Claude Sonnet-5.5

Kaiser Permanente deploys ambient AI scribes across its hospitals and medical offices to draft visit notes for physicians. Signals a shift of note-accuracy liability onto clinicians who sign AI-drafted records.

Writing · Concrete actor (hospital systems) and event (deploy AI). Lacks specific company/product names, dates, or numbers.

Radiology accounts for roughly three-quarters of FDA-authorized AI-enabled medical devices, concentrating validation demands in imaging workflows. Signals that hospital imaging departments carry local performance monitoring duties for algorithm drift across scanners and patient populations.

Judge · The FDA confirms over 1,600 AI-enabled devices, with radiology accounting for a significant majority (76.5% - 76%) of authorizations. This trend is well-documented and current.

Writing · Concrete actor (FDA, radiology), event (authorized products), and specific observation.

External validation of the Epic Sepsis Model in JAMA Internal Medicine found it missed two-thirds of sepsis cases. Indicates that vendor-supplied predictive models require local validation and ongoing monitoring before and during clinical use.

Judge · The JAMA Internal Medicine study [jamanetwork.com](https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781307) confirmed the Epic Sepsis Model missed 67% of cases, highlighting the need for local validation. A 2023 study [doi.org](https://doi.org/10.1093/jamiaopen/ooae133) in EDs found suboptimal performance, reinforcing this trend.

Writing · Excellent specificity: actor (Epic), event (validation), publication (JAMA), and quantitative anchor (two-thirds).

A 2024 JAMA Network Open randomized trial found GPT-4 access did not significantly improve physician diagnostic reasoning versus conventional resources. Indicates that benchmark performance of language models does not translate automatically into measurable clinician outcome gains.

Judge · The JAMA Network Open study (2024) directly supports the claim that LLM access did not significantly improve physician diagnostic reasoning. The LLM alone performed better, but physician-LLM collaboration did not.

Writing · Concrete actor, event, and quantitative/temporal anchor. Active voice. Excellent specificity.

The EU AI Act entered into force in August 2024, and high-risk obligations for AI in regulated medical devices apply from August 2027. Signals that hospitals deploying clinical AI face dual compliance under the AI Act and the Medical Device Regulation.

Judge · The EU AI Act classifies AI medical devices as high-risk, with obligations for these applying from August 2027. Deployers (hospitals) face documentation, human oversight, and logging duties.

Writing · Concrete actor, event, and anchor, but lacks a specific product/filing. Contains some generic forecast.

FDA finalized guidance in December 2024 allowing manufacturers to pre-authorize planned AI model updates through predetermined change control plans. Signals that cleared AI models update after authorization, increasing hospital responsibility for version tracking and revalidation.

Judge · The FDA's PCCP guidance details the necessity for manufacturers to specify and validate modifications, implying hospitals must track and potentially validate updates.

Writing · Concrete actor (FDA), event (finalizing framework), and measurable shift (new regulatory pathway) are present. Lacks a temporal anchor.

The HHS final rule under Section 1557 requires covered entities to identify and mitigate discrimination risk in patient care decision support tools, effective May 2025. Indicates that hospitals hold direct accountability for bias in vendor-supplied algorithms.

Judge · HHS Section 1557 final rule applies non-discrimination to AI/emerging tech in patient care decision support tools, requiring identification and mitigation of discrimination risks. Compliance starts July 2024 and May 2025.

Writing · Concrete actor (HHS), specific rule (1557), measurable shift (May 2025). Deduct for "indicates" (passive).

The European Health Data Space Regulation entered into force in March 2025, setting rules for primary and secondary use of health data, including AI training. Indicates that EU hospitals face new data-access and interoperability obligations affecting AI development and vendor contracts.

Judge · The EHDS regulation (EU) 2025/327 establishes frameworks for EHR and secondary data use for AI development, distinguishing clinical access and research via data permits.

Writing · Concrete actor, event, and temporal anchor. Some specific mentions of 'phased rules' but still light on quantifiable data.

Lawsuits against UnitedHealth and Humana allege the nH Predict algorithm drove Medicare Advantage post-acute care coverage denials. Indicates that payer-side AI shapes hospital revenue cycle and length-of-stay management, requiring dedicated appeal capacity.

