Artificial intelligence in healthcare already reads radiology scans, evaluates coverage requests, and even translates medical instructions. But there's a key difference between what the technology can do and what healthcare AI regulation — and patient safety — allow it to do without human oversight.
AI in Healthcare Today: How Much Is Already in Use
The growth is real and measurable. The FDA has authorized 1,451 AI/ML-enabled medical devices on a cumulative basis through the end of 2025, with 295 new authorizations in 2025 alone. Radiology accounts for 76% of them, but use is expanding steadily into cardiology, neurology, and other specialties.
This is no longer emerging technology — it's clinical infrastructure in daily use, in diagnostics and, increasingly, in administrative decisions like coverage.
U.S. Healthcare AI Regulation: What HHS and CMS Require
Healthcare AI regulation in the U.S. is running behind that scale, but it's moving. Since May 2024, HHS/OCR's final rule under Section 1557 of the Affordable Care Act prohibits discrimination by "patient care decision support tools" — including clinical AI — and requires organizations to audit and mitigate bias risk in those tools, with compliance required as of May 2025.
On the coverage side, CMS rule 4201-F has required, since 2024, that Medicare Advantage medical-necessity decisions be based on the individual patient's circumstances — not population data generated by an algorithm alone. Since January 2026, CMS has also been piloting AI-assisted review within Original Medicare.
The Global Picture: EU AI Act and Healthcare Compliance
Outside the U.S., the European Union has taken a stricter path. The EU AI Act classifies AI used for diagnosis, clinical decision support, triage, and patient monitoring as "high-risk," with full compliance obligations — conformity assessment, technical documentation, and human oversight. Those obligations were originally set to apply from August 2026; following the "AI Omnibus" simplification agreement reached in mid-2026, the deadline was extended to December 2027 for most high-risk healthcare AI, and to August 2028 for AI embedded in regulated medical devices.
The common thread across both frameworks, differences aside, is clear: the more automated a health decision becomes, the more organizations are required to be able to explain it, audit it, and — above all — have a person overseeing it.
Medical Translation and AI: Why Clinical Language Still Needs Human Interpreters
That same principle — human oversight where it matters — applies to medical translation and interpretation, and it's where many organizations still improvise. When a professional interpreter isn't available, it's common for care teams to fall back on Google Translate or another general-purpose machine translation tool.
The evidence on the risks of AI medical translation is consistent. A clinical study using pediatric emergency scenarios found Google Translate's accuracy ranged from 83.5% (Urdu) to 95.4% (French), with nuance errors even in the best-performing languages, and concluded that its use carries legal risks around privacy, consent, and professional liability. An analysis of mental health content translations found 17% of Persian-language outputs were incomprehensible, and documented errors that directly reversed clinical meaning — including one case where "exclude" was translated as its opposite, "include." A widely cited emergency medicine study comparing professional interpreters, ad hoc interpreters (family members or untrained staff), and no interpreter at all found that errors made without a professional interpreter were significantly more likely to carry real clinical consequences.
A 2026 scoping review of AI-based communication in primary care, covering studies from 2017 to 2024, found no evidence that Google Translate has ever been tested for patient safety in general practice, and warned of the risk that it displaces trained medical interpreters without ever going through that check.
HIPAA and FDA Compliance: What Certified Medical Translation Requires
This isn't an anti-technology stance — it's what healthcare compliance already requires. For informed consent in clinical trials, the FDA requires that participants with limited English proficiency receive the same information as English-speaking participants, backed by documentation of translator qualifications and a signed certification of accuracy. For higher-risk or more complex studies, Institutional Review Boards (IRBs) often require back-translation — an independent translation back into the source language by a second translator — as a quality control step. And a signed Certificate of Translation from the translator or agency is, in practice, an audit requirement for the FDA and EMA on critical clinical documents.
None of these three controls can be met with a machine-translation output that has no certified process behind it.
That's ultimately the conversation running through AI in healthcare: it's not about choosing between technology and people, but about knowing at what point in the process a decision — clinical, administrative, or linguistic — needs to stay in human hands, documented and auditable. At Arcadia, we work that exact point every day: certified medical and regulatory translation services, delivered by linguists specialized in healthcare, within ISO-aligned, HIPAA-compliant workflows that ensure traceability and controlled handling of sensitive information — exactly what regulation, on both sides of the Atlantic, is starting to require.
References
U.S. Food and Drug Administration www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices
HHS U.S. https://www.hhs.gov/civil-rights/for-individuals/section-1557/faqs/index.html
CMShttps://www.cms.gov/newsroom/fact-sheets/2024-medicare-advantage-and-part-d-final-rule-cms-4201-f
European Commission https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
PMC https://pmc.ncbi.nlm.nih.gov/articles/PMC11762056/
"Errors of medical interpretation and their potential clinical consequences: a comparison of professional versus ad hoc versus no interpreters."