The AI Clinic · Issue 014AI in patient care
AI is already in the treatment room.
Therapists are already using AI with patients
Seven stories cleared the bar, and the one I keep coming back to is not about a new model. It is about clinicians quietly working AI into real care while policy and training try to catch up.
- In a survey of 278 Australian mental health professionals, 43% used general-purpose AI daily for admin work, and one in three used it daily for client-facing tasks.
- Documents reportedly show problems in Medicare's AI-assisted prior-authorization pilot.
- Researchers tested whether AI can turn clinical notes into language patients understand without changing the meaning.
The Big Story
Therapists are already using AI with patients
A mixed-methods study surveyed 278 Australian mental health professionals after interviewing 12 clinicians. Forty-three percent reported using general-purpose AI every day for administrative or clinician-support work. One in three reported daily use for client-facing tasks.
The uses included documentation and planning, along with translation and research. Some clinicians also used it for emotional-support tasks. They described AI as a supervised support tool that saved time and reduced cognitive load. They also named privacy, accuracy, client acceptance, weak workplace rules, and loss of human judgment as unresolved problems.
Mark's read: Mental health and pain care overlap too much to treat this as somebody else's issue. Your staff may already be using ChatGPT or another general tool for session prep, notes, or patient communication. Ask what they are using before you write a policy for a workflow that only exists on paper.
The Rounds
Medicare's AI prior-auth pilot is already raising questions
Newly surfaced documents reportedly show problems in Medicare's AI-assisted prior-authorization pilot. The available RSS item did not include the specifics from the full STAT report, so I am not going to pretend we know more than that headline supports.
Mark's read: If you fight prior authorizations for patients, keep an eye on this one. Faster decisions do not help if they simply produce denials faster or make the appeal harder to understand.
The Rounds
A faster, cheaper AI for routine decisions
A new class of "judgment model" skips text generation and returns a probability instead. The claim is that these models can be 20 to 200 times faster and cheaper than a standard language model for narrow decisions.
Mark's read: This is speculative and not ready for your clinic. But it is easy to imagine this kind of scoring inside practice-management software, flagging claims likely to be denied or patients likely to miss an appointment. Watch the direction. Do not buy the promise yet.
The Rounds
The people building AI are arguing about slowing down
MIT Technology Review covers a new round of public disagreement among AI lab leaders about how quickly frontier systems should advance. The concerns include cyberattacks, biological threats, and economic disruption.
Mark's read: This is broader context, not a clinical finding. The companies building the models underneath scribes, billing tools, and patient chatbots are openly arguing about whether they are moving too quickly. Remember that the next time a vendor describes every new AI feature as an obvious improvement.
In the Clinic
Can AI make your notes readable to patients?
A systematic review and experimental evaluation examined whether large language models can translate clinical notes into language patients can understand. The legal backdrop was Germany's Patient Rights Act, which gives patients access to their records.
Mark's read: Patient portals already expose plenty of raw clinical language. A good simplifier might reduce confused portal messages and help a patient understand the plan. The test is whether it improves readability without quietly changing the clinical meaning.
In the Clinic
AI-simulated patients for reasoning practice
An early study tested a multi-agent system that creates virtual clinical encounters for students to practice reasoning. The researchers examined performance, consistency, and the problems they encountered while building it.
Mark's read: Programs rarely have enough standardized-patient time. This could add practice reps, but it is a supplement. Students still need to gather a real story from a real person and feel what happens when the encounter does not follow the script.
Academic Corner
AI risk, fabricated references, and disclosure
One academic-AI roundup covers warnings from mathematicians, a test that exposed fabricated references, and the argument over how researchers should disclose AI use in their work.
Mark's read: The citation problem is the part you can act on today. If AI helped with a manuscript, literature review, or course resource, verify every reference against the actual paper before it reaches a student, reviewer, or editor.
Try This
Find where AI touched patient work
Ask your team one question: "Where did AI touch patient-facing work this week?"
Do not start with punishment. Ask about session prep and notes. Ask about portal messages, handouts, or translation. Then find out what else they tried because the normal workflow was too slow. Decide which uses need a safer approved path, more review, or a clear stop sign.
The last word
What are people already doing with AI in your clinic that nobody has written down yet? Hit reply and tell me.
Get the next issue.
One email a week. The AI that matters for clinical work, in plain language.