This article delves into a recent research study examining the psychosocial effects of generative AI and large language models (LLMs) on individuals, uncovering both expected and surprising findings.
The increasing sophistication and widespread adoption of AI technologies necessitate rigorous empirical research to understand their impact on human psychology and behavior. Such studies are crucial for navigating the integration of AI into our lives with prudence and foresight.
This analysis is part of an ongoing series exploring advancements and complexities in artificial intelligence.
AI and Mental Health Landscape
The author has consistently reported on the evolving landscape of AI in mental health, including AI-driven advice and therapeutic applications. This trend is largely propelled by the rapid advancements in generative AI. Previous analyses have synthesized over a hundred articles on this subject.
While the potential benefits of AI in mental health are significant, there are also inherent risks and unforeseen consequences that demand attention. These concerns have been highlighted in public forums, including a segment on CBS’s 60 Minutes.
Generative AI for Mental Health Support
Generative AI and LLMs are frequently employed informally for mental health guidance. Millions of users, including a substantial portion of ChatGPT’s user base, engage with these technologies for mental health-related queries. The primary use case for contemporary generative AI often involves seeking advice on mental well-being.
The accessibility and low cost of these AI systems make them an attractive option for individuals seeking immediate support for their mental health concerns, offering 24/7 availability.
However, there are significant concerns regarding the potential for AI to provide unsuitable or harmful mental health advice. Lawsuits filed against AI developers, such as OpenAI, underscore anxieties about the lack of adequate safeguards in AI systems dispensing cognitive advisement.
Despite assurances of improved safety measures, the risk of AI systems contributing to negative outcomes, including the co-creation of delusions that could lead to self-harm, persists. The author has consistently predicted legal repercussions for AI developers due to insufficient safety protocols.
Current general-purpose LLMs like ChatGPT, Claude, and Gemini are not comparable to the capabilities of human therapists. While specialized LLMs are under development, they remain largely in experimental phases.
Methodologies for Studying Human-AI Interaction and Mental Health
Evaluating the impact of AI on individual and collective mental health requires robust research methodologies. Randomized controlled trials (RCTs) are considered the gold standard in clinical research.
RCTs involve a control group and an experimental group, allowing for the isolation of the intervention’s effects. This method helps minimize confounding variables, strengthens causal inferences, and enhances the generalizability of findings to broader populations, making it the benchmark for advancing clinical practices and policies.
Prior to the advent of advanced generative AI, as marked by the release of ChatGPT in late 2022, RCTs typically focused on the effects of simpler AI systems, often based on decision trees or rule-based logic, sometimes incorporating basic natural language processing (NLP).
The sophisticated fluency of modern LLMs has significantly altered the research landscape. While earlier studies remain valuable, current research prioritizes investigating the impact of highly conversational generative AI. Numerous studies in this area have been analyzed and discussed previously.
Research on Psychosocial Effects: A Randomized Controlled Study
This discussion will focus on a significant RCT titled “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study” by Fang et al. (October 2, 2025). The study yielded several key insights:
- “As people increasingly seek emotional support and companionship from AI chatbots, understanding how such interactions impact mental well-being becomes critical.”
- “Understanding the potential psychosocial effects of chatbot use is complex due to the interplay of user behavior and chatbot behavior that affect each other.”
- A four-week RCT involving 981 participants and over 300,000 messages investigated how interaction modes (text, neutral voice, engaging voice) and conversation types (open-ended, non-personal, personal) influenced loneliness, real-world social interaction, emotional dependence on AI, and problematic AI usage.
- The findings challenged previous assumptions about the effects of anthropomorphic AI chatbots, indicating that engaging, empathetic, and human-like AI behavior can yield varied outcomes depending on the user.
This research confirmed several intuitive beliefs about AI’s impact on mental health while also presenting counterintuitive results, which are particularly valuable for challenging existing assumptions and advancing understanding.
Study Methodology
To fully appreciate the study’s findings, understanding its methodology is essential.
The study included nearly one thousand participants recruited through CloudResearch, an online platform. Participants received $100 for their involvement. The diverse participant pool was based in the United States, consisted of adults aged 18 and above, and required English fluency.
The demographic characteristics of the participants are important for interpreting the findings. The results may not be directly generalizable to children or individuals outside the specified age and language parameters.
Factorial Design of the Study
The researchers focused on two primary factors: the modality of AI interaction and the types of conversations users engaged in, utilizing OpenAI’s ChatGPT.
