Four in five physicians already use AI. The skill that separates them is knowing what to hand off and what to sign.

AI is not competing with your clinical judgment. It is competing with your clipboard.

The benchmark time-and-motion study still frames the problem. Across 57 ambulatory physicians in four specialties, Sinsky and colleagues found 27% of the day went to direct face time with patients and 49% to EHR and desk work. Even inside the exam room, 37% of time went to the computer, and physicians who kept diaries reported one to two hours of after-hours work each night. It has not improved. A 2024 follow-up at an academic system found total EHR time up 7.8% and inbox time up 24% between 2019 and 2023.

One in three adults asked a chatbot about their health this year, and many never followed up with a doctor

A KFF poll fielded in March found 32% of U.S. adults used an AI chatbot for health information in the past year. Among those using it for physical health, 42% did not follow up with a clinician, and 41% of all users uploaded personal medical information. OpenAI says more than 230 million people ask ChatGPT health questions every week and launched a consumer ChatGPT Health product in January that connects to medical records.

The ambient scribe evidence is in, and it is smaller and more real than the marketing

Two randomized trials published in the past year both found benefit, none of it dramatic. At UCLA (238 physicians, about 72,000 encounters), Nabla cut documentation time per note by 9.5% against usual care, Microsoft DAX did not reach significance, and burnout improved roughly 7% in both arms. A UW Health stepped-wedge trial of Abridge found burnout down about 20% and 30 fewer minutes of work outside work. The largest real-world study, in JAMA this April, followed about 1,800 adopters across five academic systems: 16 fewer documentation minutes per eight hours of patient time, about half a visit more per week, and no significant change in after-hours EHR time. Only 32% of adopters used the scribe in at least half their visits.

Disclosure: the editor serves as Clinical Director at Nabla Technologies, one of the products in the UCLA trial. The pilot figures are internal and unpublished.

The consumer chatbot on your phone is not a place for patient information

Free and personal tiers of ChatGPT, Claude and Gemini do not come with a business associate agreement, and OpenAI says it will not sign one for its consumer ChatGPT Health feature. The healthcare versions launched this year do: OpenAI for Healthcare (January 8) offers a BAA for ChatGPT for Healthcare and its API, Anthropic covers Claude Enterprise plans, and Google's Workspace BAA covers only the Gemini features on its HIPAA-included list. Texas now also requires electronic health records to be stored in the United States as of January 1, 2026.

A chatbot answers. An agent works the whole problem.

Most clinicians have only used the first kind. A chatbot needs a prompt at every step. An agent is given a goal and runs the workflow: check eligibility, draft the prior authorization letter from your last twenty approved ones, route the referral, and follow up if the payer has not answered in 48 hours. Tools that let a clinician describe a workflow in plain language and get a working version back are now good enough that a prior-auth builder or a post-op instruction generator tuned to your own protocol is an afternoon project, not an IT ticket.

In two Stanford-led trials, doctors with GPT-4 did no better than GPT-4 alone

In a randomized trial of 50 physicians working through diagnostic vignettes, those given GPT-4 scored a median 76% against 74% for those using conventional resources, a difference that was not significant. GPT-4 working alone scored 92%. A follow-up in Nature Medicine with 92 physicians on management reasoning found GPT-4 assistance did help, by 6.5 points, but physicians with the model still did no better than the model by itself. A March 2026 study found a better-designed workflow, with the AI giving a first or second opinion, closed much of that gap.

When the AI is wrong, experienced clinicians follow it too

In a mammography study, very experienced radiologists read at 82% accuracy on their own. When shown deliberately incorrect AI suggestions, they fell to 46%, and less experienced readers fell to about 20%. In a 2025 trial, 44 physicians trained in AI literacy received GPT-4o advice with errors planted in half the cases, and their accuracy dropped 14 points when the advice was wrong. Separately, an observational study at four Polish centers found endoscopists' adenoma detection without AI fell from 28.4% to 22.4% after AI assistance became routine.

Even the newest models invent references, and the orthopaedic data is recent

When ChatGPT-5 was asked to support 70 recommendations from AAOS clinical practice guidelines, it produced 2,736 references. About 7% were fabricated outright, only 49% were fully correct, and PubMed IDs were wrong 37% of the time. That is an improvement on 2023, when a Cureus study found 47% of ChatGPT-3.5's medical references did not exist. It is not good enough for a letter to a payer or a slide at a society meeting.

One in twenty AI scribe notes carried an error serious enough to matter

UC Davis audited 356 of 7,545 notes generated by 31 physicians. Omissions appeared in 18%, hallucinated content in 11.5% and wrongly included material in 9.3%. Most errors were mild to moderate, but 5.3% of notes contained an error rated as posing serious risk. In a simulated test of five commercial platforms, notes averaged a 26% error rate. The most common real-world failures at UCLA were omissions and misattributed pronouns.

