On September 25, Dr. Rachael Bedard, a geriatrician and palliative care physician, published a guest essay in The New York Times that opens with a confession: "A.I. is making me a better doctor. And I worry that it's making doctors-in-training worse." She uses a clinician-facing evidence tool several times a day. It helped her choose the most indicated antibiotic for one patient, interpret unusual bloodwork for another, and reminded her to consider migraine in her differential diagnoses for a woman who came in with dizziness. Many of us use AI the same way: as clinical thinking partners and decision support, not the decision maker.
Her worry is that her residents are relying too heavily on AI that they are missing out on actually practicing to become a great clinician. Adam Rodman, who directs AI programs at Beth Israel Deaconess's Shapiro Institute, gave her a distinction that has shaped much of this discussion. In experienced hands, decision support "might cause a little de-skilling." In a learner, "the cognitive offloading can cause never-skilling."
A 2025 NEJM review and a May 2026 Nature Medicine perspective treat skill erosion as a supervision problem medicine has to evolve around. Robert Wachter at UCSF doesn't see all skill loss as a bad thing, writing that "some of AI's productivity gains will come through humans ditching skills they no longer need." The catch, in his words, is that "we'll be working with AI tools that are good enough to be useful but not good enough to be entirely trusted." That gap falls hardest on learners, who haven't yet built the judgment they'd need to catch the tool when it's wrong.
Most of our readers are in Bedard's position: experienced, using AI every day, and starting to train newer providers. Rodman's distinction gives us two jobs. For ourselves, keep the de-skilling small by doing our own thinking before the tool does. For the people we bring on, make sure they build the skill in the first place, so AI sharpens their judgment instead of standing in for it.
What the survey data shows
In the AMA's 2026 physician survey, 81% of physicians reported using AI at work, up from 38% in 2023, and 88% said they were at least mildly concerned about losing skills. Our practices add their own load, with long visits, biomarkers tracked across years, and research that moves faster than guidelines, which is the "ever-evolving universe of facts" Bedard describes.
Two clinicians we spoke with recently described the same habit. One functional medicine practice owner told us what he wants: "What am I missing? Here's what I'm thinking. Tell me where I'm wrong. Tell me my blind spots." An integrative physician described checking the chart and labs, landing on "this is what I think's going on," and then asking, "what else am I missing here?" Both bring a hypothesis before the AI says anything, and its answer gets tested against that.
Experience helps, up to a point
In a 2023 Radiology study, radiologists read mammograms with an AI suggestion beside each one, and on some cases the researchers made the AI wrong on purpose. When the AI was right, readers at every experience level got about 80% of cases right. When it was wrong, the most experienced radiologists got only 45.5% right, and the newest got 19.8%, about one in five. The researchers call this automation bias: trusting the machine's answer over your own. Experience made it weaker but didn't make it go away.
A March 2026 randomized trial in npj Digital Medicine gave 70 physicians six diagnostic cases. With AI, they got 82% to 85% right, compared with 75% using their usual references. The order barely mattered. Physicians who saw the AI's answer before forming their own did slightly better than those who worked the case alone and then checked it. But the AI in this trial was mostly right, and in practice nobody flags which answers are the wrong ones. The radiology study shows what happens then. Writing your own read before you prompt won't raise your score on a good day. It's what keeps you sharp enough to catch the bad one.
The closest thing we have to long-term data comes from colonoscopy. In a 2025 study of 19 endoscopists who used AI for months to help spot precancerous polyps, their detection rate in colonoscopies done without AI fell from 28.4% to 22.4%. That's about a fifth fewer polyps found. Lean on a tool long enough, and the clinical skills that can only develop from practicing, can fade.
None of these studies looked at functional or longevity practices, and two involved imaging and endoscopy AI rather than the reasoning tools most of us use. So we don't read the numbers as predictions for our practices. We read them as signals that point the same way: AI makes us better when it's right, pulls us off course when it's wrong, and skills we stop practicing can fade.
