I just came back from a trip where I gave a talk on LLMs and AI to healthcare professionals as part of an Innovations in Cancer 2026 program. I expected curiosity. I did not expect the level of energy in the room.
The questions were not "how do we stop this?" They were "how do I start?" "What can I use safely?" "How do I know when to trust it?" "How do I teach my fellows?" "What should my clinic be doing now before the institution gives us a policy two years late?"
That is why I keep finding the standard AI-in-healthcare headline a little stale. You know the one: physicians are apprehensive, worried, skeptical, resistant. There is truth in it, but it is not the whole truth, and at this point it may not even be the most useful truth.
The AMA's 2026 physician survey found that more than four in five physicians now use AI professionally, more than double the 2023 rate. It also found real concern: privacy, validation, skill loss, liability, and the patient-physician relationship. That is not contradiction. That is what serious adoption looks like. The doctors are not anti-AI. They are asking the questions anyone responsible for another human being should ask before putting a probabilistic system into the workflow.
If anything, the room I saw was ahead of the headline.
The Interesting Mood Is Cautious Enthusiasm
I am surprised, again and again, by how much enthusiasm there is when clinicians get a practical introduction to AI.
Not enthusiasm in the Silicon Valley sense. Not "replace all clinicians with agents" enthusiasm. The energy is different. It is closer to relief mixed with impatience. People can see the administrative burden. They can see how much time is wasted searching, rewriting, summarizing, reformatting, and re-explaining. They can also see how many medical AI conversations are still trapped between breathless demos and institutional paralysis.
That gap is where clinicians are living right now.
The same AMA survey is useful because it refuses the caricature. Physicians are using AI, confidence is growing, and medical research summarization and clinical documentation are among the most common use cases. At the same time, physicians want safety and efficacy validation, clear liability frameworks, and a role in adoption decisions. The headline should not be "doctors fear AI." It should be "doctors are already using AI, and they want it governed like something that matters."
That is a much better starting point.
It also matches what I see in my own work. The tools that stick are not the ones that promise magic. They are the ones that remove a small amount of friction from a real day: finding a paper, cleaning up a paragraph, checking a reference, writing a quick script, turning a half-formed clinic workflow into a usable interface, or making a patient explanation less terrible.
The future of clinical AI may be spectacular. The present is mostly better plumbing. That is not a criticism. Medicine needs better plumbing.
The Evidence Layer Is Moving Closer to the Bedside
One of the more important announcements this cycle was Cedars-Sinai's partnership with OpenEvidence. The obvious read is that another health system is adopting another AI tool. The more interesting read is that clinical AI is moving from generic question-answering toward contextual evidence work.
OpenEvidence is already one of the medical AI tools I use and watch closely. The reason is not that it answers questions. Many systems answer questions. The reason is that its center of gravity is the medical literature, and the product keeps moving toward the actual shape of clinical reasoning: What is the evidence? How strong is it? What patient does it apply to? What changed recently? What do the data show when you are not just reading a sentence but trying to make a decision?
Cedars-Sinai describes the integration as a way to bring evidence into clinical decision-making with patient context from the EHR. That sentence contains most of the hard part. The value is not "AI can summarize UpToDate but with a different logo." The value is whether an evidence layer can know enough about the patient, the question, and the workflow to make a clinician faster without making them sloppy.
I also recently listened to the NEJM AI Grand Rounds episode with OpenEvidence's Travis Zack, and the theme that keeps recurring there is the right one: healthcare AI is not mainly a model problem. It is an evidence, workflow, implementation, and trust problem. If a tool cannot survive contact with the clinic, the model score does not matter.
That is the lens I would use for OpenEvidence. The question is not whether it can produce a fluent answer. The question is whether it can become a reliable evidence companion in the middle of a real clinical day.
The Patient Is Getting Their Own Notetaker
Ambient AI has mostly been framed as a clinician tool: record the visit, draft the note, save the doctor from the keyboard. That is still important. Documentation burden is real, and anything that gives clinicians more eye contact and less after-hours charting deserves attention.
But the more interesting shift may be patient-facing ambient AI.
Kin Health raised $9 million to build an AI notetaker for patients. The concept is simple: a patient records a medical visit, gets a summary, sees next steps, and can share the output with family or caregivers. TechCrunch reports that the app is intended to help patients parse medical advice and track what comes next.
That matters because patients forget things. Families are not always in the room. The plan changes across specialists. Instructions get buried in portals. A clinic visit can be emotionally loaded, especially in oncology, and the patient may leave with only fragments of what was said.
A patient-facing notetaker is not just a consumer app. It changes the information geometry of the visit. The physician's note is written for billing, communication, medico-legal documentation, and future care. The patient's version of the visit needs to answer a different question: What happened, what does it mean, and what do I do next?
