Generative artificial intelligence (AI) use and depressive symptoms: why interpretation requires caution

The recent study by Perlis and colleagues examined the association between generative artificial intelligence (AI) use and depressive symptoms in a large, nationally representative sample of United States adults (1). Using survey data from 20,847 participants, the authors reported that more frequent AI use was associated with modest but statistically significant increases in depressive symptoms, anxiety, and irritability. Notably, the association appeared strongest among individuals reporting personal use of AI, and age-stratified analyses suggested that the relationship was more evident among adults aged 25–44 and 45–64 years than among the youngest or oldest groups. These findings are important because they move the discussion of AI and mental health beyond anecdote and toward population-level evidence. At the same time, several conceptual and methodological issues deserve closer consideration before these results are interpreted as indicating that AI use itself worsens mental health.

First, the cross-sectional design does not allow the direction of association to be established. This limitation is acknowledged by the authors, yet it is particularly important in the context of personal AI use. An equally plausible explanation is that individuals with greater baseline depressive symptoms, loneliness, social withdrawal, or reduced confidence in interpersonal interaction may have been more likely to use AI for personal purposes in the first place. In such cases, AI may function not as the source of depressive symptoms, but as a low-demand substitute for human feedback, advice, or companionship (2,3). Individuals who are already distressed may turn to AI because it is always available, nonjudgmental, and emotionally undemanding. The observed association may therefore reflect reverse causality or selection effects as much as any potential negative effect of AI exposure. This issue becomes especially salient because the study found a significant association for personal use, but not for work- or school-related use, suggesting that the social and emotional context of engagement may matter at least as much as the technology itself.

Second, the finding that personal use was associated with depressive symptoms invites more careful interpretation of what “personal use” actually represents. This category is likely heterogeneous, potentially including casual curiosity, emotional support seeking, advice seeking, entertainment, or prolonged conversational engagement. These uses are psychologically distinct and may not carry the same implications for well-being. Without more detailed information on why participants used AI, how long they interacted with it, and whether their use replaced or supplemented human interaction, it is difficult to determine whether AI served as an emotional crutch, a coping aid, a productivity tool in informal settings, or simply a conversational novelty. Future studies should distinguish between instrumental use and relational or affective use, because the latter may be more closely tied to mood, loneliness, and patterns of social withdrawal. This distinction is not merely descriptive. Personal AI use may include at least three psychologically distinct forms: instrumental use, in which AI is used to solve practical problems or obtain information; advisor-like use, in which AI is consulted for guidance, reassurance, or interpretation of personal problems; and companion-like use, in which users return to AI for conversation, emotional validation, or a sense of social presence. These forms of engagement are likely to differ in their mental health implications. A testable prediction is that companion-like and emotionally validating AI use will show stronger associations with depressive symptoms, loneliness, and interpersonal withdrawal than instrumental use, particularly when AI interaction substitutes for rather than supplements human contact. This interpretation is also consistent with emerging empirical evidence that many users, including individuals with mental health conditions, turn to large language models (LLMs) for psychological support, emotional processing, or therapy-like purposes, suggesting that “personal use” may increasingly include quasi-therapeutic forms of engagement rather than only casual or informational use.

Third, the research identifies meaningful age differences, but these findings are not discussed in sufficient depth. Importantly, the authors did not merely list age groups descriptively; they tested age as a moderator and found that the association between AI use and depressive symptoms differed by age group, with significant associations observed among those aged 25–44 and 45–64 years, but not among those aged 18–24 or 65 years and older. This is a notable result because it suggests that generative AI should not be treated as a uniform exposure across the life course. Yet the paper provides limited discussion of why these age differences might emerge. One possible interpretation is that middle adulthood may represent a particularly consequential stage for AI’s emotional and social effects. Adults in their midlife years may experience a combination of work stress, caregiving burden, family transitions, marital strain, or social isolation (4), and these pressures may shape how AI is used and experienced. In some individuals, personal AI use may become a convenient alternative to emotionally effortful human interaction. In others, it may provide a form of companionship or immediate feedback during periods of stress and disconnection. The same technology may therefore serve very different functions depending on life stage, social role, and emotional need. This is precisely why age should not be treated simply as a control variable, but as a meaningful modifier of both motivation for AI use and its potential psychological consequences.

At the same time, the absence of a statistically significant association in the oldest group should not be taken to mean that AI is irrelevant to later-life mental health. On the contrary, older adults may engage with AI in qualitatively different ways. For some individuals, particularly those experiencing bereavement, social isolation after retirement, limited mobility, or separation from family, conversational AI may function not primarily as a risk factor but rather as a compensatory source of social connection (3,5). It may function as a lightweight social presence, offering responsiveness without the demands of conventional relationships. That possibility is also consistent with the broader literature cited by the authors, including evidence that some AI systems specifically designed for mental health applications may have beneficial effects. The key point is that AI use may be harmful, neutral, or supportive depending on who is using it, why it is being used, and what social resources are otherwise available.

The null finding in the youngest age group is also intriguing. One might have expected younger adults, who are more digitally immersed and more accustomed to mediated interaction, to show the strongest negative association if AI use were simply eroding social capacity. Yet that is not what this study found. This suggests that digital familiarity alone is insufficient to explain the relationship between AI use and depressive symptoms. Younger adults may use AI more fluidly as one tool among many, whereas midlife adults may be using it under different emotional or situational pressures. Alternatively, the relevant mechanism may not be age per se, but differences in coping style, social role demands, loneliness, or perceived availability of human support across age strata. Future studies should therefore move beyond age categories alone and examine age-linked mechanisms, including loneliness, interpersonal avoidance, attachment to AI, unmet emotional needs, and perceived social support.

