Six physician organizations publicly reject the claim that AI is "better informed" than doctors
October 5, 2026 · 5 items
Six physician organizations publicly reject the claim that AI is "better informed" than doctors
Medical Economics · October 1, 2026Practice operations
On Sept. 30 the AAFP, AAP, ACOG, ACP, ACS and AMA issued a joint statement saying claims that AI is inherently better informed than physicians, or that physicians can't be trusted to decide without consulting it first, "diminish physician expertise and risk undermining patients' trust." The statement landed a day after the health secretary praised AI second opinions.
The article notes the statement responds to public assertions from federal leadership and AI industry representatives that characterized AI consultation as essential — and implied malpractice risk for physicians who diagnose or prescribe without AI input.
Medical Economics summarizes the liability picture as currently asymmetric: following AI recommendations that deviate from the standard of care may increase liability when outcomes are adverse, which is expected to slow adoption until standards evolve.
The piece draws a usable line: low-risk operational uses (ambient documentation, summarization, billing) are widely acceptable, while decision support, ordering, renewals and patient messaging require rigorous validation beyond vendor claims.
Why it matters for an independent practice: This is the cleanest public cover a practice has for an AI policy that says "ambient scribe yes, autonomous clinical advice no." For MMR, it is also a content opening: a physician-authored post explaining why your clinic uses AI for paperwork and not for diagnosis is exactly the trust-building, physician-signed content AI summaries and patients both reward.
Clinicians say agent permissions should attach to the task, not the model — with stop rules and an audit trail
Healthcare IT News · October 2, 2026Practice operations
Dr. Nipa R. Shah (University of Florida College of Medicine at Jacksonville) frames the shift bluntly: "A summary or recommendation can be reviewed before it affects the patient. If an agent sends a message, changes an appointment, starts an authorization or places something into the chart, the error may have already reached the patient or changed the workflow."
Her rule of thumb is risk- and reversibility-based: retrieving records, checking whether a test was completed, routing a message and prepping material for review can run with independence; starting or stopping a medication, changing a treatment plan, diagnosing or giving clinical advice "should still involve the physician directly." Permissions attach to the task, not the model — "An AI system might be allowed to obtain a pathology report, but that does not mean it should also be allowed to interpret the report and change treatment on its own."
Dr. Talha Shaikh (New York Cancer & Blood Specialists/One Oncology) names four non-negotiables: physicians must know what the system is allowed to do, what information it used, what actions it already took — plus a clear audit trail, an easy override, and clear accountability when something goes wrong.
Joy Xu, UCLA medical student and lead author of the July 2026 npj Digital Medicine paper "AI agents in clinical practice: An evidence map," says agents remain unreliable for complex cognitive tasks like diagnostic and management deliberation, and that decisions vulnerable to bias should not be delegated.
Why it matters for an independent practice: This is a ready-made template for an independent practice's agent policy: write permissions per task (intake, records retrieval, recall reminders = autonomous; anything touching the chart or the plan = physician-approved), require a logged action trail from any vendor, and define the stop condition before go-live rather than after the first bad message goes out.
Apple is adding controls around macOS "Full Disk Access" specifically because AI agents now use it
TechCrunch · October 2, 2026Buildable AI
Apple announced additional controls around the macOS "Full Disk Access" permission, warning that increasingly capable AI agents make broad access to users' files, messages, mail and browsing history riskier.
The move follows a journalist's claim that Meta's Muse app on Mac read their private messages — a claim Meta disputed, per TechCrunch's Sept. 30 report.
The permission was originally designed so things like backups could function properly; the concern is that a setting built for utilities is now being granted to software that reads and acts on everything it can see.
Why it matters for an independent practice: Every front-desk or marketing Mac in a practice that has an AI assistant installed is a device where one toggle can expose downloaded records, patient emails and imaging exports — outside the EHR's access controls and outside any BAA. Audit which apps hold Full Disk Access on staff machines this week, and make that a standing question before any AI desktop tool is installed.
A 24-minute breakdown of Microsoft putting an always-on agent inside Outlook, Teams and Excel — and the habits that decide who gets value from it
Nate B Jones · October 2, 2026Buildable AI
The video covers Microsoft bringing an always-on AI agent built on OpenClaw — "Copilot Autopilot" — into Outlook, Teams and Excel, and argues the real question is not which model is smartest but what data your agent can see and how well you direct it.
Chapters indicate he walks through why 450 million Microsoft seats matter (02:17), Microsoft's edge in trust and distribution (10:13), auto routing and the cost of AI work (16:51), and where Anthropic and Meta challenge Microsoft for enterprise work (21:34).
Five stated takeaways are habit-shaped rather than tool-shaped: define what good looks like, give the model the right data, make useful work repeat, match reasoning to difficulty, and improve the next run.
Note this is the creator's own description, not a transcript — it tells you what the video covers, not what it concludes.
Why it matters for an independent practice: If a practice runs Microsoft 365, this agent arrives by default rather than by decision — so the two things to settle in advance are which mailboxes and SharePoint folders it can read (a PHI-exposure question) and which repeating jobs you actually want it doing, like drafting recall emails or reconciling a referral spreadsheet. The "make useful work repeat" framing is the leverage point: one well-defined recurring task beats a dozen one-off prompts.
Insilico nominates an AI-designed GIPR blocker as a muscle-sparing partner to GLP-1 therapy
Longevity.Technology · October 2, 2026Regenerative medicine
Insilico Medicine (HKEX: 3696) has nominated ISM1354, a small-molecule preclinical candidate, for obesity and related conditions such as type 2 diabetes.
The headline number: at the same dose, one of the new molecules delivered at least 18 times more drug into the bloodstream than a clinical-stage comparison compound, per the company.
The thesis is that pairing a GIPR antagonist — a drug that blocks the receptor — with GLP-1 therapies could help people lose fat while keeping muscle. Insilico flags the catch itself: several GIPR candidates have recently stumbled in clinical trials on toxicity and tolerability, so the test is whether a molecule can be both effective and safe.
This is a different bet from Lilly's retatrutide, described as an agonist that switches GIP and other receptors on; Insilico nominated an earlier GIPR antagonist, ISM0676, back in January.
Why it matters for an independent practice: For clinics and compounding partners already running weight-management programs, muscle preservation is the objection patients raise most — and the pipeline answer is still preclinical, with a named safety history of failures. That is the honest line to use in patient education now: the combination is promising, nothing is proven, and your program handles lean-mass protection with protein and resistance training today, not a future molecule.
