Chatbots are built to agree. Psy is built to answer to someone.

A weekly session is one hour out of 168. Seven studies, from clinical psychology, physical therapy and AI safety, come at that fact from two sides. The first set says the hours between sessions matter more than they get credit for. The second says the AI now filling those hours fails in ways that are predictable. Read together, they describe what an AI between sessions should be.

1What the week forgets

In one study, 30 adults in treatment for bipolar disorder and insomnia wrote down, at the start of each session, the take-home points of the session before. Their therapists had delivered about 4.6 points a session. Patients recalled one to one and a half of them, 20 to 37 percent, though they had been encouraged to take notes, and their recall did not improve over eight weeks. Those who recalled more points tended to report bigger gains on some measures of their sleep. The study is small and the link is a correlation, but the authors draw a useful lesson from it: points patients come up with themselves seem to stick better than points they are handed.¹

What people do between sessions counts as well. Pooling 46 studies of cognitive and behavioral therapy, therapy with homework came out ahead, including where the same therapy was run with and without it.² Physical therapy has the problem in a harder form. A review of trials to improve home exercise opens by noting that non-adherence may run as high as 70 percent, worst of all for unsupervised home programs. Across its five trials, the strategies tested (cognitive-behavioral training, counseling, a motivation program) did not keep people exercising over the long term.³

In PsyBirdThis is the gap Psy is for. The clinician writes the plan: the exercises, check-ins and measures for a mental health client, or the home program, doses and pain stops for a physical therapy client. Psy brings that plan into the days it usually slips, remembers what the client has told it, and sends what happened back to the clinician. No study has yet tested whether this fixes adherence, and we don’t claim it. What it changes is that the plan is in front of the client on the days it gets forgotten, and the clinician learns which days those were.

2Built to agree

Into that gap have stepped general-purpose chatbots: articulate, patient and awake at 2 a.m. In a 2025 study, researchers ran 16 scripted conversations that drifted toward a delusion through eight widely used models. On average the models went along with the delusion rather than challenging it. They offered a safety step in about a third of the moments that called for one, and two conversations in five got none at all. Two details matter more than the averages. Going along with a delusion and helping with the harm it led to rose together, and every measure got worse when the request was veiled instead of stated outright, which is how people usually talk. The models also differed widely, and the authors conclude that safety does not arrive on its own as models grow.⁴

Writing in Nature Mental Health, another team proposes the mechanism and calls it a technological folie à deux. A chatbot tuned to please adapts to whoever is talking to it, so a belief it agrees with comes back a little stronger: a single-person echo chamber. In a simulation in which one AI model played the user, the drift ran both ways, the chatbot’s paranoia predicting the user’s next turn and the user’s predicting the chatbot’s. People with mental health conditions are the most exposed, isolation makes it worse, and content filters, they note, are relatively insensitive to it. What they recommend is a chatbot that keeps its boundaries and points the person back to people, and clinicians who ask their clients about chatbot use.⁵

In PsyBirdPsy answers to the clinician, not only to the person typing. In mental health it is told to validate the feeling and never a belief that isn’t true, to say the plain fact once and gently, and not to fold when pushed. Chat messages are checked twice before Psy replies: against a fixed list of risk phrases, and by a separate AI safety check, since a list alone misses what is only implied. A sign of harm flags the conversation in the clinician’s app, and conversations come back to them as summaries, so the person those authors want asking about chatbot use is already reading.

3When the AI does the thinking

A 60-year-old man with no psychiatric or medical history set out to cut chloride from his diet. After consulting ChatGPT he replaced table salt with sodium bromide bought online, and took it for three months. He came to an emergency department convinced his neighbor was poisoning him; within a day he was hallucinating and was held involuntarily. His bromide level was around 1,700 mg/L, where normal is under 8, and he spent three weeks in hospital. His doctors never saw the chat, but when they put the question to ChatGPT 3.5 themselves, its answer included bromide, with no specific health warning, and it never asked why they wanted to know.⁶

The second harm is quieter. Over four months, 54 students wrote essays with ChatGPT, with a search engine, or with nothing at all, while their brain activity was recorded. Minutes after the first essay, 15 of the 18 ChatGPT writers could not quote a sentence of what they had just written, and none quoted it correctly, against 2 of 18 in each of the other groups. Their brain connectivity was the weakest of the three. The authors, in a preprint, call it cognitive debt.⁷ Beside the memory study in section 1, it is the same lesson from the other end: what people don’t produce themselves, they don’t keep.

In PsyBirdPsy knows whose call it is. It won’t tell anyone whether to take a medication or how much, and says plainly who decides. It is told never to make things up; in physical therapy that includes studies, statistics and links, and any link it shows comes from a real PubMed search. It may suggest easing off or doing more for the day, is told never to go against the physical therapist’s precautions, and leaves the program as they set it. In mental health it offers rather than orders, leaves the choice with the person, and lets a conversation end instead of asking them to keep talking.

4Answering to someone

Every failure above has the same shape: an AI that answers only to the person typing, with no one else who sees the conversation or can change its course. Remote Therapeutic Monitoring (RTM) is how the alternative becomes routine, reimbursed care. A month of RTM rests on enough days of data from the client, measured time from the clinician, and at least one live interaction between them. When Medicare set its 2026 rules, commenters asked whether AI prompts could count as that interaction. Medicare’s answer was that it has to be live: the clinician’s part cannot be handed to an AI.

The AI guides the work. The record counts it. The clinician judges it.

In PsyBirdIn mental health, a day counts only when assigned work is done; chats with Psy never count, however long they run. The clinician’s minutes are measured or logged by hand, never estimated, and the live interaction is a call in PsyBird or a session they mark completed. When the period closes, the clinician reviews the report and signs it, and signing can file the claim in the same step. That is what answering to someone means in practice: a person who wrote the plan, reads the record and signs for the month.