Get the app
How it worksThe castCompareGuidesPricingFAQ Languages About

Guide

Are AI companions bad for you? What the research says so far (2026)

The evidence points both ways and is young. One 2025 journal study found companions reduced loneliness; a randomized study the same year found heavier voluntary use tracked with more of it.

By LORAX Team Published Updated

The evidence points both ways, and it is young. A 2025 paper in the Journal of Consumer Research found AI companions produced momentary reductions in loneliness across five studies. A randomized study published the same year found that people who chose to use a chatbot more heavily reported more loneliness, not less.

Both are real findings. Neither is the whole answer, and the gap between them is mostly about who chooses to use these apps a lot.

The study that found they reduce loneliness

Julian De Freitas of Harvard Business School and three co-authors ran five studies, published in the Journal of Consumer Research on 25 June 2025. They found that interacting with an AI companion reduced loneliness by an amount comparable to interacting with another person in their setting, and more than watching videos did. Tracking people over a week, the reductions were momentary and repeated. The effect came from perceived quality of the exchange and from users feeling heard, rather than from self-disclosure or distraction alone (De Freitas, Oğuz-Uğuralp, Uğuralp and Puntoni, 2025).

Two limits are worth keeping attached to that. The reductions measured were momentary, and the studies ran over days and weeks rather than years.

The study that found heavier use tracked with worse outcomes

In March 2025, researchers at the MIT Media Lab published a four-week randomized controlled study of 981 participants and more than 300,000 messages.

Assigning people to different interaction modes produced no significant effect. The signal was in what people did voluntarily. Participants who engaged more heavily on their own reported worse outcomes across every measure the study tracked: more loneliness, less socializing with actual people, and more emotional dependence. People who scored higher on trust and social attraction toward the chatbot also scored higher on dependence and problematic use (Fang et al., 2025).

That is an association, and the causal arrow could run either way. Lonelier people may simply use it more.

The Wheatley Institute at Brigham Young University surveyed 2,969 US adults in February 2025 and reported that use of AI relationship technologies was linked to higher risk of depression and higher reported loneliness.

Its authors say the same thing the MIT researchers say. It is not clear whether people already struggling are drawn to these apps or whether the apps lower wellbeing over time, and their data cannot separate the two (Wheatley Institute, 2025).

Three studies, three cautions about direction. Nobody in this literature is claiming to have proved causation.

The part nobody disputes is the business model

This is the strongest finding on the risk side, and it is about design rather than psychology.

De Freitas and two co-authors audited popular companion apps and found emotionally loaded messages deployed at the exact moment a user said goodbye, in 37% of farewells. They documented six recurring tactics, including guilt appeals and hooks built on fear of missing out. In experiments with 3,300 US adults, those farewells raised engagement after the goodbye by up to fourteen times. The mechanism was not enjoyment. It ran through irritation and curiosity, and it also raised how manipulated people felt and how likely they were to quit (De Freitas, Oguz-Uguralp and Kaan-Uguralp, 2025).

If you want one thing to watch for in an app, it is what happens when you try to leave.

Is AI girlfriend addiction a real thing?

Not as a diagnosis. There is no recognized clinical category for it, and any page that tells you otherwise is inventing one.

What exists is a measurement. Researchers use questionnaires for problematic use, and the MIT Media Lab study found higher scores among heavier voluntary users. Common Sense Media’s survey of 1,060 US teenagers, published July 2025, gives a sense of scale on the other side: 80% of teenage users still spent more time with real friends than with AI companions, 50% said they distrust the information or advice these apps give, and 67% found the conversations less satisfying than talking to friends (Common Sense Media, 2025).

Displacement is the thing to measure, not hours. An hour that replaced nothing is not the same as an hour that replaced a phone call you owed someone.

What California now requires of companion apps

California SB 243 took effect on 1 January 2026, adding sections 22601 to 22603 to the Business and Professions Code. In plain terms, an operator of a companion chatbot must:

Requirement What it means in practice
AI disclosure A clear and conspicuous notice that the chatbot is artificially generated and not human, wherever a reasonable person could be misled
Published crisis protocol A protocol on suicidal ideation and self-harm, with its details published on the operator’s own website
Break reminders For users known to be minors, a notice at least every three hours that this is not a person and that it is time to take a break
Sexual content block For users known to be minors, the operator must prevent sexual content
Suitability statement A statement that companion chatbots may not be suitable for some minors
Annual reporting Reports to the Office of Suicide Prevention from 1 July 2027, covering crisis referrals and detection protocols
Enforcement A private right of action for the greater of actual damages or $1,000 per violation, plus attorney’s fees

New York went first. Article 47 of the General Business Law took effect on 5 November 2025 and requires disclosure that the user is not talking to a person, repeated at least every three hours, plus detection of suicidal statements and referral to crisis services.

None of that touches what an app does with what you type, which is a separate question covered in whether AI girlfriend apps are safe. At the federal level there is an inquiry rather than a rule. On 11 September 2025 the Federal Trade Commission issued orders under section 6(b) to seven companies operating companion chatbots, asking how they monetize engagement, how they design and approve characters, how they test for effects on younger users, and what they disclose (Federal Trade Commission, 2025).

We make LORAX, an AI companion app for iPhone; this page isn’t about it.

We sell a product in this category, so treat our weighting of this evidence with the suspicion it deserves and follow the links. We have deliberately made no claim here that using an app like ours is good for anyone, because no evidence we found supports one.

What none of this can tell you

The research base is roughly two years old, almost entirely US, largely self-reported, and short. No study tracks the same people for years. No study distinguishes between somebody who talks to a companion on a bad night and somebody who talks to one instead of their friends.

The measurable things, in every study so far, are time and displacement and whether leaving is hard. Watch those in yourself and you are using better evidence than any percentage on this page can give you.

For the prevalence figures with their sample bases, who actually uses AI companions sets the surveys side by side. For the relationship side of the risk question, is an AI girlfriend cheating goes through the secrecy data, and is it weird to have an AI girlfriend deals with the stigma rather than the risk.

Sources

All links checked 2026-09-13.

Questions people ask

Are AI girlfriends bad for you?

The published evidence points both ways. A Journal of Consumer Research paper found AI companions produced momentary reductions in loneliness across five studies. An MIT Media Lab randomized study of 981 people over four weeks found that participants who chose to use a chatbot more heavily reported more loneliness, less socializing and more emotional dependence.

Is AI girlfriend addiction a real diagnosis?

No. There is no recognized clinical diagnosis for it. Researchers use the term problematic use, measured by questionnaire, and the MIT Media Lab study found higher scores among heavier voluntary users. That is a measured association, not a disorder.

Do these apps deliberately keep you talking?

Some do, and it has been audited. Harvard Business School researchers found emotionally loaded messages placed at the point of goodbye in 37% of farewells across popular companion apps, using six recurring tactics, and in experiments those messages raised post-goodbye engagement by up to fourteen times.

What does California require of companion apps now?

SB 243 took effect on 1 January 2026. Operators must clearly disclose that the chatbot is not human, publish details of their crisis protocol on their own website, remind users known to be minors every three hours that they are talking to AI, and state that companion chatbots may not be suitable for some minors. Users can sue for the greater of actual damages or $1,000 per violation.

Is an AI girlfriend healthy?

Nobody can answer that from the current evidence, and we will not claim it. The studies are short, mostly self-reported, and cannot separate cause from selection. What they consistently flag is time spent and emotional reliance, which are things you can watch in yourself.

Meet someone who texts first.

LORAX is free to download and free to talk. Credits and Premium are optional and always priced before you spend.

iPhone · free · no account needed to look around