The most peculiar aspect of visiting a contemporary American hospital in 2026 is how imperceptible artificial intelligence has become. Before a nurse even looks up, it hums beneath everything, silently reading scans, identifying sepsis risk, writing patient messages, and organizing triage notes. It’s not announced by the screens. Seldom do the patients know. However, dozens of models are making recommendations that influence the course of events somewhere between the radiology suite and the discharge desk.
Because medical AI’s capability side has advanced rapidly, improving diagnostic precision, reducing administrative burdens, and revealing patterns that would be invisible to the human eye, while the safety side has hardly advanced at all. This winter, a CUNY scoping review examined 390 clinical AI models published between 2020 and late 2021 and discovered an unsettling finding: only 9% of the models explained how they would be updated over time. Merely 12% adhered to established reporting guidelines. Furthermore, a startling 84% of respondents never disclosed the racial or ethnic composition of their training data. These figures are not marginal. The foundations are these.
| Field | Details |
|---|---|
| Topic | The capability–safety gap in clinical AI across U.S. hospitals |
| Core Concern | Rapid deployment of clinical AI tools without matching safety, validation, and oversight standards |
| Key Study Reviewed | Scoping review of 390 clinical AI/ML models published March 2020 – December 2021 |
| Models Still in Research Phase | 98% |
| Models Actually Deployed in Real-World Care | 2% |
| Models Following Established Reporting Standards | 12% |
| Models Failing to Report Racial or ethnic data breakdown | 84% |
| Models With a Plan for Updating Over Time | 9% |
| Frameworks Cited | PAST (Poison, Abuse, Steal, Trick); CHARMS checklist (simplified six-item version) |
| Notable Voice | Azizi A. Seixas, PhD, University of Miami Miller School of Medicine |
| Regulatory Context | World Health Organization guidance on accountability, transparency, and public interest |
| Patient Sentiment | Around 60% of Americans report discomfort with providers relying on AI in personal care (Pew, 2023) |
| Article Length | 500–600 words |
| Date | May 2026 |
Here, it’s difficult to ignore the cultural mismatch. American medicine has always been wary of strange situations, meticulous about medication labels, picky about consent forms, and strangely at ease with the silent software that directs clinical care. A medication is removed if it performs differently in subgroups. A press release is frequently issued for a model that performs differently in subgroups.
Azizi Seixas, the director of the University of Miami’s Media and Innovation Lab, frames the problem with an acronym he calls PAST, which stands for poison, abuse, steal, and trick. However, his more intriguing observation is the one he makes almost casually. According to him, a model may appear superb in development but turn dangerous in real life. Populations change. Workflows change. The tool is used by a clinic outside of the context in which it was validated. Drift occurs. Then one morning, a sepsis alert goes off too frequently, doctors begin to ignore it, and a genuine warning is overshadowed by a hundred false ones. It’s not theoretical. In locations that no one is naming on the record, that is already taking place.
Speaking with individuals within health systems gives the impression that the sector is aware of the gap and is hoping someone else fills it. Hospitals are indicated by vendors. Regulators are cited by hospitals. Regulators cite the WHO, which continues to release thoughtful guidelines that no one is obliged to abide by. The patient at the end of the chain doesn’t know how to find out if the algorithm reading her mammogram has ever been audited for someone who looks like her while the models continue to ship.
The deeper discomfort is related to trust, which is nearly impossible to regain in the medical field once it has been lost. According to a 2023 Pew survey, 60% of Americans would feel uneasy if they knew their provider used AI. That figure predates the current generation of agentic systems—those that act rather than merely recommend. The figure may have become softer. Perhaps it hasn’t, and people have just stopped inquiring.
The capability question isn’t what feels urgent right now. Strangely, the easiest question is the one about capability. The question of whether American medicine can construct the cumbersome, unglamorous infrastructure—registries, update protocols, subgroup reporting, and explainability requirements—that transforms a clever model into a secure one remains unanswered. As of right now, the truth is no. Not just yet. And while everyone waits for someone else to close the gap, it continues to grow.
