The 2 a.m. Problem

Read this week's mental health headlines closely and you notice something odd: almost none of them are about the therapy hour itself.

They're about 2 a.m. They're about the intake phone call. They're about the routing decision that happens before a patient ever meets a clinician, and the twelve weeks after a referral when nobody is watching. AI has not shown up in the consulting room. It has shown up in the gaps — the parts of care that were already thin, already under-resourced, already invisible to the clinician. And it turns out those gaps are not empty space. They are where the therapeutic relationship either holds or breaks.

Start with the most concrete piece of reporting of the week. In The American Prospect, co-published with Capital & Main, Nick Romeo interviewed more than a dozen therapists and clinicians in California and Wisconsin about what happens when AI-powered systems take over triage. A Kaiser triage clinician in Walnut Creek described going from a team of nine to a team of three in three years, with a third of her daily calls now coming from patients who have simply struggled to get to the right place. A Kaiser therapist in Oakland put the design flaw more sharply than any policy paper could: "Lack of motivation and lack of follow-through are the most common symptoms of depression on the planet... We wouldn't tell a paraplegic, 'Hey, walk down the hall in order to get your wheelchair.'" A clinical psychologist described referrals for serious depression getting filtered into group anxiety classes. Kaiser and Rogers Behavioral Health both told the publication that licensed clinicians make all placement and treatment decisions, and that their technology supports rather than replaces clinical judgment.

I want to be careful here, because the interesting failure is not "the algorithm was wrong." It's that the system optimized the handoff — speed, throughput, routing — while quietly deleting the person who used to carry context across it. What the clinicians describe as "service recovery" is the cost of that deletion, paid in emotional labor, weeks of delay, and occasionally a missed prior attempt.

Now look at the same pattern from the patient's side. Three UMass Chan clinicians — two psychologists and a child psychiatrist — opened their piece in The Conversation last week with a scene: a teenage girl awake at 2 a.m., not wanting to wake her parents, typing into a chatbot instead. Their data is sobering. Seventy-two percent of teens have interacted with AI companion chatbots. Sixty-three percent of teens and young adults who used a chatbot for mental health advice told no one. And the safety of those conversations, they note, tends to degrade the longer the conversation runs — so the 2 a.m. rumination becomes a riskier interaction at 3 a.m.

That is the same gap. Not the session. The space around it.

The industry response this month was to try to make the gap safer. OpenAI launched ChatGPT for Teens on August 18, with age-appropriate protections on by default and expanded parental controls — after, as TechCrunch's Sarah Perez pointed out, the product had already scaled to 900 million weekly users. Regulators are trying too, unevenly: Quartz's survey of state law found six states split four ways on AI therapy chatbots, with the same product legal in one jurisdiction and criminally exposed in another. In California, Assembly Bill 2575 would protect health care workers from retaliation for overriding an AI recommendation; a second bill would bar clinical decisions made solely on AI output.

Guardrails are good. But notice what all of this is: a defensive posture. We are trying to make an unsupervised space less dangerous rather than asking why it is unsupervised.

The clinicians themselves have been unusually clear about what they actually want. The APA's 2026 Chatbots and Mental Health Survey of 1,242 licensed psychologists found 94% saying chatbots can't treat conditions with appropriate nuance — but also 40% optimistic that these tools could help patients when a professional isn't available. Read together, that isn't ambivalence. It's a specification. APA CEO Arthur C. Evans Jr. put it plainly: these tools "work best when used to complement a relationship with a licensed, human professional who understands how to treat a person, not a prompt."

That's the throughline, and it's the one I keep coming back to as a founder. The gap between sessions is not dead time to be automated or filled with a synthetic companion. It's where the clinical picture actually forms — the sleepless night, the fight with a friend, the week the meds started working or didn't. A therapist walks into every session partially blind to it, and reconstructs it from memory, in a room, with a person who may not remember accurately either.

So the design question is narrow and, I think, answerable. Can the 2 a.m. moment become signal that reaches the person who is qualified to act on it — with consent, and with the clinician's judgment as the decision point rather than the thing being routed around? Can intake carry context forward instead of stripping it, so that a first session starts warm rather than cold, and the match between patient and therapist is made on more than a dropdown menu of insurance and availability?

Build it that way and AI strengthens the relationship the entire field agrees is doing the therapeutic work. Build it the other way and you get exactly what the Kaiser and Rogers clinicians are describing: efficiency at the seams, and a patient at the end of the phone who has been sent all over the place.

A former Rogers therapist gave The American Prospect the sentence I'd put on the wall. "Therapy works because you build a relationship with your therapist. AI can't build that connection in the same way that an actual human being can."

She's right. Which is why the job isn't to build the connection. It's to keep it from breaking in between.