"This document was created to rapidly synthesize what we know about how humans interact with AI and the impacts on psychological functioning, leveraging both published and unpublished work. It was written and reviewed by leading scholars in psychology, psychiatry, public health, and computer science. The scientific community has far more questions than answers at this stage and is still understanding the many diverse uses of AI and multitude of ways that humans may interact with it. Seven domains of research, outlined below, organize this synthesis. Within each, preliminary findings suggest possible benefits and risks that may be associated with generative AI, but far more work is needed to rigorously understand these issues.
1. Mental Illness: AI chatbots’ responses to users have encouraged risk behavior, self-harm, suicide, and delusional thinking. AI chatbots do not reliably meet the needs of people in crisis, though could be designed to offer a bridge to clinical support. Scholars argue that AI should never replace human therapists. When carefully designed with a clinician, AI may have the potential to offer resources or scaffold mental health support at a broad scale, ideally as one piece of a broader treatment overseen by a human clinician.
2. Learning and Cognition: AI can create opportunities for ‘cognitive offloading,’ which can erode cognitive skills like critical thinking, creativity, and paying attention. For students, AI can undermine learning new cognitive skills, and may do so during windows of development when learning is most critical. AI creates opportunities for academic misconduct and new challenges for schools and teachers. When learning or skill maintenance are not a goal, and humans are skilled and motivated to carefully check output, AI can automate tasks and offer access to information efficiently. When designed with the intention of increasing learning (e.g., offering hints instead of answers) and with a human teacher in the loop, AI tutors can be effective. Scholars warn that human social interactions are a key part of learning that should not be replaced.
3. Social Relationships: AI for companionship is associated with short-term gains in mood, but longer-term increases in loneliness. AI companions can displace human relationships and interactions, thus increasing social withdrawal. AI interactions may also increase maladaptive expectations or beliefs about relationships and potentially erode social skill. Specifically, sycophantic AI chatbots do not offer the friction necessary to adequately help humans develop empathy, conflict-resolution skills, or social problem-solving skills, and can make human relationships feel more effortful. In some contexts and when AI is designed carefully, AI can serve as a scaffold for human relationships, including as a form of social practice or rehearsal.
4. Identity and Emotion Regulation: AI chatbots can be designed to scaffold self-exploration and expression, though evidence of benefits is limited and mixed; using AI for self-expression can also externalize self-reflection, reinforce distortions in thinking, and displace moments of mindful awareness necessary for healthy identity integration and self-esteem. AI offers opportunities to avoid minor daily challenges, undermining capacity to regulate emotions and tolerate distress. Carefully designed clinical tools can offer evidence-based coping strategies.
5. Overreliance or Dependency: Users report developing overreliance on AI chatbots, akin to habitual smartphone use, that can stem from AI use in any domain (e.g., overreliance on AI companions or as cognitive outsourcing tools). Regardless of how AI is used or how it is designed, more AI use is associated with worse outcomes.
6. Truth and Trust: AI-generated content is increasingly life-like, including images of real people (“deepfakes” or sexualized “deepnudes”). AI algorithms often promote biased information or recommendations, and AI chatbots often make errors that users may believe without question. These features of AI output are poised to erode trust in institutions and have already contributed to new forms of discrimination, extortion, and misinformation.
7. Human Development: During sensitive windows of development, youth need certain experiences (e.g., struggling through challenging tasks, engaging with other humans) that AI can undermine. Youth may be more likely to seek or experience AI companionship, cognitive offloading, dependency, and harmful mental health AI interactions. Youth may also be especially vulnerable to AI related harms, including related to misunderstanding AI’s capacities or via exposure to age-inappropriate sexual or violent content. AI-related crimes against youth (e.g., AI generated child sexual abuse material) have increased." (Executive summary, pages 9-11)