The study therefore raises a broader conceptual issue: generative AI should not be understood as a single exposure analogous to screen time. This point applies not only to purposes of use, but also to the heterogeneity of AI systems themselves. A general measure of generative AI use may collapse across tools that differ substantially in interaction style, design goals, optimization targets, and relational expectations. General-purpose LLMs such as ChatGPT, Claude, and Gemini may be used for information seeking, writing, problem solving, or personal advice, whereas dedicated companion chatbots are often designed to sustain emotionally engaging and relationally framed interaction. These differences matter because the psychological meaning of AI use may depend not only on how often people use AI, but also on what kind of system they use and what type of relationship the system invites. Generative AI is therefore better conceptualized as a family of socially responsive technologies whose psychological meaning depends on both the system being used and the context of use. This framing is consistent with several established perspectives in human–machine interaction. Uses and gratifications theory suggests that the effects of media technologies depend on the motives users bring to them, such as information seeking, emotional regulation, entertainment, or companionship. Parasocial interaction theory further helps explain why repeated interaction with a responsive nonhuman agent may come to feel socially meaningful, even when the relationship is not reciprocal in a human sense. Similarly, the Computers Are Social Actors (CASA) paradigm suggests that people often apply social rules and expectations to computers and other interactive systems. Together, these perspectives imply that generative AI should be examined not only as a digital tool, but also as an interactive system that can invite social interpretation, emotional reliance, and perceived companionship. A person who uses AI to draft emails at work is engaging with a fundamentally different phenomenon from a person who repeatedly turns to AI for advice, reassurance, conversation, or emotional validation. The present findings are valuable precisely because they begin to signal that not all AI use is alike. The significant result for personal use is therefore not a minor subgroup observation, but one of the most conceptually important findings in the paper.

Future research should build on this contribution in several ways. Longitudinal studies are needed to clarify temporal ordering and disentangle whether AI use precedes depressive symptoms, follows them, or both. Studies should distinguish among age groups not only descriptively, but theoretically, examining whether AI functions differently for young adults, middle-aged adults, and older adults. Future research should also examine whether these patterns vary across cultural contexts. The study by Perlis et al. was based on a US sample, and the meaning of personal AI use may differ in societies with different norms regarding emotional disclosure, family support, help-seeking, technology adoption, and access to mental health services. In some contexts, AI may be used mainly as an informational or productivity tool, whereas in others it may become a more acceptable or accessible source of advice, emotional support, or companion-like interaction. More assessments of personal use are also needed, including emotional motives for use, duration and intensity of conversational engagement, whether AI replaces or supplements human contact, and whether users perceive AI as a tool, an advisor, or a companion. Future longitudinal studies could test whether these subtypes have different psychological trajectories: instrumental use may be weakly or inconsistently related to depressive symptoms, whereas advisor-like or companion-like use may be more strongly associated with depressive symptoms when it reflects unmet emotional needs, loneliness, or substitution for human support. Furthermore, future studies should examine full AI conversation data, such as chat logs donated by consenting study participants, because these records may reveal interaction patterns, emotional tone, advice-seeking, reassurance-seeking, and companion-like engagement that cannot be captured by frequency measures alone. Such data are currently difficult for researchers without affiliation to chatbot developers to access, but they could become valuable shared research resources if collected with informed consent and released only after robust anonymization procedures. Finally, intervention studies should test not only whether AI can improve productivity or access to information, but also whether certain forms of use alleviate loneliness in some groups while exacerbating withdrawal or depressive cognition in others.

In that sense, the most important contribution of this study may not be the conclusion that more frequent AI use is associated with greater depressive symptoms, but rather the recognition that this association is unlikely to be simple, uniform, or universally harmful. The findings should prompt a more nuanced agenda that considers the purpose of use, reverse causality, and life-stage heterogeneity. As generative AI becomes increasingly embedded in everyday life, the more important question is not just whether AI affects mental health, but who is affected, in what situations, and by what psychological and social processes.

Acknowledgments

During the preparation of this work, a generative AI tool was only used for language editing and refinement. The tool was not used to generate the intellectual content, arguments, or conclusions.

Footnote

Provenance and Peer Review: This article was commissioned by the editorial office, Journal of Medical Artificial Intelligence. The article has undergone external peer review.

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0070/prf

Funding: None.

Conflicts of Interest: The author has completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0070/coif). The author has no conflicts of interest to declare.

Ethical Statement: The author is accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Perlis RH, Gunning FM, Uslu AA, et al. Generative AI Use and Depressive Symptoms Among US Adults. JAMA Netw Open 2026;9:e2554820. [Crossref] [PubMed]
2. Luo X, Wang Z, Tilley JL, et al. Seeking Emotional and Mental Health Support From Generative AI: Mixed-Methods Study of ChatGPT User Experiences. JMIR Ment Health 2025;12:e77951. [Crossref] [PubMed]
3. Ta V, Griffith C, Boatfield C, et al. User Experiences of Social Support From Companion Chatbots in Everyday Contexts: Thematic Analysis. J Med Internet Res 2020;22:e16235. [Crossref] [PubMed]
4. Montgomery SC, Short JR, Ledermann T, et al. Work and Family Transitions Throughout Adulthood and the Impact on Health: A Systematic Review. J Popul Ageing 2023;16:519-64.
5. Brownson-Smith R, Ananthakrishnan A, Hagen O, et al. Conversational Systems for Social Care in Older Adults: Protocol for a Scoping Review. JMIR Res Protoc 2025;14:e72310. [Crossref] [PubMed]