Are the mental health risks of AI companions being underestimated?

What are the promised benefits of AI companions, and what hidden risks do experts see?

AI companions, often called 'Social AI,' are promoted as a solution to overburdened healthcare systems and widespread loneliness [1]. They offer conversation-based emotional support that is accessible and low-cost, potentially reaching people who face stigma or high costs in traditional mental health care [2][3]. However, a 2025 expert workshop convened at Cambridge University [1] concluded that these same systems carry serious risks of user manipulation and the unintended development of emotional dependencies—risks that could actually worsen the very problems of isolation and poor mental health they are meant to solve. The experts warned that these dangers are often underestimated by both developers and policymakers.

Do the studies agree that dependency on AI companions is a real danger?

Yes, all three papers converge on the same core warning: emotional dependency is a genuine and serious risk. The 2025 workshop report [1] explicitly states that Social AI may lead to 'unintended development of dependencies,' potentially exacerbating mental health issues. A 2023 review [2] echoes this, calling for 'responsible regulation' to minimize potential harms, including the risk that users become overly reliant on AI for emotional support. A 2021 review [3] adds that while chatbots can be used alongside formal treatments, they are 'not a substitute'—implying that over-reliance could lead people to delay or forgo professional care. The agreement across these sources, from a 2021 academic review to a 2025 expert consensus, strengthens the conclusion that dependency is not a theoretical worry but a documented risk.

Where do the studies conflict or highlight gaps in our understanding?

The papers do not directly contradict each other, but they highlight different aspects of the risk, which reveals a gap in the evidence. The 2025 workshop [1] focuses on policy-level risks like manipulation and systemic dependency, while the 2021 review [3] emphasizes practical concerns like privacy, data security, and AI limitations. The 2023 review [2] bridges these by calling for 'Human–Artificial Intelligence' (HAI) principles to integrate human values and empathy into AI design. The key gap is that none of the papers provide long-term, quantitative data on how many users actually develop harmful dependencies or how severe the consequences are. The risks are identified and agreed upon, but their real-world scale remains uncertain—which is itself a reason to be cautious.

About These Sources

This answer is built on 3 studies (2 peer-reviewed, 1 preprint) — published from 2021 to 2025, 1 from 2024 or later, collectively cited 112 times — selected as the most relevant from 3 studies that passed quality screening, drawn from 38 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions