Editorial policy
How this work is researched and reviewed
AIMOCS writes for clinicians and health leaders about a field where being wrong has consequences. These are the standards every briefing, guide, and data page is held to before it reaches you.
What we cover, and for whom
We write about artificial intelligence in healthcare — the evidence behind clinical tools, how they are evaluated, and how they behave once deployed. Our readers are practicing clinicians, health-system leaders, and the people building this technology. We write in plain language, and we assume a reader who wants the reasoning behind a finding as much as the finding itself.
Sourcing standards
Every factual claim in a published piece is tied to a primary source: a peer-reviewed study, a regulatory filing, an official dataset, or a document from the organization that produced the information. We prefer the original over any summary of it. Each article carries a numbered source list, and each figure in the text points to the source it came from.
- Statistics carry the date they reflect and the study or filing they are drawn from. Where a number changes over time, the page is a living one and shows when it was last updated.
- Clinical and regulatory claims are held to the primary literature. Marketing material and secondary reporting may point us toward a source, but they are never the source themselves.
- An automated check resolves every citation against Crossref and the live web, and checks each referenced work for a retraction notice. It runs on every proposed change to the library, and a source that cannot be resolved, or that has been retracted, fails the check.
Who is responsible for this work
Ibrahem Albalkhi holds editorial responsibility for everything published on this site. He is an AI and healthcare researcher with 36 PubMed-indexed publications as of 13 August 2026; the citations are linked on the about page. He decides what publishes, what is held back, and what is corrected. Disputes about anything on this site reach him at the corrections address below.
Human review
A draft becomes a published page only after a person reads it against its sources. Every piece is checked for accuracy, clarity, and tone before it publishes, and the editorial owner named above signs off on what goes live.
Where a named subject-matter reviewer has independently reviewed an article, that person is credited in the article’s byline block. Where no such review has happened, we say nothing — we do not credit a review that was not performed. Most of the library carries editorial review only today; independent clinical and regulatory review is being added article by article as reviewers join, and the credit appears when it has been earned.
Independence
Our editorial choices are made on the merits. When we write about a tool, a vendor, or a study, coverage is not exchanged for payment, and a commercial relationship — where one exists — is disclosed on the page. Guest contributors and faculty are identified as such, and their affiliations are stated.
Corrections
When we get something wrong, we fix it in the open. A substantive correction is noted on the page along with the date it was made, and logged below, so the record stays honest. If you believe a page contains an error, write to corrections@aimocs.com with the page and the specific claim, and we will review it against the source.
A correction changes a fact, a figure, or a conclusion. Fixing a typo, a broken link, or a formatting error is not a correction and is not logged. Routine scheduled updates to living pages are shown by the update date on the page itself.
Corrections log
- — FDA-cleared AI devices by year and specialty: Led with 692 devices by 2023 and 2024 as the record year, and said no FDA-authorized device used a large language model; it now leads with the FDA list's 1,430 AI-enabled devices through 2025 (331 in 2025) and notes the December 2025 clearance of a device with a patient-facing LLM.
- — The FDA AI-enabled device list: a statistics tracker: Said large language models were absent from the list while also reporting a December 2025 LLM clearance, and led with the FDA's January 2025 figure of more than 1,000 devices; it now scopes the absence to a 2025 taxonomy and leads with 1,430 devices through 2025.
- — AI in healthcare statistics (2026): Gave Med-PaLM 2's MedQA score as 86.1% and said no FDA-authorized device relied on a large language model; it now gives the published 86.5% and notes the December 2025 clearance of a device with a patient-facing LLM.
- — AI in radiology: what is cleared and evidenced in 2026: Said none of the FDA-authorized devices used large language models; it now says none of the 1,016 devices in a 2025 taxonomy did, and notes the December 2025 clearance of a device with a patient-facing LLM.
- — AI in cardiology: what is cleared and evidenced in 2026: Said none of the FDA-authorized devices used large language models; it now says none of the 1,016 devices in a 2025 taxonomy did, and notes the December 2025 clearance of a device with a patient-facing LLM.
- — AI in pathology: what is cleared and evidenced in 2026: Said none of the FDA-authorized AI devices used large language models; it now says none of the 1,016 devices in a 2025 taxonomy did, and notes the December 2025 clearance of a device with a patient-facing LLM.
- — Imaging AI marketplaces compared: Said none of the FDA-cleared devices used large language models; it now says none of the 1,016 devices in a 2025 taxonomy did, and notes the December 2025 clearance of a non-imaging device with a patient-facing LLM.
- — AI in Emergency Care: Summarized a 2025 FDA device taxonomy as finding that no devices use generative models; it found that none of its 1,016 devices used a large language model, while more than 100 generate data such as images.
- — AI in oncology: what the evidence actually shows: Summarized a 2025 FDA device taxonomy as finding that no devices use generative models; it found that none of its 1,016 devices used a large language model, while more than 100 generate data such as images.
- — AI in primary care: exam scores versus patient outcomes: Cited the 86.5% MedQA score to the paper that introduced Med-PaLM, which reported 67.6%; the score is now credited to Med-PaLM 2 and its 2025 Nature Medicine paper.
