AI Clinical Decision Support Gains Ground as FDA Loosens Oversight

Strong diagnostic results arrive as regulators step back

AI-driven clinical decision support (CDS) had a busy 2026, with progress on two fronts at once. Researchers published diagnostic results that would have seemed unlikely a few years ago. Over the same months, the U.S. Food and Drug Administration redrew its rules so that fewer of these tools need its approval before they reach hospitals. The two trends meet inside hospitals, where adoption is moving faster than the evidence or the oversight. Taken together, the coverage points to one conclusion: hospitals buying clinical AI now carry much of the responsibility for checking whether it works.

What happened at the FDA

On January 6, 2026, the FDA issued revised final guidance on Clinical Decision Support Software, replacing the 2022 version. Commissioner Marty Makary presented it as a way to cut unnecessary regulation and encourage innovation.1 The agency put out a revised version on January 29, which is why some coverage gives the later date.929

The main change is about tools that give a single answer. Under the 2022 approach, software that offered one specific recommendation instead of a list of options was more likely to count as a regulated medical device. Developers responded by padding their outputs with extra options, even when only one made clinical sense.2 The FDA now says it will use enforcement discretion for tools that give one clinically appropriate recommendation, as long as they meet the other non-device criteria.15 The guidance also says software that drafts a radiology report summary from a radiologist's own findings may fall outside device rules if a clinician stays in the loop.1

The underlying legal test did not change. Under section 520(o)(1)(E), software escapes device status only if it meets four conditions:

• It does not analyze medical images or physiological signals.
• It presents recognized medical information.
• It is aimed at a healthcare professional, not a patient.
• It lets that professional independently review the basis for each recommendation.4

Analysts mostly agree the fourth condition decides most cases. This is where large language models (LLMs) run into trouble. At a March 2026 town hall, the FDA reportedly confirmed that LLM-based software can qualify as non-device CDS, but only if clinicians can actually check how it reached its recommendations.4

Some lines did not move. Any function that analyzes medical images to produce a diagnostic recommendation is still regulated, and so are direct-to-consumer AI image interpretation services.3 The College of American Pathologists expects most predictive AI in pathology to stay under FDA oversight because of how images are classified.10

Where the readings diverge

The legal commentary splits on emphasis more than on facts. Covington calls the revision a set of useful clarifications, not a major rewrite.16 Others describe a change in philosophy. One physician commentary says the agency is accepting more low-risk innovation and leaning harder on professional judgment and post-market accountability.2 Orrick calls it a strategic rebalancing that keeps scrutiny on black-box, time-critical or directive tools.5

The sharpest disagreement is over what the guidance says about AI. Orrick reads the expanded discretion as covering certain generative AI features, provided clinicians can understand and verify the logic.5 Covington notes that although the FDA announced the guidance as focused on AI, the document itself says nothing directly about AI, leaving questions about clinician-facing chatbots and AI-drafted prescription renewals unanswered.1 A compliance guide written in Februrary takes a stricter view. It lists unexplainable AI/ML models and sepsis prediction tools whose basis cannot be reviewed as examples of regulated devices.7

Both readings are partly right. The guidance makes the rules more predictable for transparent, guideline-based tools. It does not settle how probabilistic LLM output can meet a standard built around independent review. Coverage also agrees that the guidance does not address patient-facing symptom checkers and chatbots.12

The most useful framing comes from a procurement-focused analysis. Many hospitals treated an FDA marker as a free signal of quality, and the share of decision-support tools carrying one is likely to shrink just as adoption speeds up.6 The same analysis notes that the guidance was issued without a public comment period. It also cites concerns that clinicians often lack the time or tools to do the independent review the exemption depends on, which could leave them relying on software no regulator has examined.6

The diagnostic evidence: impressive, with conditions

The best-known result of the year came from Harvard Medical School and Beth Israel Deaconess Medical Center, published in Science. In 76 real emergency room cases, OpenAI's o1 reasoning model reached the exact or a very close diagnosis at initial triage 67% of the time. The two attending physicians it was compared against scored 55% and 50%.15 The researchers stressed that the model used the same electronic record information the physicians had at the time, without any pre-processing.1512 Mount Sinai's chief clinical officer, who was not involved, said the technology was possibly ready for prime time.12

