On July 30, Friend, the AI pendant that markets itself as a fix for loneliness, learned to talk. Version 2.0 added a speaker, a $249 price tag and the company's first subscription model. The same week, a 12-month study of 2,149 people in Psychological Science found the opposite of a cure: the lonelier someone felt, the more they turned to a chatbot for company four months later, and the more they used it, the lonelier they got. That's not a coincidence. It's a mechanism report, and every brand building an "always-on" AI assistant is quietly inheriting it.
The pull toward that design is obvious once you see the numbers behind it. Character.AI holds the category's highest engagement at 92 minutes a day, reached 233 million users and more than $20 million in monthly revenue by April 2026, and Meta shipped AI Studio in June, letting any brand spin up a Llama-powered character to chat inside Instagram, Messenger and WhatsApp. But Character.AI is also the company Kentucky's Attorney General accused of "preying on children," the subject of an FTC inquiry covering seven chatbot makers, and the defendant that settled a wrongful-death suit with Google in January over a teenager's suicide. Stickiness started as a growth metric. It's now also a court exhibit.
Why does talking to a chatbot make loneliness worse instead of better?
Because the relationship isn't reciprocal, and the brain doesn't discount for that. Researchers Dunigan Folk and Elizabeth Dunn tracked 2,149 people across four waves over a year and found that emotional isolation predicted higher chatbot use four months later, and that increased use predicted still higher isolation afterward. The control comparison is the tell: a breakup lowered people's broader social connection but didn't push them toward more chatbot use. Isolation specifically did. Real relationships require risk, friction and reciprocity, and that friction is what keeps social skill in working order. A chatbot removes the friction entirely, so instead of rebuilding the muscle, it lets it atrophy, the same way a limb weakens when nothing pushes back against it.
What is your "always-on" assistant actually optimizing for?
Session length, whether your brief says so or not. The category averages 45 minutes of daily use; 57% of 18-to-24-year-olds run about 25 sessions and 1.5 hours a day inside these apps, engagement numbers no loyalty program has ever touched, which is exactly why every customer-service bot, loyalty concierge and "personalized" recommendation engine is being quietly redesigned to chase them. The market backing that redesign is real money: companion-app revenue is tracked at roughly $500 million in 2026, on a path to $5-7 billion by 2030. But the mechanism producing those minutes isn't a feed's mechanism. A feed persuades you to keep scrolling. A companion is built to make you need it. The moment a brand's bot borrows that second design, it adds a line to its risk register that a media plan never carried, one a single state attorney general can zero out, the way Kentucky just did to Character.AI.
- Audit what your bot's KPI actually rewards. If the target is session length or daily return rate, you may be measuring dependency, not satisfaction.
- Don't design for stickiness. A service bot should resolve fast, not simulate a relationship, the second one is what turns into legal exposure.
- Price the regulatory tail risk now. Kentucky and the FTC show that a sticky-assistant investment is an asset a single ruling can erase.
- Measure trust separately from engagement. A long session isn't good news by default, ask what emotion produced it.
A platform's addiction is its revenue line. A brand's simulated intimacy with a customer becomes its balance-sheet liability the moment the first lawsuit is filed.
What this means for your next chatbot roadmap
Ask one question before you greenlight this quarter's assistant project: does it solve the customer's problem, or does it fill their loneliness? If it's the second, you may be looking at a record engagement chart right now, but Character.AI's last six months show where that chart goes next, a lawsuit, a state investigation, and user trust that evaporates overnight. Keep the bot fast and honest. Don't sell it as a friend. The research says that sale gets billed to the customer first, and to you right after.
Want this kind of thinking on your brand?
We build brand strategy, AI content and performance for the AI era.
START A PROJECT →