Featured

Labels

Dr. Algorithm Will See You Now: The Trump Administration’s Great AI Health-Care Experiment


I have some exciting news for anyone who has ever waited six weeks to see a specialist, three hours in an urgent-care lobby, or seventeen minutes on hold while an insurance company explained that your suffering was not formatted correctly.

The federal government has apparently located the cure for American health care.

It is artificial intelligence.

Of course it is.

We could have guessed this was coming. When a country has too few doctors in many communities, exhausted nurses, rural hospitals fighting to remain open, medical bills that resemble ransom notes, electronic records that refuse to communicate with one another, and insurers that can turn the word “no” into a forty-page administrative process, the obvious response is to add a chatbot.

The Trump administration is moving to push AI deeper into medical care, presenting the technology as a tool that can modernize a system almost everyone agrees is expensive, fragmented, slow, and unnecessarily hostile to the people trapped inside it. The broad vision is easy to sell: connect health information, reduce paperwork, give patients more control over their records, help clinicians make sense of oceans of data, and use AI-powered applications to guide people through chronic conditions and everyday medical decisions.

On paper, this sounds wonderful. Then again, on paper, my insurance plan says I have coverage.

I am not opposed to AI in medicine. Quite the opposite. I think the technology could become one of the most useful tools modern health care has ever adopted. It can help doctors summarize visits, flag dangerous drug interactions, detect patterns in scans, reduce clerical work, translate medical language, and help patients arrive at appointments better informed. If AI can stop a physician from spending half the visit staring into an electronic record designed by people who apparently hated both medicine and computers, I am prepared to applaud.

But I also know how Washington and Silicon Valley behave when they discover a shiny new object at the same time. Washington wants a historic announcement. Silicon Valley wants market share. Hospital executives want efficiency. Insurers want lower costs. Investors want growth. Every institution arrives with a dream, a slide deck, and a financial incentive.

The patient arrives with a body.

That difference matters.

The Plan Sounds Great Because All Plans Sound Great Before They Meet Reality

The administration’s health-technology campaign has emphasized interoperability—the ancient and apparently mystical idea that patients should be able to access their own medical information and move it between systems without hiring a private investigator. The federal push has brought large technology companies, health-care organizations, and digital-health businesses into an effort to build apps and services around shared data standards.

Some of those applications are supposed to help people manage conditions such as diabetes and obesity. Others may use conversational AI to answer questions, organize care, simplify intake, or help patients navigate a system whose natural state is administrative fog.

I understand the appeal. American medicine contains enough paperwork to qualify as a renewable resource. A patient can disclose the same allergies to four different employees in one afternoon, only to discover that the specialist never received the records sent three weeks earlier. A doctor can spend years in medical school and then devote a large portion of the workday to clicking boxes so a billing system feels emotionally supported.

If AI can reduce that waste, bring scattered records together, and give clinicians more time with patients, then let it work. Give it a keyboard. Give it a tiny digital stethoscope. Name it Employee of the Month.

The problem is that “AI in health care” covers an enormous range of uses, and public discussion tends to toss them into one glittering bucket. An automated transcription tool is not the same as a system recommending a diagnosis. A scheduling assistant is not the same as an application interpreting symptoms. Software that reminds a patient to take prescribed medication is not the same as software deciding whether that patient needs urgent care.

The stakes rise sharply as an AI system moves from organizing information to influencing clinical judgment. A typo in a restaurant recommendation might send me to a disappointing sandwich. A fabricated medical claim might delay treatment, encourage the wrong treatment, or persuade someone that chest pain is merely the body expressing a creative difference of opinion.

That is why I become suspicious when deployment is described with the breathless language of a product launch. Medicine is not a beta test in which the most inconvenient outcome is having to reinstall the app.

We Are Still Arguing About the Guardrails While the Car Is Leaving the Driveway

The timing is particularly remarkable. The Food and Drug Administration has acknowledged that generative-AI-enabled medical devices may create risks that differ from those posed by traditional software and earlier AI tools. In August 2026, the agency released a discussion paper seeking feedback on risk assessment, premarket evaluation, and postmarket monitoring.

Read that sequence again. The federal government is promoting broader AI integration while the federal regulator responsible for medical-device safety is still asking foundational questions about how some of these products should be reviewed and watched after release.