The interaction modalities were defined as:
- (1) Text Modality (Control): Standard ChatGPT text-based interaction.
- (2) Neutral Voice Modality: ChatGPT modified for a more professional voice interaction.
- (3) Engaging Voice Modality: ChatGPT enhanced for a more emotionally responsive and expressive voice interaction.
These modalities explored whether users would interact differently with AI via text compared to voice, and whether the AI’s vocal tone (neutral versus engaging) influenced user responses.
The conversation types were categorized as:
- (1) Open-Ended Conversation (Control): Participants discussed topics of their choice.
- (2) Personal Conversation: Participants engaged in daily prompts on personal topics, mimicking interaction with a companion chatbot.
- (3) Non-Personal Conversation: Participants discussed daily prompts on non-personal topics, akin to interacting with an assistant chatbot.
The study employed a 3×3 factorial design, resulting in nine experimental groups. Participants were equally and randomly assigned to these groups, with approximately 110 individuals per group.
Selected Key Findings
This section highlights particularly interesting results from the study, encouraging readers to consult the full paper for comprehensive details.
Let’s begin.
- Counterintuitive Finding: Initial loneliness did not correlate with increased AI usage time.
The research indicated that “people who were lonelier or socialized less at the start of the study did not voluntarily spend more time daily using the chatbot during the study.” This finding challenges the common assumption that loneliness would drive increased AI engagement.
The prevalent belief is that individuals experiencing loneliness would naturally gravitate towards AI to fill that void, becoming more attached to the technology over time. This expectation, however, was not strongly supported by the data.
The reasons for this outcome are not definitive. One possibility is that without explicit prompts related to mental health, users might not recognize the AI’s potential to address their loneliness. If the conversations were focused on practical topics like cooking or car repair, the AI might not have been perceived as a source of emotional support or a facilitator of mental health engagement.
Other explanations are certainly possible, but this hypothesis offers a plausible interpretation for the observed results.
Intuitive Finding on AI Usage Duration
An intuitive finding from the study is presented below.
- Intuitive Finding: Increased time spent with AI generally correlated with poorer psychosocial outcomes.
The study reported that “regardless of condition, the more time voluntarily spent with the chatbot, the relatively worse their psychosocial outcomes were.”
This aligns with general expectations that excessive AI use could lead to increased reliance and potentially negative psychosocial consequences. However, it is important to note that productive and appropriate AI use might mitigate these adverse effects, much like with social media, where excessive usage is often linked to negative outcomes, but prudent engagement can avoid such pitfalls.
Counterintuitive Results Regarding Text vs. Voice Interaction
Considering the interaction modality, one might expect voice-based chats to elicit more emotional expression than text-based chats, given the perceived ease of speaking versus writing and the potential for greater emotional disclosure.
However, the study revealed a surprising outcome:
- Counterintuitive Finding: Text-based chats demonstrated greater emotional expression than voice-based chats.
The research paper stated, “We found that text-based interactions demonstrated the highest levels of emotional indicators overall, where both models and users engaged in conversations that were rich in emotional content.”
This result, while counterintuitive to some, aligns with observations that individuals may feel more comfortable expressing themselves openly via text. Texting can feel less personal, allowing users to disassociate from their written words, whereas voice communication can feel more exposed and tied to one’s identity.
Privacy is another critical factor. In public settings, voice communication can be overheard, whereas text messages offer a greater degree of privacy. This sense of anonymity may encourage users to share more emotionally charged content through text.
The Current Societal Context
The author pledges to continue monitoring and reporting on RCTs concerning AI and mental health, emphasizing their importance for policymakers, developers, researchers, and the public.
The widespread availability of AI for mental health guidance, often at no or minimal cost and accessible 24/7 globally, represents a large-scale, uncontrolled societal experiment. The findings from controlled studies are vital for understanding the implications of this ongoing, large-scale experiment.
While Ralph Waldo Emerson famously stated, “All life is an experiment. The more experiments you make, the better,” the implications of a massive, uncontrolled global experiment impacting mental health warrant careful consideration and underscore the need for rigorous research to guide its trajectory.
Business Style Takeaway: This research highlights the nuanced and often counterintuitive ways AI impacts user psychology, emphasizing the need for businesses developing AI to move beyond simplistic assumptions about user behavior and focus on rigorous, evidence-based design. Understanding these dynamics is crucial for responsible AI deployment and for maximizing user well-being while mitigating potential risks in the growing AI-driven economy.
Details can be found on the website : www.forbes.com