That is powerful. It is also delicate. Consent, privacy, accuracy, clinical review, and discordance between the physician note and the patient summary are not edge cases. They are the product surface. The tool is only helpful if it makes the visit more legible without quietly creating a second medical record that nobody owns.
Still, this feels directionally right. Patients should not need a medical degree, a perfect memory, and three portal logins to understand what just happened to them.
My Own Stack Is Getting More Mobile
The tools I am actually using have shifted in a way that feels worth naming.
I am using Codex with mobile more. That is not a small change. The old version of coding with AI still assumed you were sitting at a desk, opening an IDE, and spending a block of time on the problem. The newer version is closer to dictating a small idea into a coding agent from your phone, checking the result later, and steering the next pass when you have a minute.
That fits the rhythm of medicine much better than the old developer workflow. A lot of clinical software ideas do not arrive while you are sitting in front of a blank repository. They arrive between patients, after a frustrating Epic click path, while rewriting the same message for the fifth time, or when you realize a spreadsheet has become a shadow EHR. Mobile access to a coding agent makes those moments capturable.
The same thing is happening in research. I have been using Paperpile's AI integration more because it lets the paper library connect to the AI assistant instead of forcing me to manually shuttle PDFs, citations, and context around. That sounds like a small convenience until you are working across dozens of papers. Then it becomes the difference between "I should read that later" and "I can interrogate the actual source right now."
This is where AI is becoming less theatrical and more useful. Not one giant assistant that does everything. A set of smaller bridges between the places where the work already lives: the paper library, the phone, the codebase, the notebook, the clinic note, the EHR, the meeting transcript.
I am increasingly convinced that the best personal AI stack is not the fanciest model. It is the stack with the fewest copy-paste seams.
For a more detailed version of this, see my May 2026 tech stack and the vibe coding guide for physicians.
Google I/O Was Really About the Agent Layer
Google I/O 2026 was the clearest signal yet that the next AI battleground is not the chat window. It is the layer underneath search, phones, documents, video, code, shopping, and personal workflows.
Google announced new Gemini models, deeper AI Search, information agents, generative UI, Gemini Spark, Daily Brief, upgrades across AI Studio and developer tools, and research features like Literature Insights built with NotebookLM. The details matter less than the pattern. Google is trying to make AI less like a destination and more like an operating layer.
For clinicians, that matters in two ways.
First, the general-purpose tools are going to arrive before the healthcare-specific governance. Physicians will use them for learning, summarization, writing, coding, and administrative work whether or not the hospital has an AI strategy. That is already happening.
Second, the boundary between "consumer AI" and "clinical-adjacent AI" is going to get blurrier. A patient may use AI Search to understand a biopsy report. A resident may use NotebookLM to study guidelines. A physician may use an agent to monitor papers in a narrow disease area. A clinic may use a coding agent to build an internal tool. None of that looks like a traditional FDA-cleared medical device, but all of it touches the medical information environment.
That is why clinicians need literacy now. Not because every doctor needs to become a prompt engineer. Because the default AI layer of the internet is becoming part of the context in which patients learn, clinicians work, and institutions make decisions.
Quick Hits
Kin Health is building a patient-facing AI notetaker. The important shift is not transcription. It is the possibility that patients leave visits with a clearer, shareable, action-oriented version of what happened. The hard questions are consent, privacy, accuracy, and who owns the summary when it affects care.
OpenEvidence and Cedars-Sinai are bringing evidence retrieval closer to the EHR. This is the kind of clinical AI deployment worth watching because it moves beyond generic chatbot behavior toward patient-contextual evidence work.
Google I/O pushed AI deeper into the operating layer. The most relevant healthcare point is not any single Gemini demo. It is that AI agents, AI Search, and generative interfaces are becoming ambient infrastructure for everyone, including patients and clinicians.
Paperpile now connects research libraries to AI assistants. This is the kind of quiet workflow feature that matters for academic medicine: less dragging PDFs around, more source-grounded reading and synthesis.
The AMA data are better than the anxiety headline. More than four in five physicians report professional AI use, but they still want validation, privacy protection, liability clarity, and a say in adoption. That is not fear. That is governance trying to catch up to use.
The Doctors Are Not Waiting
The thing I came back with from the cancer innovation meeting was not a new benchmark or product take. It was a mood.
Clinicians know AI is coming into the clinic because it is already in their lives. They are using it to read, write, summarize, teach, code, search, and think. They are watching patients use it too. They are appropriately cautious about anything that touches diagnosis, treatment, documentation, or trust. But caution is not the same as refusal.
The doctors are not waiting. They are asking for a map.
That is what the next phase of medical AI education should provide: not hype, not prohibition, but practical literacy. What is safe to use? What is not? What belongs in the EHR? What should never touch PHI? Which tools are good enough today? Which claims still need evidence? How do we teach trainees to use AI without outsourcing judgment?
If the room I spoke to is any indication, healthcare is ready for that conversation.
Until next time, reach out if any of this sparks something.
- Ramez