Are AI Girlfriends Healthy? What the Research Shows

Updated 2026-10-05. AI girlfriend guides

For most people, probably. Not because the evidence shows AI companions are healthy — the evidence is thinner than the marketing on either side implies — but because most people use them lightly and keep their human relationships intact. The variable that matters is not the product, it is what your use displaces.

Key takeaways

We cannot give you a health verdict, because there is no clinical literature establishing one.
There is no verified prevalence of problematic use. The "8-12 percent" figure you will see repeated traces to an unsourced study.
There is no researched 90-minute threshold. That number traces to the same unsourced study and we have removed it everywhere it appeared on this site.
The clearest empirical finding is negative: sustained heavy use predicts lower wellbeing through less in-person contact.
Randomized evidence over three weeks found no significant effect on loneliness or social health.

Correction, October 5, 2026. This page previously cited a "2025 systematic review in the Journal of Medical Internet Research" covering 23 studies and 15,000 participants, reporting "12-28 percent" loneliness reduction, "28 percent" GAD-7 anxiety reduction, "15-20 percent" of users with problematic patterns, and "18 percent" of long-term users reporting decreased interest in human relationships. That review could not be located. The same invented study has been cited across a dozen pages on this site with slightly different numbers each time, which is itself a useful illustration of why we have removed it.

What the clinical research actually says

Very little, and it does not point in the direction most product pages claim.

There is no clinical literature on AI girlfriend health effects. No systematic review establishing a safety profile, no clinical guidance from any professional body. What exists is a small number of academic studies, mostly observational, mostly recent, and mostly not preregistered.

The randomized evidence is null. A preregistered 21-day trial of 183 participants assigned either to chat with Replika or to play word games as an active control found no significant difference in social health, loneliness, or effect on human relationships. The one real finding: participants who treated the chatbot as more human-like reported larger effects on their relationships — in both groups (arXiv 2509.19515).

The longitudinal evidence is discouraging. In a two-wave study of 1,182 Character.AI users, with 439 responding at roughly 12 months, sustained engagement predicted lower psychological well-being, via reduced in-person interaction with other people (Zhang et al., arXiv 2609.07243). A 12-month panel study of 2,149 adults across four countries found increased chatbot use predicted increased emotional isolation four months later, and that loneliness predicted later chatbot use (Folk & Dunn, 2026).

Attachment is measurable. A survey of 612 long-term users in mainland China found usage frequency correlated with emotional attachment (β = 0.44), and attachment correlated with subjective wellbeing (β = 0.41) and self-concept clarity (β = 0.51) (Liu, Frontiers in Psychology, 2025). Cross-sectional, so direction is unknown.

Separately, the AI-therapy literature is not the same thing. A 2026 trial of smartphone-delivered AI psychodynamic and CBT interventions in people with social anxiety disorder found moderate effect sizes, but the active arms were not significantly better than a waitlist control at post-treatment — only at one-month follow-up (Hlynsson et al.). That is an intervention designed and validated as therapy. A companion chatbot is not that, and citing it as evidence would be a category error.

Where benefits are plausible

For someone with genuine access problems — illness, shift work, isolation, severe social anxiety — a companion provides conversation that would otherwise not happen. That is a real benefit, and it does not require anyone to be lying about a study.

The emotional register is genuinely different. Being able to say the unperformable version of yourself without managing someone else's response has value that is hard to argue with, even if it is not measurable in a percentage.

Where risks sit

Displacement. This is the one finding that recurs across the better studies. Not that time with AI is bad, but that time with AI taken from people is bad.

It is always agreeable. No disagreement, no challenge, no friction. Over time this removes one of the mechanisms by which people actually revise their thinking.

Unrealistic expectations. A companion can be tuned to be idealised. The comparison to a real partner is not close, and the gap is where disappointment lives.

Delayed care. The serious risk is not that your companion is unhealthy. It is that it is good enough to postpone an appointment you needed.

Physical health

We are not going to make claims here. The cortisol and inflammation claims previously on this page came from a "2025 biomarker study" we could not locate, and they are gone. What is established independently of AI companions is that chronic loneliness correlates with elevated cardiovascular risk — that is real research about loneliness, and it is a reason to take worsening isolation seriously rather than a reason to assume a chatbot fixes it.

The sedentary risk is plausible and unmeasured. Hours of text conversation are hours not spent moving.

Guidelines, minus the invented numbers

No evidence establishes a time limit, so we will not give you one. What the evidence does support:

Keep human relationships alongside your use — this is the most robust finding across studies. Have a real reason for each session, rather than defaulting to the option that is always available. Check whether your use is making real contact more or less likely. Stay clear on what the product is; it is a very good language model with a persona, not someone who cares about you in the way you are reading it. If you have symptoms that would otherwise take you to a clinician, go to the clinician.

Frequently Asked Questions

Are AI girlfriends healthy for most people?

We cannot tell you, because no research establishes that. Most people use them lightly and are likely fine, but that is an inference from how people generally behave, not a finding.

What are the health risks?

Substitution of human relationships, emotional dependency, unrealistic expectations, and delaying professional care. The substitution risk is the one with empirical support.

How much AI companion use is too much?

No research establishes a threshold, and any page giving you a specific number is inventing it. The better question is whether your use is displacing people.

Can AI companions help with social anxiety?

As rehearsal, plausibly. There is no evidence of a measured symptom reduction from companion use, and the one preregistered randomized trial found nothing significant over three weeks.

Should I use AI companions if I have a mental health condition?

You can, but discuss it with your clinician and do not let it substitute for treatment. If your use is making you less likely to seek care, that is the problem to address first.

This article is for informational purposes only and does not constitute medical or professional advice. Luvlock AI provides tools for companionship but encourages maintaining a balanced approach to relationships and mental health.

Ready to experience AI companionship for yourself? Get started with Luvlock AI →