Is talking to an AI healthy? What the research shows

What does the research currently show?

The field is young, the studies are small, and most are correlational. Confident claims in either direction — “AI companions cure loneliness” or “AI companions cause loneliness” — exceed the evidence. What does exist is a small set of carefully designed studies pointing in seemingly opposite directions, and the more useful question is why.

The encouraging finding

In a 2024 study published in npj Mental Health Research, Maples, Cerit, Vishwanath, and Pea followed more than a thousand students using the companion app Replika. The students were measurably lonelier than typical peers. The headline finding cut against the dominant cultural expectation: roughly three times as many students reported the AI stimulated their human relationships as reported it replaced them, and a small but meaningful share reported the AI had interrupted suicidal thinking. For this subset, the AI functioned as a bridge — a low-stakes outlet that produced enough emotional regulation to facilitate human contact, not as a substitute for it.

The cautionary finding

In a 2025 study, OpenAI and the MIT Media Lab combined a four-week controlled trial with analysis of millions of real ChatGPT interactions. The pattern: the heaviest daily users reported more loneliness, more emotional dependence, and less in-person socialization. The users whose interactions were most intimate or self-disclosing reported the highest loneliness. The direction of causation is genuinely unresolved — lonelier people may simply use AI more — but the correlation is consistent with a substitute-use risk.

The dividing line between help and harm

Read together, the two studies sketch the same line: AI conversation tends to help when it operates as a supplement that sends the user back toward human connection, and tends to hurt when it operates as a substitute the user disappears into. The variable is not whether the AI is sophisticated or affectionate or convincing. The variable is the shape of the habit — and the design features of the product that push the habit one way or the other.

The design features that matter

The cross-study pattern points to several specific product-design features as relevant to outcomes.

1. Session structure. A phone call ends by design. An infinite chat thread does not. Bounded sessions remove a design lever that engagement-optimized products typically use to extend session length.
2. Re-engagement mechanics.Streaks, badges, intermittent notifications, and “the AI misses you” framing are documented engagement-amplification patterns in consumer software. Their presence in an AI companion product correlates with the substitute-use pattern.
3. Romantic and parasocial framing. Products framed as romantic partners produce qualitatively different use patterns than products framed as friends or conversation tools. The OpenAI / MIT study's “personal” high-intimacy use cluster overlaps substantially with romantic framing.
4. Crisis handoff. Whether the product refers to human crisis services (988 in the U.S.) when distress is detected, or whether it attempts to process the distress itself, is a structural indicator of intended use.
5. Honesty about being AI. Products that acknowledge their artificial nature when asked produce different long-run use patterns than products that obscure it.

Where voice-based AI fits, honestly

Voice-based AI conversation tools — including CallByrd, a phone-based AI designed for unstructured conversation — inherit several of the bridge-favoring design features above by virtue of the format itself. A phone call ends. There is no infinite thread. There is no avatar to bond with parasocially. CallByrd specifically declines romantic framing, discloses that it is AI when asked, and refers users to 988 when distress is detected. These design choices do not guarantee bridge-style use; they make it more accessible. The user's actual use pattern is what determines the outcome.

The bottom line

The early research on AI conversation does not support a simple verdict. AI used as a supplement that returns users to human connection appears to help; AI used as a substitute appears to hurt. Design features that encourage bounded, honest, non-romantic, crisis-aware use make the bridge pattern more accessible. The user's habit is the determining variable. Anyone experiencing thoughts of self-harm should contact the 988 Suicide and Crisis Lifeline rather than an AI tool.