Articles and Information

Literature

1) El-Sayed, S., et al. (2024). A Mechanism-Based Approach to Mitigating Harms from Persuasive Generative AI. Google DeepMind, arXiv.

Summary: This Google DeepMind paper establishes a systematic framework for studying and mitigating the harms associated with persuasive generative AI systems, distinguishing between rational persuasion (based on evidence) and manipulation (exploiting cognitive biases). To address the growing risks, the authors shift the focus from mitigating harmful outcomes to mitigating harmful processes—specifically by mapping the underlying mechanisms of manipulation that erode a user's cognitive autonomy and decision-making integrity

2) Rousmaniere, T., Zhang, Y., Li, X., & Shah, S. (2025). Large Language Models as Mental Health Resources: Patterns of Use in the United States. Practice Innovations.

Summary: A survey of U.S. residents with mental health conditions found a substantial rate of adoption, with nearly half of all users turning to general-purpose Large Language Models (LLMs) for support with issues like anxiety, depression, and personal advice. The majority of these users reported that the LLMs improved their mental health, and more than a third found the AI support more beneficial than traditional human therapy, despite a small percentage encountering harmful responses.

3) Zhang, Y., Li, X., Zhu, J., Sheng, Z., & Rousmaniere, T. (2025). What happens, what helps, what hurts: A qualitative analysis of user experiences with large language models for mental health support. Unpublished manuscript.

Summary: This is a qualitative study on 243 adults with mental health concerns using LLMs for emotional support. The research highlights significant limitations and risks, including the LLMs providing non-actionable, overly generic, or even risk-inducing advice, underscoring the need for clinical safeguards despite their perceived utility.