Judge · Multiple sources, including court documents and news articles, confirm lawsuits against UnitedHealth and Humana regarding nH Predict. The high reversal rate of denials on appeal suggests the need for dedicated appeal capacity.

Writing · Names specific actors (UnitedHealth, Humana), an algorithm (nH Predict), and concrete event (lawsuits/denials).

Epic integrates generative AI features for message drafting and charting directly into its EHR, which serves the largest share of US acute-care hospital beds. Signals that AI procurement follows EHR vendor roadmaps, narrowing negotiating leverage and independent model selection.

Judge · Epic has officially rolled out generative AI features, including in-basket drafting, chart summarization, and a native ambient tool called Art for Clinicians. This shifts procurement leverage to EHR vendors.

Writing · Names Epic and its EHR, specific features, and anchors to US acute-care hospital beds. Uses active voice.

The Joint Commission and the Coalition for Health AI released joint guidance in September 2025 on responsible AI use in healthcare delivery organizations. Signals that accredited hospitals face expectations for AI governance structures, performance monitoring, and incident reporting.

Judge · Joint Commission and CHAI's guidance, dated 2026, directs hospitals to establish AI governance structures, including committees and incident reporting processes. This indicates a clear push towards AI governance as an organizational competency.

Writing · Concrete actors, event, and temporal anchor. Specific requirements outlined.

The 2024 Change Healthcare ransomware attack disrupted claims processing across US providers, exposing concentrated dependencies on third-party technology vendors. Indicates that AI vendors with access to clinical data extend the third-party risk surface that hospital continuity plans must cover.

Judge · The Change Healthcare attack highlighted existing third-party dependency risks. AI vendors extend this, but specific widespread AI vendor disruptions aren't yet documented.

Writing · Concrete actor (Change Healthcare), event (ransomware attack), and temporal anchor (2024) are excellent. The AI vendor link is a forward-looking implication, but still specific enough.

California AB 3030 requires health facilities to disclose generative AI use in patient clinical communications, effective January 2025. Indicates that state-level disclosure duties reach hospital patient portals and messaging workflows, including human contact options.

Judge · California law [leginfo.legislature.ca.gov] mandates disclosure for AI-drafted clinical communications starting Jan 1, 2025. EU guidance [rm.coe.int] emphasizes transparency, but without a specific law.

Writing · Concrete actor, event, and temporal anchor. Excellent specificity and active voice.

A 2023 Pew Research Center survey found 60 percent of US adults feel uncomfortable with providers relying on AI for diagnosis and treatment. Indicates that majority patient discomfort constrains consent practices and communication strategies for clinical AI.

Judge · Pew Research (2023) confirms 60% discomfort. Other sources (2024, 2025) corroborate widespread desire for notification and racial disparities in trust/bias concerns, directly supporting the signal.

Writing · Concrete actor, event, and temporal anchor. 'Most' is slightly vague but survey-based. Active voice.

Obermeyer and colleagues in Science found a widely used US risk-prediction algorithm underestimated Black patients' needs by using healthcare cost as a proxy. Indicates that documented bias cases sustain patient scrutiny of hospital algorithms and support regulatory equity requirements.

Judge · Obermeyer et al. (2019) published in Science demonstrated racial bias in a widely used US algorithm. This supports patient scrutiny and regulatory needs for equity.

Writing · Concrete actors, event, and measurement; active voice. Minor deduction for 'widely used' and 'patient scrutiny'.

The UK Information Commissioner ruled in 2017 that Royal Free NHS Trust unlawfully shared 1.6 million patient records with DeepMind. Indicates that consent and transparency over secondary data use remain trust and legal exposure for hospital AI partnerships.

Judge · The UK ICO ruled in 2017 that Royal Free NHS Trust unlawfully shared 1.6M patient records with DeepMind without sufficient consent/transparency, confirming legal exposure. [ico.org.uk](https://ico.org.uk/for-the-public/ico-40/google-deepmind-and-class-action-lawsuit/) [bbc.co.uk](https://www.bbc.co.uk/news/technology-40483202) [medconfidential.org](https://medconfidential.org/wp-content/uploads/2018/05/royal-free-undertaking-03072017.pdf)

Writing · Concrete actor, event, and quantitative anchor. 'Remain trust' is slightly awkward.