Not every skill is worth protecting. Wachter points out that "very few of us know how to maneuver a horse-drawn carriage or use an abacus these days," and no one misses them. He asks three questions to decide: how likely you are to need the skill again, what happens if you need it and don't have it, and what it costs to keep it up. Formatting a lab summary or drafting a patient handout can go. Building a differential, reading a panel against a patient's history, and noticing when an answer doesn't fit can't.
The cost in time
A concierge physician told us he had about 40 results in his inbox one weekend and spent four hours on two of them doing deep dives with AI. He called it "great for quality," and he also gave up most of a weekend to it. Writing out your hypotheses, differential, and working plan before you open the tool helps you decide which results deserve that kind of deep dive.
Where never-skilling shows up for us
Many of us are onboarding someone, whether new to clinical practice or new to functional medicine. One practice owner estimated a new provider "might take them like five years to kind of get good in an IFM model because it's just very nuanced." With AI, that provider can produce a plausible protocol in the first week. Whether she can explain or defend her clinical reasoning behind it to the patient takes longer, and teaching her that skill is on us.
It is also why AI-drafted text in our platform is visibly marked in the chart for clinician review. One practice owner saw it during a demo and said, "it's very clear what was generated... It's on me or the provider, whoever the provider is, to review it."
The label shows what the AI wrote. It can't show whether the person signing off understands why. That understanding gets built the slow way, case by case, with someone experienced checking the work. They call it clinical practice for a reason.
What to Try This Week
Write out your clinical thinking before you prompt. One line in your working note: your leading hypothesis and the one finding that would change your mind. Then open the tool, and keep anything with patient details inside a HIPAA-compliant clinical platform. When the AI disagrees, that line tells you where to look. It also tells you which results deserve a deep dive and which you can sign off on your own.
Ask for blind spots. Prompt for what you might be missing, what argues against your read, and which single test would most change your plan. When the AI is wrong, a confident answer pulls you toward it, and the mammography data shows how hard it pulls.
Open one source. For any answer that will change a protocol, click through to the cited study and check the population and design. Longevity evidence is uneven, and a confident sentence from AI can rest on a mouse study or a population decades younger than the patient sitting in front of you.
Keep AI-drafted text marked until a clinician signs it. Lab parsing counts. For the first few weeks, spot-check extracted values against the original report, the way one physician told us she plans to by "initially double checking." When a newer provider writes an AI-assisted protocol, the supervising clinician reviews it before it reaches the patient.
Run one AI-free case review a month with newer providers. Rodman suggested AI-free clinical reasoning sessions for learners, and a monthly one fits a small team. A script to open it: "Before anyone opens a tool, tell me what you think is going on and what you'd order. Then we'll see what the AI adds and where it disagrees with you." Close by having them explain the plan the way they'd explain it to the patient.
Pick one skill to protect and one to hand off. Run your own work through Wachter's three questions: how likely you are to need the skill, what happens if you don't have it, and what it costs to keep up. Hand off the low-stakes work, and keep practicing the reasoning you can't afford to lose.
Be a modern clinician with the help of Ultralight, the AI-native EHR built specifically for functional, integrative, and longevity medicine. We've recently launched wearables integrations and improved AI-native clinical workflows. Get in touch to see the latest updates.
Live October 8: Advanced AI Prompting for Functional & Longevity Clinicians
This issue is about bringing your own read before you prompt. On October 8, we'll take it further in a live session for functional, integrative, and longevity clinicians on advanced prompting with Ultralight AI.
Live · Thursday, October 8 · 11am PT / 2pm ET
Hosted by Sunita Mohanty, CEO of Ultralight, and Karina Tsareva, Practice Ops at Ultralight, with panelists Marwa Shoeb of Marwa Shoeb MD, Sheila F. Marshall of Next Chapter Medicine, and Lyssa Jaye of FLIMD.
In the news
Lowering Lp(a) did not lower cardiovascular events in the first large outcomes trial. Novartis reported top-line results from Lp(a)HORIZON on September 4: pelacarsen lowered Lp(a) in 8,323 patients with elevated Lp(a) and established cardiovascular disease, but did not significantly reduce major adverse cardiovascular events. Announced in early September and worth the late look, with full results still to come. Lp(a) stays useful for sorting risk. For a patient waiting on a drug to fix her number, the conversation comes back to apoB, blood pressure, and the risk factors we can move today.
