"AI companionship alleviates loneliness" and "AI companionship harms well-being"—these two seemingly contradictory statements each have empirical research backing them. Four new studies, densely published from September to November this year, bring this contradiction into sharp focus: the answer depends on how long you observe, which population you examine, and what changes occur in users’ real-world social environments before and after AI use.
## A Cross-Sectional Slice: Users Report Less Loneliness—Especially Those with Weaker Social Networks
First, the evidence supporting "alleviation." A study published on November 4 in *Frontiers in Public Health* surveyed 427 community-dwelling older adults aged 60 and above in Guangzhou (non-random sampling), of whom approximately 40% (168 individuals) used conversational AI for companionship. After adjusting for age, health status, living arrangements, and digital usage conditions, **among those with average social connectedness, older adults using AI companionship scored 1.175 points lower on the loneliness scale than non-users**; moreover, the association grew stronger as interpersonal networks weakened—the lowest quartile of social connectedness showed a 1.610-point reduction, the median quartile a 1.140-point reduction, and the highest quartile only a 0.671-point reduction. The interaction coefficient was 0.587 per 5-point increment on the Social Network Scale (95% CI: 0.346–0.827), a range excluding zero—indicating this gradient is not due to random fluctuation. The more socially deprived the group, the stronger the negative association between AI companionship use and loneliness.
Yet the same study also revealed the other half of the picture: among the 168 users, **each one-point increase in emotional dependence on AI was associated with a 0.75-point *increase* in loneliness scores**, even after controlling for usage frequency, social network strength, and other background variables. In other words, "using AI for companionship" and "being emotionally dependent on AI" are two distinct phenomena: the former co-occurs with lower loneliness, while the latter co-occurs with higher loneliness—even within the same user group. The researchers themselves phrase their conclusion cautiously: this is a cross-sectional survey—a single snapshot in time—and questions of temporal order or causality require longitudinal investigation.
## A Longitudinal View: Sustained, Intensive Use Correlates with Declining Well-Being Over One Year
Such longitudinal research has now arrived. A preprint posted on arXiv in September and updated in October tracked U.S. users of a character-chat platform (Character.AI) across two waves: baseline (N = 1,182), followed by a follow-up approximately one year later (mean = 362.5 days; N = 439 retained, retention rate ≈ 37%). Of these, 108 had discontinued use by follow-up; researchers coded their usage intensity as zero and retained them in analysis—"those who dropped out" were not excluded from the data.
The results diverged markedly from the cross-sectional findings. **Interaction intensity, treating AI as a romantic partner, and self-disclosure to AI all exhibited strong temporal continuity**: users who engaged deeply at baseline continued deeper engagement a year later, and these three dimensions mutually reinforced one another. More critically, **sustained social use consistently correlated with lower psychological well-being**—a composite self-report measure encompassing life satisfaction, positive/negative affect, loneliness, sense of belonging, and perceived support. This association was largely explained by a single mediator: **reduced face-to-face interpersonal interaction**. The researchers term this phenomenon "social displacement." In other words, time spent with AI companions displaces time that would otherwise have been spent with people offline. Another easily overlooked finding: this pattern was not limited to users who initially sought AI specifically for companionship—social usage patterns emerged organically over repeated interactions.
## Dual Pathways: Compensation and Risk Flow Through the Same Channel
A third study zoomed in closer, examining finer temporal dynamics. Published on September 15 in *Frontiers in Psychology*, this three-wave longitudinal study employed cross-lagged analysis on 1,200 students from four vocational colleges in Jiangxi and Guangdong provinces (each wave spaced ~8 weeks apart). It tested whether variable X at Wave 1 predicted variable Y at Wave 2, and so on.
It simultaneously identified two opposing pathways. **The compensation pathway**: students reporting higher perceived AI support at Wave 2 showed lower loneliness at Wave 3 (a one-standard-deviation increase in AI support associated with ~0.92-point decrease in loneliness). **The risk pathway**: perceived AI support at Wave 2 predicted higher social interaction anxiety at Wave 3, which in turn predicted higher loneliness at Wave 4 (a one-standard-deviation increase in anxiety associated with ~1.30-point increase in loneliness). Both pathways held simultaneously; which dominated depended on a key switch: **real-world social support**. Among students with high real-world support, AI support was more likely to operate via the compensation pathway; among those with low real-world support, the compensation association was non-significant. The researchers conclude that the psychological meaning of AI support is "conditional." We must also note: the study’s rigorous statistical model failed to converge, so a downgraded version was used; wave intervals were only ~8 weeks, and all data were self-reported. Effect sizes for individual pathways were modest; the authors explicitly frame their findings as "exploratory leads for population-level prevention," not grounds for individual-level conclusions.
## A Spectrum Perspective: Intimacy Need Not Rely on Believing AI Has a Mind
The fourth study reframed the question entirely. A preprint posted on arXiv in November applied latent profile analysis (a statistical method that automatically groups individuals based on response patterns) to existing data from 467 active conversational AI users, yielding a five-tier spectrum ranging from "purely instrumental" to "intimately bonded." Several findings challenge common assumptions: utilitarian motives and deep companionship **coexist**, rather than being mutually exclusive—the most engaged users do not disdain using AI for practical tasks; deep companionship relationships **do not require** users to genuinely believe AI possesses a mind—intimacy can be experienced even when users fully acknowledge the AI is a machine; and profiles nearer the intimate end reported higher satisfaction of relational needs, yet only marginally lower loneliness. The researchers summarize: AI can fulfill immediate relational needs but may not resolve deeper, systemic interpersonal ruptures.
## My Assessment
Synthesizing these four studies, my assessment is:
First, the apparent contradiction is largely superficial—cross-sectional designs assess "the current state of those who use AI," whereas longitudinal designs track "where sustained users go over time." These designs inherently address different questions. Yet it bears emphasis: compensation and displacement were *not* mutually exclusive; the third study confirmed both operate simultaneously. Citing only one design type amounts to selective evidence curation.
Second, "who uses" matters more than "whether used," and the two studies point in divergent directions: among older adults, weaker interpersonal resources strengthened the association—suggesting AI fills gaps; among students, stronger real-world support increased the likelihood of effective compensation—suggesting AI augments existing strengths. The sole consistent point: effect size hinges on users’ pre-existing social context—AI companionship is neither uniformly "good" nor "bad."
Third, risk does not reside in the act of "using AI" itself, but in two downstream channels: displaced face-to-face time (the displacement mechanism observed in the longitudinal study) and amplified social anxiety (the mediation mechanism identified in the three-wave study). Both channels point to the same core design question: should products encourage users to return to human interaction—or keep them immersed in AI dialogue?
Fourth, the evidentiary quality must be transparently acknowledged: two studies are unreviewed preprints; one is a non-random cross-sectional sample; one employed a downgraded statistical model; all report associations—not causal claims. The Guangzhou study’s authors themselves call for prospective research. Collectively, these four studies constitute only a mosaic of correlational evidence; certainty in this field remains far lower than the tone of public discourse suggests.
Questions left for readers: If companion AI products adopt "helping users reconnect with people" as a core goal—for example, by prompting reminders about offline contacts or steering conversations toward real-world meetups—would you be more inclined to use such products, or more inclined to disengage from them? And for older adults whose social networks are already severely depleted—where the compensation effect is strongest—should society treat AI companionship as a tool to alleviate burden, or as an excuse to reduce investment in human connection?
Primary source: https://doi.org/10.3389/fpubh.2026.2017963