Nothing inspires confidence like installing the diving board while the committee studies whether the pool needs water.

To be fair, regulation often develops alongside technology. If government waited until every uncertainty disappeared, we would still be debating whether electricity looked trustworthy. Medical innovation always involves some risk, and excessive caution can carry its own human cost. Delaying a useful diagnostic system may mean delaying real benefits for real patients.

But there is a canyon between demanding perfect certainty and insisting on basic evidence. We should know what a clinical AI tool is designed to do, what data shaped it, how it performs across different populations, how often it fails, what kind of failures occur, who monitors it, how updates change it, and who becomes responsible when its advice harms someone.

Those questions are not anti-innovation. They are what responsible innovation looks like after it grows up and stops using “disruption” as an excuse for leaving broken furniture everywhere.

President Trump has recently dismissed prominent concerns about advanced AI and argued that excessive restrictions could weaken the United States in its competition with China. That instinct—go faster because a rival may go faster—is familiar in technology policy. It is also an awkward foundation for bedside care. I do not want my medical treatment calibrated to a geopolitical footrace. I want it calibrated to whether the tool works.

National competitiveness matters. So does surviving the appointment.

AI Can Be Brilliant, Wrong, and Completely Unembarrassed

One reason humans find generative AI so persuasive is that it can communicate uncertainty with the tone of a person announcing train times. It may produce an answer that is polished, orderly, and false without experiencing even a moment of hesitation. Confidence is part of the interface, not proof of the conclusion.

In ordinary life, this can be annoying. In medicine, it can be dangerous.

An AI assistant might overlook a rare condition, misunderstand an incomplete record, confuse correlation with causation, or generate information unsupported by the evidence it was given. It might perform well in a controlled evaluation and then drift as clinical practices, patient populations, or the underlying model change. It may struggle with ambiguous symptoms, multiple chronic illnesses, unusual medication combinations, or the simple fact that human beings are terrible at describing what hurts.

And then there is bias. Medical data reflects the health system that created it. If certain communities have historically received less testing, later diagnoses, undertreatment, or inconsistent access to specialists, their records do not provide a neutral portrait of biological reality. They also contain the fingerprints of unequal care.

Train a system on that history without careful correction and it may automate yesterday’s failures with tomorrow’s processing power.

The machine does not need to hold prejudice in any human sense. It only needs to learn patterns from a system that already distributes attention unevenly. An algorithm can be completely indifferent and still produce discriminatory results. In fact, indifference at scale is one of automation’s specialties.

This is why performance averages are not enough. A tool that looks impressive across a large population may perform poorly for a smaller racial group, for women, for older adults, for people with disabilities, or for patients whose symptoms fall outside the neat categories favored by the training data. “Mostly accurate” offers little comfort when you belong to the “mostly” that researchers studied poorly.

I want independent validation, subgroup testing, transparent limitations, continuous monitoring, and a clear process for reporting harm. I want clinicians to know when a recommendation came from an automated system and patients to know when AI materially influenced their care. I want meaningful human review, not a ceremonial human stationed nearby to absorb legal blame after the machine has already shaped the decision.

Privacy: Please Accept These Terms Before Discussing Your Tumor

The privacy question may be even more complicated than the accuracy question.

AI becomes more useful when it has context. In health care, context means diagnoses, medications, scans, laboratory results, genetic information, reproductive history, mental-health treatment, substance-use records, location data, sleep patterns, and whatever embarrassing detail you confessed because a doctor told you it mattered.

That is not ordinary consumer data. It is an intimate map of a person’s vulnerabilities.

The administration’s effort to improve data sharing could help patients escape the digital walls separating hospitals, physicians, pharmacies, and insurers. I support that goal. People should be able to retrieve their records without fax technology playing a starring role in the process.

But easier movement of data creates more routes through which data can be copied, analyzed, sold, breached, or used for purposes a patient never meaningfully understood. Health apps may operate under different privacy rules depending on what they do, who provides them, and whether they qualify as covered entities or business associates under federal law. Many consumers hear “health information” and assume “HIPAA protects everything.” Reality, as usual, has read the terms and conditions.

