RN
Two serious research groups have looked at whether talking to an AI companion makes people less lonely, and they came back with opposite answers. That disagreement is more useful than either answer on its own, because of where the two studies part company.
A twelve-month longitudinal study of more than two thousand adults across four Western countries, published in Psychological Science, found that people who felt less socially connected subsequently used social chatbots more — and that heavier use went on to predict higher emotional isolation. A separate four-week randomised trial found something compatible but narrower: voice-based interaction produced a modest reduction in loneliness, while heavy daily use tracked with greater loneliness, more dependence and less real-world socialising. Against both sits a Harvard Business School working paper arguing that AI companions reduce loneliness outright, which has not been through peer review.
Read together, the pattern is not "does it work" but "at what dose". Light use looks harmless and sometimes pleasant. Heavy use is where every study that finds harm finds it. Nobody signs up intending to be a heavy user, and the products are not built to keep you a light one.
What you are actually buying
Strip the marketing off and the product is a subscription to a language model wearing a persona, with a memory store attached. The memory is the part that makes it feel different from a chatbot, and the memory is also the product's inventory: everything you disclose becomes context that makes the next conversation land better.
That is a genuinely clever design, and it means the incentive runs one way. An app that is good at its job is an app that makes you want to type more, more often, for longer. There is no version of this where the company benefits from you closing it.
The regulatory record is not ambiguous
The privacy side of the category has been examined, and the results are on the record rather than a matter of opinion.
The Mozilla Foundation reviewed eleven romantic AI chatbots in February 2024. Ten of the eleven failed its minimum security standards. Most of the privacy policies said very little about whether conversation content trains the underlying models, and most gave users no way to opt out of that use. At the time of the review the eleven apps had racked up in the region of a hundred million downloads on Google Play alone.
Italy's data protection authority suspended Replika in February 2023 over risks to minors, and later fined its developer five million euro, finding no valid legal basis for the processing and no working age verification. The same regulator opened a further inquiry into how the model itself was trained.
In September 2025 the US Federal Trade Commission issued orders to seven companies offering AI companion products, demanding records on how they monetise interactions, how they design and sign off the characters, and what they know about the effect on minors.
None of that means the category is illegitimate. It means the intimate-disclosure part is the part regulators have been looking at, and that a reasonable person should look at it too.
The asymmetry nobody mentions
Conversations with a companion app are among the most sensitive text a person will ever produce. Sexual preferences, insecurities, health, work trouble, the specific shape of a bad week — all typed voluntarily, in a context designed to reward candour, into a service most people have not read the retention terms for.
Compare that with the porn side of the same industry, where you watch and leave. The companion product inverts it: the value flows from what you supply. That is a different transaction, and it deserves a different level of attention than picking a tube site.
If you want one anyway
Plenty of people will, and there is nothing wrong with that. A few things are worth doing first.
Find the retention and training clauses before the free trial ends, not after. The question that matters is whether conversation content trains the model and whether you can refuse — Mozilla's finding was that most services answer no to the second.
Set a rule about frequency before you have a habit, because the dose finding is the most consistent result in the literature. If checking it becomes automatic, that is the condition the studies associate with harm.
Assume the persona can change. Companies rewrite system prompts, swap underlying models, and adjust what their characters will and will not do, sometimes overnight and sometimes because a payment processor made them. Replika's users found this out in 2023 and it has happened repeatedly since.
And take the advertising for what it is. A product that promises never to disappoint you is describing a design goal, not a relationship.
The services this directory has actually inspected sit under AI sex chat sites, with the image-generation side kept separate under AI porn sites. If you have decided to subscribe, the billing mechanics and the question of what "uncensored" covers are the two things worth reading next.