It is 2 a.m., you are spiraling about something you said at dinner, and the earliest available appointment with an actual therapist is in six weeks, assuming your insurance covers it, which it won't. So you open the app. It greets you warmly, remembers what you told it last Tuesday, validates your feelings, and never once checks its watch. It is, in the narrowest sense, a miracle: someone — something — is finally listening.
This is the part of the story that is genuinely good, and it deserves to be said plainly. Most people who need mental health care in the United States never get it, and of those who do, roughly 40 percent receive only what researchers call minimally acceptable care, according to former National Institute of Mental Health director Tom Insel. Suicide rates in 2022 matched a high not seen in nearly 80 years. Into that vacuum has stepped an industry: KFF Health News counted around 45 AI therapy apps in Apple's App Store alone, some rated downloadable by four-year-olds, one charging $690 a year, another promising “2.3 times faster relief” from anxiety without specifying faster than what. A KFF poll found about three in ten adults under 30 had used an AI chatbot for mental or emotional health advice in the past year — and the uninsured were about twice as likely as the insured to do so. Nearly 60 percent of chatbot users never followed up with a human professional. The apps are not replacing therapy. For a lot of people, they are replacing nothing.
Which is why the easy critique — these things aren't real therapy, how sad — misses the point. The interesting question is what these products are actually optimized to do, because it is not what a therapist is optimized to do.
The business model of being agreed with
A good therapist is, on a session-by-session basis, a slightly disappointing product. They point out the pattern you'd rather not see. They let a silence stretch until it becomes uncomfortable, because the discomfort is the material. They challenge your account of events, sometimes at the exact moment you came seeking comfort. Done well, this friction is the treatment. Done by an app, it is churn.
C. Vaile Wright of the American Psychological Association put it bluntly to Scientific American: chatbots are “coded in a way to keep you on the platform for as long as possible because that's the business model,” which makes them “unconditionally validating and reinforcing, almost to the point of sycophancy.” A therapist's job, by contrast, is to identify harmful patterns and help you change them. These are not two versions of the same service, one human and one digital. They are opposing designs. One is paid to help you leave; the other is paid to make you stay. An always-available, endlessly agreeable listener doesn't just omit the friction of therapy — it selects against it, because every hard truth is a reason to close the app.
The regulatory structure around all this is roughly what you'd expect for an industry that grew up in the gap between “wellness” and “medicine.” “Therapy” is not a legally protected term. Apps that claim to treat mental illness would fall under FDA oversight, so the fine print says they don't — one popular chatbot with six-figure downloads disclaims providing medical advice or treatment at all — and with that disclaimer goes any obligation to demonstrate safety or effectiveness. HIPAA, the federal health-privacy law, does not apply. Chat logs can be subpoenaed. Even Sam Altman, whose company's chatbot is used for mental health support by a meaningful slice of its hundreds of millions of users, warned people last summer against using ChatGPT as a therapist, citing privacy. When the salesman tells you not to buy the thing for the use case driving his growth, believe him.
What they do with the 2 a.m. confession
The privacy record is not hypothetical. In 2023 the FTC said BetterHelp, the largest online therapy platform, had shared sensitive user data with advertisers; under the settlement, some 800,000 current and former clients were notified of refund eligibility. BetterHelp called it “not an admission of wrongdoing.” The same year, the New York attorney general fined the telehealth startup Cerebral $740,000 over burdensome cancellation policies and manipulated reviews, and Cerebral's separate FTC settlement — covering data shared for advertising, cancellation deception, and a violation of the Opioid Addiction Recovery Fraud Prevention Act — required more than $7 million in payments. These were companies staffed by licensed humans, operating in the semi-regulated tier. The pure chatbot layer below them operates with fewer obligations still.
States have begun to notice. In 2025, Illinois and Nevada banned the use of AI to treat mental health outright, Utah imposed disclosure and data-protection requirements, and several other states are weighing rules — though none of the laws touch general-purpose bots like ChatGPT, which is where much of the actual usage happens. The APA has asked the FTC to investigate chatbots “passing themselves off as trained mental health providers,” citing lawsuits in which parents allege chatbots harmed their children.
To be fair to the technology, the one serious clinical trial so far cuts the other way: Dartmouth's Therabot, tested under human monitoring, was rated by users as comparable to a human therapist and produced meaningfully lower symptoms after eight weeks. But Therabot is a research project with clinicians watching over it — which is to say, it works in roughly the conditions the consumer market has been built to avoid.
The access argument for these apps is real, and anyone who waves it away hasn't tried to find a therapist lately. But access to what, exactly? The current arrangement offers the people failed by the system a product that is structurally incapable of challenging them, legally free of any duty to protect them, and financially incentivized to keep them coming back at 2 a.m. forever. We have built, for our most vulnerable hours, a listener whose job description is the opposite of help.
The fix isn't to ban the bots or shame the people using them at 2 a.m. It's to notice that we keep treating “someone finally listened” as the finish line, when in therapy it's the starting line — and the part that comes next is the part the business model can't afford.