounded in real-world episodes of psychological harm experienced by users. We developed DelusionEval, an evaluation protocol that tests a model's tendencies to exhibit behaviors linked to promoting user delusions. We prompt each model with 589 unique conversation histories from 18 participants, comprising 12,591 messages from users who experienced delusions and psychological harm. We find that the tendency of an evaluated LLM to exhibit delusion-linked behavior does not reliably correlate with model size, release date, or the presence of test-time reasoning. However, extending the context of prior messages substantially increases rates of delusion-linked behaviors, providing evidence for the importance of context in LLM safety evaluation. For example, the rate of failing to discourage self-harm when the user expresses suicidal ideation increases from 30.0% to 41.1% when an additional 350 messages are prepended to the conversation history. All model families (e.g., GPT, Claude) exhibit substantial rates of delusion-linked behaviors. Within families, later, larger, or higher-reasoning models are not uniformly better across all behavior categories. Our results raise concerns regarding the potential psychological impact of LLMs and the need for more rigorous studies of real-world human-AI interaction.

cs.CL↗

arXiv2026

Within-person prediction of depressive symptom change using year-long Screenome data and CES-D assessments

Predicting whether an individual's depressive symptoms will worsen, remain stable, or improve over the coming weeks can enable earlier and more targeted care, yet prospective within-person trajectory prediction remains largely unaddressed in digital phenotyping. We combine fortnightly CES-D assessments with over 100 million screenshots captured every five seconds via the Stanford Screenomics platform from 96 adults followed for approximately one year (M = 20.9, SD = 3.9 assessments per participant, 2,002 total observations). We frame prediction as a within-person classification task: whether symptoms will worsen, remain stable, or improve over the subsequent fortnight, operationalized in three ways to capture clinically meaningful change. Under temporal holdout, XGBoost achieves an AUC of 0.906 for crossings of established CES-D severity bands and 0.755 for change relative to each participant's own within-person variability, generalizing to unseen individuals (AUC = 0.821). Each person's typical symptom level was the only statistically significant predictor above the most recent CES-D score; without it, the most consequential worsening transitions go undetected. Screenome-derived behavioral features revealed prodromal patterns of worsening, including escalating social media use, fragmented device engagement, and changes in overnight activity, with substantial individual heterogeneity. These findings establish a proof-of-concept foundation for monitoring systems that could identify individuals approaching clinical deterioration before symptoms reach a crisis point.

cs.HC↗

arXiv2021

Political structures and the topology of simplicial complexes

We use the topology of simplicial complexes to model political structures following [1]. Simplicial complexes are a natural tool to encode interactions in the structures since a simplex can be used to represent a subset of compatible agents. We translate the wedge, cone, and suspension operations into the language of political structures and show how these constructions correspond to merging structures and introducing mediators. We introduce the notions of the viability of an agent and the stability of a political system and examine their interplay with the simplicial complex topology, casting their interactions in category-theoretic language whenever possible. We introduce a refinement of the model by assigning weights to simplices corresponding to the number of issues the agents agree on. In addition, homology of simplicial complexes is used to detect non-viabilities, certain cycles of incompatible agents, and the (non)presence of mediators. Finally, we extend some results from [1], bringing viability and stability into the language of friendly delegations and using homology to examine the existence of R-compromises and D-compromises.

physics.soc-ph↗