HHS states that the HIPAA Security Rule requires regulated entities to use administrative, physical, and technical safeguards for electronic protected health information. Those protections matter. They do not magically answer every question created when consumer technology, clinical platforms, insurers, data brokers, and AI companies begin exchanging increasingly detailed information.

Who receives the data? How long is it kept? Is it used to improve a commercial model? Can it be combined with nonmedical information? Can a patient delete it? Can consent be withdrawn? Does refusing permission reduce access to care? What happens when a company is acquired, goes bankrupt, or quietly changes its privacy policy on page thirty-seven?

Consent is not meaningful when the alternative is being unable to use the service. Nor is it meaningful when a patient is handed a dense agreement during a medical crisis and asked to tap “accept” with the speed normally reserved for cookie banners.

If the government wants public trust, privacy cannot be a footnote under the download button. It has to be part of the architecture.

The Great Efficiency Miracle—and Who Gets to Keep the Savings

Much of the enthusiasm around medical AI rests on efficiency. That makes sense. Clinicians are drowning in documentation. Prior authorization consumes time that could be used for care. Health systems generate vast amounts of information but routinely fail to place the right information in front of the right person at the right moment.

AI could help.

But whenever I hear that a new technology will make health care more efficient, I ask a vulgar but useful question: efficient for whom?

If an AI tool saves a physician two hours of administrative work each day, will those hours become more time with patients? Or will administrators schedule more appointments, reduce staff, and convert every saved minute into a new productivity target?

If an automated system helps an insurer process claims faster, will patients receive quicker approvals? Or will the company discover it can deny care at a velocity previously unavailable to human civilization?

If hospitals save money, will they lower prices? Improve staffing? Support rural clinics? Or will the savings ascend through the organization until they achieve executive-compensation form?

Technology does not determine how its benefits are distributed. Institutions do. AI may reduce work, but it does not decide whether workers receive relief. It may lower costs, but it does not decide whether patients see the savings. It may expand access, but it does not decide whether underserved communities receive the best systems or merely the cheapest automated substitute for a human professional.

That last possibility worries me most. Wealthier patients may use AI as a supplement to attentive physicians, specialists, and concierge services. Poorer patients may be handed AI as a replacement for the people the system never hired.

One group gets a doctor enhanced by artificial intelligence. The other gets a login screen explaining that all representatives are currently assisting other patients.

That is not democratized medicine. It is two-tier care with better branding.

Doctors Need Help, Not Another Invisible Boss

I have no interest in romanticizing the current system. Doctors make mistakes. Clinicians miss patterns. Human judgment is affected by fatigue, incomplete information, time pressure, and bias. The relevant comparison is not flawless doctors versus flawed algorithms. It is real clinicians and real tools operating under real constraints.

Used well, AI can act as a second set of eyes. It can surface possibilities a clinician might have overlooked, summarize a complex history, flag changes in a patient’s condition, and handle clerical work that contributes to burnout. Those are meaningful gains.

Used badly, it can become an invisible supervisor whose recommendation is difficult to challenge. Clinicians may defer to it because they trust the model, because the hospital expects compliance, because overriding it requires extra documentation, or because a future lawsuit will ask why they ignored “the system.”

Automation bias is not solved by placing a human somewhere in the workflow. The human must have enough time, information, authority, and training to disagree.

That means hospitals cannot treat AI literacy as a fifteen-minute webinar followed by a multiple-choice quiz everyone completes while eating lunch. Clinicians need to understand the tool’s intended use, limits, common failure modes, and evidence base. They need to know when it is likely to be unreliable. They need access to explanations that are genuinely useful, not decorative technical theater.

Patients also deserve an appeal path. If an algorithm contributes to a denial, triage decision, diagnosis, or treatment recommendation, someone must be accountable for reviewing the result. “The computer said so” was an irritating excuse when it prevented a refund. It becomes intolerable when it prevents care.

Accountability Cannot Be an Easter Egg

Here is the question every grand AI initiative eventually tries to leave at someone else’s office: who is responsible when the system causes harm?

The developer may say the model only provides information. The hospital may say the clinician made the final decision. The clinician may say the approved tool was integrated into the standard workflow. The insurer may say it relied on a vendor. The vendor may point to warnings in the user agreement. The government may explain that innovation requires flexibility.

