AI Outpaces Medicine's Guardrails

The week's health-AI news cuts in two directions at once: extraordinary capability gains on one side, and a governance system visibly struggling to keep pace on the other.

JAMA Says AI Will Outperform Physicians by 2030 — and Human Oversight May Make Things Worse

A peer-reviewed perspective published in the Journal of the American Medical Association makes one of the boldest claims yet seen in mainstream medical literature: by 2030, autonomous AI systems will surpass physicians in diagnosis, drug prescribing, and chronic disease management.

The authors are not fringe voices. Ezekiel Emanuel is a bioethicist who served as a health-policy adviser in the Obama White House. His co-author, Vinod Khosla, is a billionaire venture capitalist who has long championed AI's disruptive potential in medicine. Together they reviewed more than 50 peer-reviewed studies comparing AI with clinicians across specialties.

Their most provocative finding concerns oversight itself. Rather than arguing that humans should remain in the loop as a safety valve, Emanuel and Khosla warn that requiring physician sign-off on AI decisions may actively worsen outcomes. They coin the term AI-induced deskilling: as doctors increasingly defer diagnostic thinking to algorithms, they lose the clinical judgment needed to catch the cases where the algorithm is wrong — and their intermittent interventions introduce additional error rather than filtering it out.

If the argument holds, it transforms the regulatory question. The debate shifts from "how do we make sure AI assists doctors safely?" to "at what point does mandating human review become a source of harm rather than a safeguard?"

Stanford Designs AI Viruses That Kill Drug-Resistant Bacteria — and Raises a Biosecurity Red Flag

Researchers at Stanford University have synthesised 16 functional bacteriophages — viruses that infect and kill bacteria — using genome language models trained on sequences from more than two million naturally occurring phages. The AI-designed viruses successfully killed antibiotic-resistant strains of Escherichia coli in laboratory conditions.

The models used, Evo1 and Evo2, work in a manner analogous to the large language models behind text-generation tools: they learn the statistical patterns of genomic sequences and generate novel ones that follow those patterns. The resulting phages are not found in nature. They are, in a meaningful sense, AI inventions.

The therapeutic potential is significant. Antibiotic resistance is projected to kill ten million people annually by 2050, and the pipeline of new antibiotics has been near-empty for decades. Precision phage therapy — targeting specific bacterial strains without disrupting the surrounding microbiome — has long been a theoretical alternative. AI-accelerated phage design could make it practical at scale.

But biosecurity experts are sounding alarms that go beyond this specific application. If AI can design functional novel viruses capable of killing bacteria, the same approach could theoretically be directed toward pathogens. The researchers acknowledge that their work demonstrates AI-designed biology now moves faster than the governance frameworks built to contain it.

Evaxion Launches AI-Designed Glioblastoma Vaccine Targeting Viral Antigens

Evaxion Biotech has added EVX-05 to its pipeline: an off-the-shelf cancer vaccine candidate for glioblastoma, the most aggressive and lethal form of primary brain cancer, designed by its AI-Immunology platform in collaboration with Duke University.

What makes EVX-05 distinctive is both its mechanism and its accessibility. Rather than being personalised — trained on a specific patient's own tumour — it targets endogenous retroviral antigens: remnants of ancient viral infections baked into the human genome that are overexpressed in glioblastoma tumours but largely silent in healthy tissue. Because these targets are shared across patients, the vaccine can theoretically be manufactured at scale and administered without the individualised preparation that has limited the clinical and commercial reach of personalised cancer vaccines.

EVX-05 replaces an earlier programme, EVX-03, which was discontinued. The pipeline addition comes as AI-designed oncology therapies move from proof-of-concept into early clinical stages across the industry. Whether AI-discovered shared antigens will prove immunogenic enough to produce durable responses in glioblastoma — a tumour notorious for evading immune surveillance — remains to be demonstrated in trials.

Mayo Clinic Scales Its AI Diagnostics Nationally — While Facing a Lawsuit Over the Same Technology

Mayo Clinic is deploying its diagnostic AI platform beyond its own campuses, licensing it to community hospitals as a way to extend specialist-level diagnostic capability to settings that cannot recruit subspecialists. The platform is built on a foundation of more than 12,000 clinical studies.

The strategic logic is straightforward: AI can carry institutional knowledge into hospitals that would otherwise have no access to it. For patients in rural or underserved areas, the promise is meaningful. A community hospital without a cardiologist or neuroradiologist on staff might, through this kind of partnership, offer diagnostic quality approaching what patients receive at major academic medical centres.

The timing is complicated by a lawsuit filed in July 2026 alleging that a healthcare system using Mayo's platform cut corners on implementation in ways that endangered patient care and compromised patient privacy. The case surfaces a question the industry has largely left unresolved: when a health system licenses an AI diagnostic platform from a brand-name institution and something goes wrong, who bears liability — the licensor whose model made the recommendation, or the licensee whose clinicians acted on it?

Federal Health Agencies Are All-In on AI. The Legal Framework Is Not.

A newly published analysis from scholars at Harvard and the University of Maryland finds that AI use across the major US federal health agencies surged dramatically over the past year: FDA adoption rose 148%, CDC 87%, CMS 78%, and NIH 51%. The agencies are using AI across the regulatory review process, disease surveillance, claims adjudication, and research prioritisation — tasks that influence or directly determine medical decisions for hundreds of millions of Americans.

The problem, the scholars argue, is that accountability frameworks have not kept pace. When an FDA model influences whether a drug gets approved, or a CMS algorithm determines a coverage decision, existing legal doctrines struggle to assign liability if the outcome harms a patient. Whether an AI recommendation constitutes a federal action subject to administrative review is largely untested. Disclosure requirements when AI influences a clinical decision are inconsistent or absent.

The analysis frames the situation as a systemic risk. With AI now embedded in life-and-death decisions at scale, the gap between capability and accountability is no longer a theoretical concern — it is operational.