AI in Healthcare: What’s Working, What’s Hype, and What’s Next

The most widely adopted AI in medicine doesn’t diagnose anyone. It listens to a visit and drafts the note. Meanwhile, radiology owns roughly three-quarters of all FDA-authorized AI devices, and the first generative-AI drug has entered Phase III. Large language models still repeat fabricated clinical details in a majority of adversarial tests. Here is where each category stands as of October 2026.

Quick Takeaways

• Working: Ambient AI scribes cut documentation time and burnout. Imaging AI has the deepest regulatory track record.
• Hype-prone: LLM clinical decision support. Hallucination rates in adversarial tests ranged from 50% to 82%.
• Unproven: AI-discovered drugs. Early-phase success looks strong, but no AI-discovered drug has completed Phase III.
• Next: Regulators are writing rules for generative AI devices, and ambient tools are moving into coding and billing.

The 2026 Scorecard

What’s Working: Ambient AI Scribes

Ambient scribes record a visit, transcribe it with speech recognition, and use an LLM to draft a structured note for clinician review. Adoption has been fast. A Medical Economics report (March 2026) found about 70% of physicians at UCSF using scribes in daily practice. An AJMC study found 62.6% of Epic-using hospitals had adopted an ambient AI tool as of June 2025.

The evidence is encouraging but modest:

• A multicenter JAMA Network Open study of 263 clinicians across six health systems found burnout prevalence fell from 51.9% to 38.8% after 30 days.
• A 2026 JAMA study followed 8,581 ambulatory clinicians over two years. Secondary summaries report savings of roughly 16 minutes per 8-hour shift, concentrated among heavy users.
• Epic launched native AI Charting in February 2026, which pressures standalone vendors such as Abridge.

Two caveats matter. First, the Peterson Health Technology Institute warns that scribes extending into coding may raise coding intensity, and with it spending, even if the coding is accurate. Second, consent is a live legal issue. A January 2026 class action against Sharp HealthCare alleges a patient was recorded without knowledge. A UK GP survey also flagged accuracy and equity concerns in multilingual consultations.

Verdict: Real value, modest per-clinician time savings, and unresolved billing and consent questions.

What’s Working: Imaging AI and Device Regulation

The FDA’s AI-Enabled Medical Device List passed 1,600 authorized devices in 2026. Counts differ by snapshot date and source, but the pattern is consistent.

The FDA states that its list is not comprehensive, and that “FDA-authorized” is more accurate than “FDA-approved.” Most clearances came through 510(k), which relies on equivalence to an existing device. Analyses cited in 2026 reporting found that fewer than 1% of authorizations reported patient health outcomes.

On August 18, 2026, the FDA published a discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices. Expect LLM-based tools to face a different review path than classic image classifiers.

What’s Hype-Prone: LLMs in Clinical Decision Support

Passing a licensing exam is not the same as reasoning safely over a messy chart. A Mount Sinai study (published in Communications Medicine) planted one fabricated detail, such as a fake lab test or invented syndrome, into 300 physician-validated vignettes. Researchers tested six LLMs and generated 5,400 outputs.

Three lessons follow. Lowering temperature does not fix this failure mode. Prompt design helps but does not eliminate the problem. Shorter inputs showed slightly higher hallucination odds.

Engineering a Safer Clinical Summarizer

Treat input as untrusted and force the model to ground every claim. This pattern uses the Anthropic Python SDK. Use only de-identified or synthetic text unless your deployment has a signed BAA and approved PHI handling.

pip install anthropic
export ANTHROPIC_API_KEY="your-key"

import anthropic

client = anthropic.Anthropic()

SYSTEM = """You summarize clinical notes for clinician review.
Rules:
1. Use ONLY facts present in the note. Quote the source phrase for each claim.
2. If a lab, sign, or diagnosis cannot be verified from standard medical
   knowledge, flag it as [UNVERIFIED] instead of elaborating.
3. If information is missing, write [NOT IN NOTE]. Never infer values.
4. Do not recommend treatment."""

note = "SYNTHETIC EXAMPLE: 58M, chest pain x2 days. Troponin ordered..."

response = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=800,
    temperature=0,          # reduces variance, does NOT remove hallucination
    system=SYSTEM,
    messages=[{"role": "user", "content": f"<note>\n{note}\n</note>\n\nSummarize."}],
)
print(response.content[0].text)

Annotations:

• temperature=0 improves reproducibility. The Mount Sinai results show it is not a safety control.
• The [UNVERIFIED] flag targets the exact failure tested: elaborating on fabricated details.
• Keep a clinician in the loop. Log every output for audit.

What’s Unproven: AI Drug Discovery

Capital has moved faster than clinical proof. Industry analyses estimate roughly $60 billion invested since 2019 and about 175 AI-linked clinical programs.

The flagship case is rentosertib (ISM001-055), an oral TNIK inhibitor for idiopathic pulmonary fibrosis from Insilico Medicine. The company says it is the first end-to-end generative-AI drug candidate to reach Phase III, a claim that depends on how you define “end-to-end.”

Other late-stage entrants include Generate:Biomedicines’ GB-0895, an AI-engineered antibody in Phase 3 for severe asthma since December 2025.

The honest scoreboard: a widely cited 2024 analysis found 80–90% Phase I success for AI-discovered molecules versus 40–65% historically. But Phase II failure rates look similar to conventional drugs, and none had completed Phase III at the time of reporting. AI speeds the early stages. Biology still decides the outcome.

Real-World Workflows

How a health system should deploy an ambient scribe

1. Pilot with volunteers in one specialty. Track note-edit time and after-hours “pajama time.”
2. Set consent policy before launch: patient notice per encounter, recording retention limits.
3. Audit samples weekly: compare drafts with signed notes for omissions and invented findings.
4. Monitor coding drift: track level-of-service distributions before and after rollout.
5. Review equity: test non-English and interpreter-mediated visits separately.

Prompt template for clinician-facing literature checks

Role: Clinical evidence assistant. Audience: licensed physician.
Task: Summarize evidence on [intervention] for [condition].
Constraints:
- Cite only studies you can name with year and journal; otherwise say "no source".
- Separate RCT evidence from observational and benchmark evidence.
- State what outcome was measured (clinical outcome vs. surrogate).
- End with a list of claims that need manual verification.

What’s Next

• Generative AI device rules. The FDA’s August 2026 discussion paper signals new risk frameworks for LLM-based devices.
• Scribes become billing engines. Coding support raises both revenue and audit exposure.
• Outcome evidence. The field needs trials measuring patient outcomes, not just accuracy or time saved.
• Phase III readouts. Rentosertib’s results will test the AI drug discovery thesis, though 52-week data is not imminent.

FAQ

Is AI replacing doctors?

No. The highest-adoption tools handle documentation and image triage under clinician oversight. No evidence reviewed here shows autonomous replacement of physician judgment.

What is the most successful use of AI in healthcare today?

Ambient AI scribes have the fastest adoption and the clearest early benefits, including lower burnout and less documentation time. Imaging AI has the longest regulatory record, with roughly 76% of FDA-authorized AI devices in radiology.

Are AI chatbots safe for medical advice?

Not reliably. In a Mount Sinai test, models repeated planted false clinical details in up to 83% of cases. Prompt safeguards cut the rate roughly in half but did not eliminate it. Use chatbots for general education and verify anything clinical with a professional.

Has AI discovered an FDA-approved drug?

Not yet. Rentosertib is the furthest along, with its Phase III trial dosing its first patient on September 9, 2026. It remains investigational.