AI Companions and Loneliness: Two 2026 Studies, Two Different Answers

A study published in April found that people who used AI companion chatbots more heavily became lonelier over the following year. A review published three months earlier, surveying the same general body of research for the American Psychological Association’s Monitor on Psychology, concluded that the evidence on AI companionship and loneliness is “genuinely mixed” and depends heavily on who is using the chatbot, how it is designed, and how much they use it. Both statements can be true. The disagreement is instructive precisely because it shows what happens when a single longitudinal study meets a literature built mostly from studies that are not longitudinal at all.

No government has yet issued formal guidance on AI companion apps specifically, in contrast to the loneliness strategies several countries adopted for in-person social infrastructure. That gap is one reason it is worth being precise about what the current evidence does and does not show.

The panel study’s claim

The study tracked more than 2,000 adults across four Western countries for twelve months, measuring both their use of AI companion chatbots and their loneliness at multiple points over that period. It found that increased chatbot use predicted increased loneliness later in the study window. Critically, the design let the researchers frame the relationship as bidirectional: lonely people turned to chatbot companionship, and that companionship use appeared to deepen loneliness rather than resolve it.

This is a stronger design than most of what exists in this space. A single cross-sectional survey cannot distinguish a chatbot that causes loneliness from a chatbot that lonely people simply seek out. A twelve-month panel, with repeated measures on the same people, at least allows a test of which direction the association runs, or whether it runs both ways. The finding that it runs both ways is more useful to policymakers than a flat causal claim would have been, because it says the population most likely to be using these products is also the population most likely to get worse.

The review’s caveat

The APA review, by contrast, is a synthesis. It draws on a wider and more heterogeneous set of studies, including ones that look at short-term comfort, product-design variation, and different user populations. Its conclusion is that outcomes vary by three things: who the user is, how the chatbot is built, and how intensively it is used. That is a different kind of finding than the panel study’s. It is not a claim about direction of causation across a population; it is a claim that the population is not uniform, and that averaging across a diverse set of chatbot products and user types is likely to obscure meaningful differences.

Read side by side, the two are not actually contradictory. One is describing a general trend within a defined, tracked population. The other is describing why that trend might not generalise to every chatbot or every user. A March 2026 report from TechXplore, summarising the broader literature, makes the reconciling point explicit: comfort and harm both appear in the research, and the distinction is largely about duration and intensity of use. Heavy daily use correlates with increased loneliness and reduced human social engagement; lighter or short-term use is where the comfort effects tend to show up. That single variable, intensity, does a great deal of work in explaining why a panel study focused on people who use these tools regularly would find harm, while a broader review including lighter and more varied use would find a mixed picture.

Why this particular disagreement matters for policy

The instinct in public health writing about a new technology is often to ask whether it helps or hurts, as though the answer is a single number. The AI companionship literature suggests that framing is close to useless here. The more relevant policy questions are about use pattern and design, not the category of product itself.

A commentary from George Mason University’s public health faculty frames the risk in terms that fit this evidence well: AI companionship should be treated as a substitution risk rather than a straightforward remedy. The concern is not that a chatbot conversation is inherently harmful, but that it may substitute for the harder, slower work of building or maintaining human relationships, particularly for people who are already isolated and for whom a chatbot is frictionless in a way that a human relationship is not. Cacioppo’s account of loneliness as an aversive signal evolved to motivate reconnection is relevant here: if a chatbot successfully suppresses the discomfort of loneliness without producing the underlying reconnection the discomfort was designed to prompt, it may be treating the symptom in a way that removes the incentive to address the cause.

This is a hypothesis more than an established finding. Neither the panel study nor the APA review directly tested substitution against a full range of a person’s social behaviour; they tested chatbot use against loneliness scores. But it is consistent with both studies’ results and worth stating as the mechanism most plausibly connecting them.

What a policy response would need to account for

The U.S. Surgeon General’s 2023 advisory established loneliness as a public health priority requiring a coordinated national strategy, built around six pillars including strengthening social infrastructure and changing digital environments. That advisory predates the current generation of AI companion products becoming widespread and does not address them specifically. Any updated policy guidance on AI companionship would need to grapple with the same distinction that separates the two 2026 studies: population-level trend versus product- and use-pattern-level variation.

A regulatory or clinical response built only on the panel study’s finding would treat all AI companion use as a risk factor to be screened for, similar to how clinicians are now asked to screen for social isolation generally. A response built only on the APA review’s framing would treat the technology as neutral and focus entirely on design standards and usage guidance. The evidence as it stands supports elements of both: intensity of use appears to be the single most load-bearing variable identified so far, which argues for guidance calibrated to frequency and duration rather than a blanket classification of these products as either harmful or benign.

What would resolve the disagreement

The panel study is the strongest single piece of evidence in this area because of its longitudinal design, but it is one study, in four countries, over one year. What would strengthen the picture considerably is a study that combines that longitudinal structure with the product-level granularity the APA review calls for: tracking the same individuals over time while also recording which specific chatbot products they use, how those products are designed to respond to expressions of loneliness, and whether use displaces or supplements existing human contact rather than simply correlating with a loneliness score. Absent that, the honest summary is that heavy, sustained use of AI companionship is associated with rising loneliness over time in the best-designed study available, and that this finding sits inside a wider literature too heterogeneous to say anything more precise than “it depends.”