The 167 Hours Nobody Is Regulating

The number that stayed with me this week was not the one about bans or funding. It was 63.3%.

That is the share of adolescents and young adults who used an AI chatbot for mental health advice and did not tell a parent, a friend, or another trusted adult, from a JAMA Pediatrics study covered by Becker's Behavioral Health. The same study estimates 8.2 million young Americans, roughly one in five, did this in 2025. Nearly 92% found it somewhat or very helpful. Almost 43% came back at least monthly.

Read that last figure again. A recurring monthly appointment, kept faithfully, that appears in no chart, no intake form, and no treatment plan. John Constantino, MD, who leads behavioral and mental health at Children's Healthcare of Atlanta, told Becker's that young people are drifting into conversations with AI that mirror what they would otherwise bring to a caregiver or clinician, "often without realizing that's what they're doing." His warning was blunt: "AI can be seductive in pulling, especially vulnerable youth, into that kind of a role or position or mindset."

One pattern ran through the stories I read this week, and I do not think it is the one most people are taking away.

The line is being drawn around the chair, not around the week

The regulatory story is real and it is moving fast. Becker's reported this week that five states enacted laws in 2026 restricting AI therapy chatbots, Colorado, Maine, Rhode Island, Tennessee and Vermont, joining Illinois and Nevada from the year before, according to the Transparency Coalition's mid-year legislative report. Maine's took effect July 29. Colorado's lands August 12. Most of these laws work the same way. They lean on the state's existing authority to license mental health professionals and say, in effect, this is a licensed act, and software does not hold a license.

Quartz, covering the same wave, described states as "all over the map": six states split four ways, with the same product legal in one jurisdiction and criminally exposed in another. Their framing of where the trouble starts is precise. It begins when the bot plays doctor, handing out a diagnosis, a treatment plan, or a license number it invented.

I agree with the direction of these laws. I have argued for a while that psychotherapy should stay with qualified clinicians, and it is clarifying to watch legislatures arrive at the same place from the licensure side rather than the AI-safety side. But notice what the laws settle. They settle who may occupy the therapist's chair for fifty minutes. They say almost nothing about the other 167 hours in the week.

And the 167 hours are where the demand lives.

Warmth is not the same as help

There is a tempting counterargument: maybe the chatbot is fine for the in-between, because people find it kind. Quartz took that apart this week in a piece on what chatbot empathy studies measured. Since a 2023 study in which health care professionals rated chatbot responses as warmer and more empathetic than physicians', that finding has circulated as evidence that AI might be better at the emotional core of therapy. Quartz's argument is that the high empathy scores are largely measuring agreeableness, not clinical skill, and that people in crisis often need pushback more than validation.

That maps onto something clinicians say often. A meaningful part of the work is the moment the therapist declines to agree with you: the gentle contradiction, or the question that makes the room uncomfortable for thirty seconds. A system optimized to keep you talking is unlikely to do that, because doing it costs engagement.

So the chatbot wins on availability and loses on judgment. The clinic wins on judgment and loses on availability. Both are true at 2 a.m., which is when it matters.

The market has already started voting

The capital side has reached the same conclusion. Market analysis of 2026 mental health funding from research2guidance found money concentrating on clinical infrastructure, psychiatry copilots, clinical AI assistants, and clinician-supervised platforms, while generic LLM chatbots dressed as therapy became, in their words, the most visible loser of the year. Their summary of the new bar is the sentence I would tape to a wall: "Human in the loop is now a precondition, not a feature." That analysis was published in May. I include it as the backdrop against which this week's regulatory news landed.

What they flag as underbuilt is the interesting part: measurement infrastructure and care navigation, the connective tissue between appointments.

What we are building

Put the week together. Legislatures have defined the therapist's role and defended it. Researchers have shown that what feels like empathy from a chatbot is frequently compliance. Millions of young people are holding a secret standing appointment with software because the alternative is a wait list and an awkward conversation. Investors are backing the companies that put a clinician at the center.

None of that closes the gap. It says clearly who should own it.

At Lisner we have taken the position that the answer is a better-informed clinician, not a better chatbot. Whether AI can do therapy is now largely settled, by statute in seven states. The open problem is what a therapist knows about a patient's week before the patient sits down. Today the 167 hours produce a private log no clinician sees, if they produce anything at all. They could produce signal the clinician can see, when the patient chooses to share it.

Constantino's line about a Georgia school district stuck with me. Every student's ID badge lists the name of one specific adult in the building who is their person. It is a small detail that stands in for the entire thesis. The technology worth building makes sure that person is not walking in blind.

"Mental health is health," Constantino said, arguing that health systems need to stop outsourcing it. Same instinct, different layer. AI in mental health should be the reason the clinician already knows what happened on Tuesday, not a way to route around them.