Meanwhile, the patient is still injured.

Responsibility has to be assigned before deployment, not improvised afterward by a relay team of lawyers. Health systems should document when and how AI influences care. Vendors should disclose material limitations and report serious failures. Regulators should have access to postmarket performance data. Patients should have practical routes to correction and compensation. Whistleblowers should be protected when they expose unsafe systems.

And no, proprietary software cannot become a magic phrase that ends public scrutiny. A company should be allowed to protect legitimate intellectual property. It should not be allowed to hide evidence of safety, bias, or failure behind a curtain marked “trade secret” while its product helps determine whether people receive treatment.

If an algorithm is important enough to influence my care, it is important enough to audit.

What I Would Demand Before Calling This Progress

I do not want the federal government to abandon medical AI. I want it to approach the technology with ambition disciplined by evidence. That requires more than optimistic announcements and voluntary promises.

First, AI systems that influence diagnosis, treatment, triage, or coverage should face risk-based evaluation before widespread use. The greater the potential harm, the stronger the evidence required.

Second, testing must reflect the diversity of the population that will actually use the system. Results should be reported across relevant demographic and clinical groups, not hidden inside one reassuring average.

Third, monitoring must continue after deployment. Generative systems can change through updates, new data, altered prompts, and new clinical contexts. Approval cannot be treated as a ceremonial blessing that lasts forever.

Fourth, patients should receive clear notice when AI materially shapes a decision about their health. That notice should explain the tool’s role in ordinary language and identify how to request human review.

Fifth, data collection should be limited to what is necessary. Secondary uses should require meaningful consent. Security standards should be enforceable, and breaches should trigger consequences stronger than a carefully worded apology emailed three months later.

Sixth, public agencies need enough independent technical expertise to challenge vendor claims. A regulator cannot supervise cutting-edge systems if its best employees are leaving for the companies it regulates or if its staffing is treated as decorative overhead.

Finally, the public should be able to see who is participating, what contracts exist, what performance measures are used, what failures have been reported, and how conflicts of interest are managed. Trust does not emerge because officials use the word “ecosystem” repeatedly at a podium. Trust is built through evidence, transparency, and consequences.

I Want Better Medicine, Not Better Theater

The Trump administration is right about one thing: American health care desperately needs modernization. Patients should control their records. Doctors should spend less time documenting care and more time delivering it. Useful information should move securely instead of becoming trapped inside incompatible systems. AI can help make that happen.

But urgency is not permission to become careless.

The administration’s larger posture toward AI—accelerate, compete, avoid rules that might slow industry—fits neatly with the interests of companies eager to enter one of the world’s largest markets. It fits less neatly with medicine, where the cost of a bad rollout cannot always be repaired with an update.

I refuse the lazy choice between worshiping AI and fearing it. The technology is neither a digital messiah nor a mechanical villain. It is a powerful set of tools being introduced into a health system whose incentives are already badly distorted. That is precisely why governance matters.

AI placed inside a humane, accountable system could give clinicians more time, patients more understanding, and researchers new ways to detect disease. AI placed inside an extractive, opaque system could make denial faster, surveillance deeper, and responsibility harder to locate.

The code may be new. The temptation is ancient: promise the public convenience, collect its data, reduce labor costs, privatize the gains, and socialize the mistakes.

So yes, let us use AI to improve medical care. Let it organize records, reduce paperwork, assist diagnosis, widen access, and help people understand their health. Let us test bold ideas and accept that responsible progress cannot be completely free of risk.

But let us stop pretending that asking who gets hurt, who gets paid, who owns the data, and who answers for failure is somehow hostility to innovation. Those are not obstacles placed in front of progress. They are the difference between progress and a sales pitch.

I would gladly welcome Dr. Algorithm into the examination room.

I just want to know who licensed it, who tested it, who is watching it, whether it treats everyone fairly, where it sends my records, and whether an actual human can overrule it before my insurance company teaches it the meaning of “not medically necessary.”

Apparently, that makes me difficult.

Good. Patients should be difficult when everyone else is in a hurry.

Sources and Further Reading

This article is commentary and analysis. It is not medical advice.

Labels:

Comments

Post a Comment