The FDA wants input on how to judge generative AI clinical tools, and comments close October 19. In an August 18 discussion paper, the agency's device center proposed a "competency-based" review for generative AI medical devices, modeled on how clinicians are credentialed, with periodic benchmarking and clinician review after a tool reaches the market. If AI is part of your clinical reasoning, this is a chance to tell regulators what good enough to use should mean before the rules are set.
A Nature Aging study of 104,208 adults found that men and women hit their biggest aging shifts at different ages. Researchers built sex-specific aging clocks from 172 routine clinical lab values at three Chinese medical centers, published August 25 and worth the late look. Men showed a metabolic shift between 35 and 45, with changes in BMI, triglycerides, and cholesterol. Women showed theirs between 55 and 65, coinciding with menopause and worsening lipids. The cohort is entirely Chinese, so treat the ages as a rough guide. It still supports trending a patient's standard panel over the years rather than reading each one on its own.
Upcoming Conferences & Events
Oct 6, A Clinician's Masterclass on VO2 Max · Virtual · Free · Live session with Brooks Leitner, MD PhD of VO Health and Sunita Mohanty of Ultralight on bringing VO2 max into clinical practice. Ultralight will be there!
Oct 8, Beyond the Basics: Advanced AI Prompting for Functional & Longevity Clinicians · Virtual · 11am PT / 2pm ET · Live session hosted by Sunita Mohanty and Karina Tsareva of Ultralight, with panelists Marwa Shoeb, Sheila F. Marshall, and Lyssa Jaye, on advanced prompting with Ultralight AI. Ultralight is hosting!
Oct 8-10, A4M Women's Health Summit · San Antonio, TX · The best clinical education on hormone, metabolic, and midlife women's health you will see this year. The room to be in if you are growing the perimenopause and menopause side of your practice.
Oct 17-18, Roundtable of Longevity Clinics · Buck Institute, Novato, CA · Some 250 longevity clinic leaders, physicians, and researchers working toward shared standards for longevity testing and interventions. In-person and virtual tickets are open. Ultralight will be there!
Oct 21-23, DOC (Living Room Lab) · Sonoma, CA · Salon-style sessions on longevity science and medical AI with faculty from UCSF, Stanford, and the Buck, plus validated diagnostics in the Living Room Lab. Ultralight will be there!
Oct 21-24, NAMS Annual Meeting · San Diego, CA · The single most practice-changing meeting of the year for midlife women's health. Your protocols will look different after this one.
Nov 5-8, Eudēmonia Summit · West Palm Beach, FL · One of the most talked-about longevity gatherings in the U.S. With experientials, hands-on demos, and the best place to try the emerging frameworks your patients will ask you about next year. OvationLab and Ultralight will be there!
Nov 5-7, Private Physicians Alliance Annual Meeting · St. Petersburg, FL · The gathering for independent, cash-pay, and concierge physicians navigating practice independence. Practical and peer-driven. Ultralight will be there!
Nov 8-11, American College of Lifestyle Medicine Conference · Orlando, FL · Lifestyle medicine's main annual event, with evidence-based approaches to behavior change, chronic disease, and healthspan. Growing overlap with the longevity medicine community.
Nov 10-13, Valley Forum 2026 · Napa Valley, CA · Invitation-only forum where healthcare leaders work through the industry's hard problems, AI's role in care among them. Ultralight will be there!
Dec 11-13, A4M Longevity Fest · Las Vegas, NV · The biggest longevity event in the U.S. The room spans clinicians, industry, founders, and the people building next year's platforms, and the connections from this one tend to compound through the rest of your year. OvationLab and Ultralight will be there!
Know of an event we should add? Reply and tell us.
Until next week
AI makes us sharper when our own read of the patient stays in the loop. Keeping it there is now part of running a practice, for us and for every provider we bring on.
Training someone new this year? Forward this to them, or reply and tell us how you teach clinical reasoning with AI in the room.
Until next week, keep building the practice you imagined when you started.
— Dr. G and Sunita