FDA AI/ML-enabled medical device clearances and SaMD guidance changes

ANALYZE — plain-language breakdown + contested concern

What changed for physicians building AI tools:

1. Single-output tools can now avoid device regulation ✅ — A physician building a drug recommendation system can provide one answer (e.g., “Consider lisinopril”) without listing alternatives, as long as it meets other CDS criteria. This reverses 2022 guidance requiring multiple options.

2. Clinical documentation AI gets clearer path ✅ — Radiology report summarizers and diagnostic suggestion tools qualify for non-device status if they use “well-understood and accepted sources” and the physician reviews independently (doesn’t analyze underlying images).

3. Automation bias still a regulatory concern ⚖️ — FDA maintains its position (citing a 2004 journal article) that time-critical AI may cause over-reliance. This limits regulatory clearance for tools that physicians might use in emergent scenarios without independent verification time.

4. Radiology clearance pathway is well-trodden ✅ — 79% of March clearances were radiology products. Product code QIH (anatomically agnostic imaging analysis) covered 7 submissions, indicating a mature pathway.

Confidence: verified for items 1-2,4; gated for item 3 (FDA policy position, not technical claim) | Evidence: Cooley Analysis, Covington 5 Takeaways

Contested concern — PCCP under-adoption: August 2025 guidance introduced a mechanism for pre-authorized algorithm modifications without new 510(k) submissions. Yet only 8% of March applicants used it. Possible reasons: (a) sponsors unfamiliar with new process, (b) PCCP adds upfront burden, (c) initial clearance speed prioritized over long-term flexibility. This may shift as familiarity grows, but current data suggests PCCP is not yet standard practice.

At the bedside / For clinical AI

For hospitalists and clinical AI builders:

• If you’re building CDS software (risk calculators, diagnostic suggestions, clinical documentation aids): the January 2026 guidance expands non-device pathways significantly. Single-recommendation outputs are now acceptable. Focus on criterion 4 transparency (show sources, avoid information overload, enable independent HCP review).

• If you’re using AI radiology tools: understand that 79% of new clearances are in this space. Tools are increasingly “anatomically agnostic” (one algorithm, multiple body regions). Ask your vendor if they have 510(k) clearance or rely on enforcement discretion — the line matters for liability.

• If you’re deploying time-critical AI (sepsis prediction, stroke triage): FDA maintains automation bias concerns for tools in scenarios where independent verification time is limited. This is a regulatory position, not a proven risk. Design for physician override and document clinical decision-making.

• Pediatric gap: only 3 of 24 March clearances included pediatric labeling. If you work in pediatrics, AI tools may lack validation data for your population — verify age ranges in the labeling.

Confidence: verified (regulatory pathways), gated (clinical judgment calls) | Evidence: Combined analysis from sources above

Sources

• Innolitics March 2026 AI SaMD Clearances
• Cooley LLP Analysis Jan 20, 2026
• Covington Analysis

— Jeremy Tabernero, MD · More literature watch · Get in touch