72% of Americans want to be told when AI is part of their care, and almost half do not know if it already is

Pew surveyed 3,488 adults in June. The share who want to be told was highest for scans and diagnosis (81% each) and was still 72% for note-taking. Forty-six percent were not sure whether AI had been used in their care. The wording matters: in a Duke study of AI-drafted portal replies, patients rated the drafts slightly higher than human-written ones, and their preferred disclosure was "written by Dr. T with the support of automated tools." A separate JAMA Network Open study found any mention of AI in a physician's advertising slightly lowered ratings of trust and empathy.

Disclosure is already the law in Texas and California, and more states are coming

Texas has the strictest rules for clinicians. SB 1188 (effective September 2025) requires practitioners who use AI in diagnosis or treatment to stay within scope, review AI-created records and disclose the use to patients. TRAIGA, in effect since January 1, requires that disclosure by the date of service. California's AB 3030 requires a disclaimer on AI-generated patient clinical communications unless a licensed provider reviews them first. Utah requires disclosure in high-risk interactions. Colorado replaced its stalled AI Act with a notice-based law that takes effect January 1, 2027. A December 2025 executive order seeks federal preemption, but Congress has not acted, so state laws still apply.

APMA's position: AI should support, not replace, physician judgment

APMA adopted a formal AI position statement in March and filed comments with HHS. It calls for disclosure when AI is used in coverage decisions, opposes opaque algorithms that drive denials and downcoding, and asks that podiatric physicians be included in AI design and regulation. It raised AI claims review directly with Aetna on March 30, alongside skin substitutes and modifier 25. The Joint Commission and the Coalition for Health AI issued voluntary guidance last September that also asks organizations to tell patients about AI use, and the Federation of State Medical Boards has said since 2024 that the physician remains accountable for AI-informed decisions.

Scribes are pushing E/M levels up, and payers are pushing back

At the five systems in the April JAMA study, scribe adoption added about $167 a month in E/M revenue per clinician. Trilliant Health found the share of established visits billed at 99214 or 99215 rose 7 to 12 points at six systems as AI scribing spread. Blue Cross Blue Shield Association estimates AI-enabled coding added about $2.3 billion in spending. Cigna began automatically downcoding level 4 and 5 E/M claims last October. Podiatry is already under a spotlight: OIG found 44 of 100 sampled podiatrist E/M claims noncompliant, with modifier 25 at issue.

Medicare's AI prior-authorization pilot covers skin substitutes for lower-extremity wounds

CMS's WISeR model began January 1 in traditional Medicare in New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington, with a different AI vendor in each state. Its 14 service categories include skin substitutes for lower-extremity chronic wounds (in states with an active coverage policy), electrical nerve stimulators and epidural steroid injections. The House Appropriations Committee voted in June to defund it, the Senate declined to repeal it in July, and records obtained by STAT last week describe a rushed launch and delayed decisions.

No AI malpractice verdict yet, and the physician still holds the pen

No U.S. malpractice verdict has turned on AI so far. The Doctors Company says it has no AI exclusions and has seen few AI-specific claims, and the consistent advice from carriers and boards is to vet tools, document the AI's role and your own reasoning, and have a disclosure protocol. Most clinical AI pilots that fail do so because of leadership, not technology. Adoption stalls when the organization does not reward it, clinicians are not trained, or the tool does not fit the workflow.

AI measures hallux valgus angles as well as surgeons in a fraction of a second, but fracture detection is less settled

A deep-learning model trained on 2,468 weightbearing radiographs matched surgeons on HVA (ICC 0.90) and IMA (ICC 0.89) and had fewer IMA outliers, at 0.13 seconds per image. For fractures, a 14-study meta-analysis pooled 93% sensitivity and 94.5% specificity. In a single-center head-to-head published last month, however, a commercial FDA-cleared tool reached 74% sensitivity on 701 foot and ankle exams against 84% for musculoskeletal radiologists, and struggled most with Chopart and Lisfranc injuries.

Diabetic foot AI is promising in the lab and mixed in outcomes

A UK smartphone model detected diabetic foot ulcers with 91.6% sensitivity in real NHS clinics using inexpensive Android phones. Remote temperature monitoring predicted ulcers weeks early in its original validation, but a VA evaluation of 924 veterans found no reduction in amputation or hospitalization. Spectral AI's DeepView received FDA De Novo authorization in May for burns only, and its diabetic foot program is still in study.

Chatbot answers about bunion surgery read at a college level

A June study in Foot and Ankle Surgery compared ChatGPT 5.2 and Gemini on ten common patient questions about minimally invasive bunion surgery. Quality scores were similar, readability was college level for both, and neither was very actionable. An AOFAS abstract found AI-written patient education averaged a grade 13.3 reading level against 9.5 for surgeon-written material, with lower quality and actionability. A Foot & Ankle Specialist study found GPT-4 gave little or no surgical detail in 80% of responses on 15 foot and ankle conditions.

FDA has authorized more than 1,400 AI devices, and generative AI rules are still open for comment

The FDA's AI-enabled device list passed 1,400 in March after a record 331 authorizations in 2025. In January it loosened its clinical decision support guidance for tools that give a single recommendation a clinician can independently review. In August it published a discussion paper on regulating generative AI devices, with comments due October 19. There is still no formal FDA framework for large language model clinical tools.