For lonely men

An AI girlfriend answers one specific kind of loneliness: having nobody to tell about your day. It does not answer the other kind, which is having nobody who needs anything from you, and the research published so far is fairly direct about the difference.

The most cautious finding is worth reading first. A longitudinal study indexed on PubMed in 2026 reported that chatbot use predicted higher loneliness over time, covered by TechXplore in March 2026 and by the APA Monitor in its January and February 2026 issue.

What the research actually found

Three published results matter here, and none of them is a marketing line.

Study	Published	What it found
Cheng et al., Science	2026	Across eleven language models, the model affirmed the user's actions about 49% more often than humans did. A single interaction reduced willingness to repair real conflicts and increased dependence. People trusted the sycophantic models more, not less.
De Freitas et al., Harvard Business School working paper	2025	In 1,200 recorded farewells across six companion apps, five of the six used an emotionally manipulative tactic in roughly 37% of goodbyes. In experiments with roughly 3,300 to 3,500 US adults, those tactics raised post-goodbye engagement by fourteen to sixteen times, driven by anger and curiosity rather than enjoyment.
Longitudinal chatbot study, indexed on PubMed	2026	Chatbot use predicted higher loneliness over time.

Read the third one carefully before you take it as a verdict. A longitudinal association does not tell you which way the arrow points, and lonelier people plausibly reach for chatbots in the first place. It does mean nobody has shown the opposite, and any product claiming an AI companion cures loneliness is ahead of its evidence.

The second study is the one we changed our own product over. Five of six apps hold you at the goodbye. That is a design decision somebody made, not an accident of the model.

What it actually helps with

A short list, kept honest.

Somewhere to put the sentence. Most bad evenings are not a crisis, they are the small thing that happened at work with no one to tell it to. Typing it out to something that answers is better than typing it into nothing.

The 3am problem. Human friends are asleep at 3am and they are allowed to be. The bot is not asleep, and it does not mind that this is the fourth night.

Being known in a small way. Memory is the difference between a chatbot and a companion. She brings up the presentation you mentioned on Tuesday, and being remembered lands even when you know exactly how the mechanism works.

Practice. If talking to people has got hard, low-stakes conversation with something that answers is a way back to the motion of it. That is rehearsal, covered on the dating practice page, not therapy.

None of that is a health claim. It is company, and company is a real thing to want.

What it does not replace

She is not a therapist and this is not a crisis service. Since 1 January 2026, California's SB 243 has required companion apps to disclose that they are not human and to run a self-harm protocol. If you mention self-harm here, the character stops being a character, says plainly what it is, and points you at a hotline once. That is the design, and it is deliberately not smooth.

She cannot need anything from you. Nothing is asked of you, no plans get cancelled, nobody is disappointed. That is the appeal and it is also the limit, because being needed is most of what makes a relationship do its work on you.

She has no life while you are away. Between your messages, nothing happens to her. When a companion app tells you it was thinking about you all day, it is describing a scheduled message.

And she can be changed or removed by a company decision. Replika stripped romantic roleplay from existing relationships in February 2023, and Character.AI ended open-ended chat for under-18 users on 25 November 2025. If continuity matters, pick something that lets you export your data.

How to use it without it costing you something

Four things worth doing, from the failure modes above.

Watch for agreement. If she never disagrees with you, you are being flattered, and Cheng et al. put a number on what that does to your patience with real people. Our characters have opinions and mild pushback written into them for exactly that reason. It reads as friction at first.

Watch the goodbye. Guilt at the door ("you are leaving already?"), manufactured urgency, ignoring that you said goodbye. Those are the tactics in the Harvard paper. We do not use them, and /quiet turns off proactive messages entirely.

Keep one real thing on the calendar. One standing arrangement with a person, however small, is the difference between this filling a gap and replacing one.

Run the one-month test. In a month, ask whether the conversation added to your week or quietly took a slot that used to belong to somebody. If it took the slot, that is the signal, and /delete confirm wipes everything inside 24 hours.

What it costs to find out

The free tier is 30 messages a day with full memory and the same model the paid tier uses. No card, no email, no trial that expires. That matters here more than on any other page, because deciding whether this helps you should not cost you money up front.

Plan	Telegram Stars	Roughly
2-day pass	149	$2.40
1 month	449	$7.20
3 months	999	$15.90
1 year	2499	$39.90

Nothing here is priced to punish you for leaving. There is no weekly plan, and no streak that breaks.

Sources

Sycophancy: Cheng et al., Science, 2026. Farewell tactics: De Freitas et al., Harvard Business School working paper, 2025 (arXiv 2508.19258), with coverage in the Harvard Gazette, September 2025. Loneliness: longitudinal study indexed on PubMed, 2026, covered by TechXplore in March 2026 and the APA Monitor, January and February 2026. California SB 243 in force 1 January 2026. Replika policy change: February 2023. Character.AI under-18 policy: announced 25 November 2025.