The AI Clinic · Issue 010Clinical AI in the record
Nobody reads the whole chart anymore.
The chart got too thick to read. So doctors are asking it questions.
It's Monday, August 24. The thread this week is a gap. Stanford physicians are asking an AI what's buried in charts too thick to read in a visit. The same week, The Lancet pointed out that nobody actually tests these tools on that job.
The tools are moving into the room faster than the tests that would tell us to trust them. Most weeks that gap is abstract. This week you can see both sides of it.
- Health systems, Stanford among them, are rolling out AI tools that answer questions from the full patient record. Faster chart review, and a quieter way to miss things.
- A Lancet Digital Health comment argues medical AI benchmarks test exam knowledge, not real clinical work like reconciling a messy chart. The authors want resident-level task benchmarks instead.
- Google added a student hub to Gemini: flashcards, quizzes, study notebooks, syllabus reminders. Your students will find it before you do.
The Big Story
The chart got too thick to read. So doctors are asking it questions.
STAT reports on health systems deploying AI tools like ChatEHR, built at Stanford, that search and summarize a patient's entire record. Physicians are using them to pull answers out of charts too dense for any human to fully review during a visit.
The pitch is honest, which is what makes it work. Nobody reads the whole chart. There isn't time, and for a complex patient there may be years of it. A tool that answers "when did this patient last have imaging" in three seconds is solving a real problem, not an invented one.
Mark's read: The failure mode here is quiet. A wrong summary doesn't look wrong, it looks like a clean answer. If your EHR grows one of these features, and it will, decide now which details you always verify in the source note before acting. Med lists and imaging reports are my start. Treat the AI's answer like a student's chart review: genuinely useful, probably right, and checked before it changes your plan.
From Work, Understood
Not sure where the friction in your practice actually lives?
Ten plain questions, no email required, five minutes. It scores where the friction in your practice actually is instead of you guessing at it. If it points at something real, that's exactly the conversation a workflow audit is built to solve.
Research Watch
Every medical AI passes the exam. Nobody tests the job.
A Comment in The Lancet Digital Health argues that most medical AI benchmarks measure exam-style knowledge, vignettes and multiple choice, rather than the messy multi-step work of an actual clinical day, like reconciling conflicting data scattered through an EHR. The authors propose building benchmarks from the tasks residents actually do.
Mark's read: Read this next to the story above. The chart tools are already deployed. The tests that would tell us they're safe for that work don't exist yet. So when a vendor quotes an accuracy number, ask what task the number was measured on. Passing a board exam is not reconciling a med list at 4:45 on a Friday. Until the benchmark looks like the second thing, treat the score as marketing.
Second Opinion
One model to read every scan? The field is betting the other way.
A review of foundation models in medical imaging pushes back on the idea that one large model will soon handle every imaging task. The authors argue data scarcity and medicine's steady drift toward sub-specialization both point the opposite direction, toward more specialized tools, not fewer.
Mark's read: A useful counterargument from inside the field for the next time someone tells you a single AI will shortly read every scan type. The realistic near future is narrower tools doing specific jobs well. That's less exciting and easier to evaluate, which is probably good news.
In the Classroom
What nurse educators actually think of generative AI
A systematic review pulled together the research on nurse educators' own experiences with generative AI, an angle studied far less than what students are doing with it. The pattern: real enthusiasm and real time savings on one side, worry about assessment integrity and workload on the other.
For faculty: If you teach, this is a benchmark. The hesitation you feel about assessment integrity isn't a personal quirk, it shows up across the literature. So do the time savings. Worth a skim before your next curriculum meeting so you're arguing from data instead of impressions.
Try This
Spend fifteen minutes inside your students' new study tool
Google rolled out a student hub inside Gemini: study notebooks, flashcard and quiz generation, calendar reminders built from an uploaded syllabus, and Deep Research through Gemini Live. Your students will show up already using it. So get there first. Load one of your own lecture PDFs and have it build a quiz, then grade that quiz yourself. You'll learn two things fast: what the tool is genuinely good at, and where it flattens your material into generic test prep. Then your AI policy for the term can say something specific instead of something nobody believes.
The last word
One more if you have twenty minutes. A video making the case that knowledge work is shifting from doing tasks to directing AI that does them, with five skills for the shift: mapping what the tools can actually do, managing context, prototyping, spotting opportunities, and picking up new skills fast. I keep thinking about it as a staffing question. The hire that matters in an AI-heavy clinic may not be the person who does the task best. It may be the person who can direct the tool and catch it being wrong.
Has an AI chart tool, or an AI-briefed student, shown up in your week yet? Hit reply and tell me about it. I read every one.
Get the next issue.
One email a week. The AI that matters for clinical work, in plain language.