Your staff already found their own AI.

Some clinicians are quietly using consumer AI tools that their organizations never approved. They're drafting letters, summarizing notes, looking up clinical questions, and trying to get through a workload that still doesn't fit inside the day.

JMIR calls this "shadow AI." The report cites a survey of more than 500 health care professionals in which 58% said they had used free, general-purpose AI tools at work. That gap between what staff need and what an organization has approved creates problems with privacy and accuracy. It also leaves hard questions about cybersecurity, patient consent, and who is accountable when the output is wrong.

The Trump administration is pointing to AI, including AI-based avatars, as part of the answer for rural health care. Rural leaders interviewed by STAT were much less certain. They described tools that can reduce documentation burden and help with communication, but not the financial and staffing losses putting rural systems at risk.

OpenAI's GPT-6 Astra is getting mixed reviews. It appears unusually strong at computer use and 3D work, while trailing other models on some general reasoning and design benchmarks. That makes it less of an obvious replacement for the model you already use and more of a look at where agents are going.

A new viewpoint asks what ambient AI scribes could do to medical education. The concern is straightforward: if trainees use AI-written notes before they learn to build a note themselves, they may miss some of the clinical reasoning that documentation forces them to practice.

Researchers evaluated 27 chronic-pain apps relevant to Australian occupational therapy practice. The average objective quality score was 3.33 out of 5, rated acceptable, but scores ranged from poor to good. Engagement scored poorly on average, and the mean subjective quality score was 2.67 out of 5.

A narrative review of 21 publications mapped where large language models are being used in systematic reviews and meta-analyses. The authors propose roles for AI across six stages, starting with question development and searching. The framework also covers screening and data extraction, along with analysis, with retrieval and human oversight built into the workflow.

Reported accuracy across the included studies ranged from 61% to 99%, but that isn't one clean performance number. The studies tested different models on different tasks, and the proposed framework still needs real-world validation.

Ask your team one question this week: "Which AI tools are you already using for work?"

Don't start with punishment. Start with the actual workflow. What are they putting into the tool? What are they getting back? Where does patient information enter the process? You can't write a useful AI policy for an imaginary clinic.