- — What is agentic AI in healthcare: Called MedAgentBench's 69.67% best-model score the ceiling for clinical agents; it is now dated to the 12 models tested in the benchmark's 2025 publication.
- — The best AI in healthcare courses in 2026, compared: Described AIMOCS as offering sixteen module artifacts, a public curriculum currency log, a case-study library and author conversations; it now lists what members get: ten self-paced courses (129 lessons, each ending in a hands-on exercise) with certificates of completion, and a live cohort by application.
- — The best AI in healthcare communities, societies, and networks in 2026: Described AIMOCS membership as a weekly research briefing, deployment case studies, live sessions and a quarterly-reviewed curriculum; it now lists the discussion feed, member directory, events, ten self-paced courses with certificates, the full briefing and the open library.
- — How to get into AI in healthcare: a clinician's guide: Described AIMOCS membership as a weekly research briefing, deployment case studies, live sessions with senior leaders and a quarterly-reviewed curriculum; it now lists what membership includes today.
- — Designing human oversight for clinical agents: Said every AI-drafted note in the cited ambient-scribe study required review before signing, which the study does not report; the passage now describes the five-site study of 1,809 adopters as published.
- — Agents in nursing workflows: documentation, handover, and the virtual-nursing evidence: Presented 631 to 875 flowsheet entries per 12-hour shift as a 2025 study's own finding; it is an earlier figure that study cites, and the study's own measure (31% of the shift in flowsheets) is now reported separately.
- — Revenue-cycle and coding agents: accuracy, denials, and the compliance stakes: Gave KFF's insurer denial-rate range as 1% to 54% and CAQH's electronic adoption rates as 81%, 78% and 40%; the published figures are 3% to 36%, and 80% (claim status), 77% (claim payment) and 35% (prior authorization).
- — Prior-authorization agents: what the CMS rule and the evidence actually say: Gave CAQH's electronic prior-authorization adoption rate as about 40%; the published fully electronic rate is 35%.
- — Clinical AI trial results tracker: Combined the best algorithm from a 250-person diabetic-retinopathy validation with the referable-disease score from a later 343-patient deployment into one result; the two are now reported separately.
- — Global AI in health regulation tracker: Said the EU's Digital Omnibus was still only proposed, and quoted a TGA phrase from the wrong TGA page; the Omnibus entered into force on 27 July 2026, moving high-risk dates to 2 December 2027 (Annex III) and 2 August 2028 (Annex I), and the TGA point is now sourced to the right release.
- — AI in pediatrics: a 2026 guide: Said about 18.7% of pediatric-labeled devices described validation on data that included children, a figure the cited audit does not contain; the page now gives the audit's finding that 6 of 42 pediatric-labeled devices (14.3%) carried a registered clinical-trial identifier.
- — AI in nursing: a 2026 evidence guide: Reported the CONCERN trial's mortality benefit without its 24.9% rise in unanticipated ICU transfers; both results are now reported.
- — US state laws on AI scribes: a tracker: Said Colorado's AI Act (SB 24-205) would take effect on 30 June 2026; it was repealed and reenacted as SB 26-189, signed 14 May 2026, a narrower framework effective 1 January 2027.
- — Patient triage chatbots compared: Described a vignette study's 82.1%, 70.5% and 23.5% figures as the share ranking the correct condition first; they are top-3 diagnostic accuracy.
- — AI scribe head-to-head comparison: Reported a pooled Mini-Z burnout figure that the trial did not calculate, and a '50+ specialties' figure for Nabla with no source; the page now gives DAX (+2.83) and Nabla (+2.69) separately and drops the specialty count.
- — Medical coding AI compared: Described the second study's task under a single code system; it now states both readings, since the abstract labels it ICD-9 prediction while the methods score against ICD-10.
- — Imaging AI marketplaces compared: Said every Sectra Amplifier Marketplace app integrates through the Sectra viewer; the source describes a separate Amplifier Integrators track contracted per vendor.
- — Software as a Medical Device (SaMD): Referred to the FDA's Clinical Decision Support Software guidance as the September 2022 edition; the FDA revised it as a January 2026 final guidance.
- — The Kaiser Permanente 2.5-million-encounter ambient scribe deployment, analyzed: Gave the physician-survey figure as 88% and cited a 35-transcript, 48-of-50 accuracy result found in neither source paper; the survey figure is 84%, and the accuracy result was removed.
- — Documentation time and burnout: what the evidence shows: Reported combined Mini-Z and PFI-WE burnout results across both scribes that the source did not report; the page now gives the paper's per-tool estimates.
- — Hallucination and omission rates in AI scribe notes: what the studies measured: Cited a 35-transcript, 48-of-50 accuracy result that the cited source does not contain; it was removed.
Freshness
A field this fast makes accuracy a moving target. Pages carry the date they were last updated, and we revisit them on a schedule — more often for statistics and regulation, less often for stable definitions — and whenever a development on the ground makes an update warranted.
Questions about how we work? Write to hello@aimocs.com.