The caveats were significant. The authors called for prospective trials in real patient care rather than claiming the model was ready for life-or-death decisions.15 The AI worked from text only, while real clinicians also rely on images, sounds and nonverbal cues.12 Emergency physician Kristen Panthagani pointed out that the comparison group was internal medicine attendings, not emergency physicians, and said the headlines were overhyped.15 Some reported differences also lacked statistical significance where confidence intervals overlapped.19

In June, Nature published two more systems, MIRA and Google's AMIE. MIRA reached 87.8% diagnostic accuracy against 78.1% for a six-physician panel. AMIE matched primary care physicians on management reasoning in simulated multi-visit cases and did better on guideline alignment.14 Both teams said real-world validation is still needed.14

Other findings point the other way:

• Safety benchmark: Stanford and Harvard's NOHARM benchmark found that even advanced models produced severely harmful recommendations in up to 22.2% of 100 real consultation cases.13
• Broad meta-analysis: Across 83 studies, generative AI averaged 52.1% diagnostic accuracy and trailed specialist physicians by 15.8 percentage points.11
• Doctors using a chatbot: In a University of Virginia trial, physicians using ChatGPT Plus did no better than colleagues using conventional resources, even though the chatbot alone scored above 92% on the same cases.11
• Automation bias: A JAMA study found that physicians shown deliberately biased AI predictions lost 11.3% in accuracy, even when the AI explained its reasoning.11

The automation bias finding matters for regulation. The FDA's revised framework relies on clinicians being able to review a tool's reasoning, yet this evidence suggests a visible explanation does not reliably stop doctors from following the AI.

The gap between benchmark and bedside is a recurring theme. One industry analysis cites hospital studies in which diagnostic AI lost 20 to 35 percentage points of accuracy once it met live EHR data. That claim is reported second-hand and should be treated with caution.20 The pattern makes sense, though. Narrow imaging tools trained on one task report sensitivity in the high 80s to high 90s. General-purpose reasoning models are much more variable.13

Hospitals are deploying anyway

Adoption is not waiting for these debates to end. The American Medical Association reported in March 2026 that 81% of U.S. physicians use AI in their practice, up from 38% in 2023.13 Aidoc says its imaging AI runs in nearly 2,000 hospitals. One analysis points out that this is a company-reported deployment figure, not proof of better patient outcomes.28 Intermountain Health uses AI for decision support, and Cleveland Clinic has partnered with PathAI on diagnostics.21

Another complication is that the tools people actually use are not always the ones their institutions approved. Wolters Kluwer experts warned that health system executives are catching up with clinicians who have already adopted generative AI apps on their own, often without being able to spot answers that sound authoritative but are clinically wrong. They also raised the risk of clinicians losing skills over time.25

Paying for these tools is its own barrier. Most CDS tools have no dedicated reimbursement and are funded by health systems themselves.29

Some claims deserve skepticism. Vendor marketing citing 20–30% mortality reductions across hospital AI CDS comes from companies selling the software, not from independent trials.26 Even the count of FDA-authorized AI tools varies by source, from "700+" to 1,524.2613 That spread shows how unreliable the numbers in this market can be.

Why it matters

Outside the U.S., regulation is getting stricter. Obligations for high-risk medical AI under Article 6(1) of the EU AI Act apply from August 2, 2027.28 In the U.S., the FDA has moved toward post-market monitoring and fewer pre-market gates.6

The result is a split that tracks the technology. Narrow image-analysis tools, which have the strongest evidence, stay under FDA review. Text-based reasoning tools, whose results range from beating physicians to giving severely harmful advice in more than one case in five, are the ones getting more room to operate.

The central point is that the FDA did not deregulate diagnostic AI as a whole, and the strong study results do not mean autonomous diagnosis is here. What did change is who has to do the checking. Hospitals now need their own validation, monitoring and governance, because a federal clearance will cover fewer of the tools they buy.