OpenEvidence

AI clinical search is the entire product.

Oversight rests on citation integrity and on explicit negative scope, and both are real and independently corroborated, which is unusual. The terms of use state that the service is in no way intended to serve as a diagnostic service or platform, to provide certainty with respect to a diagnosis, to recommend a particular product or therapy, or to substitute for the clinical judgment of a qualified healthcare professional.

The company describes the product across its site as an experimental technology demonstrator that does not provide medical advice, diagnosis or treatment. Answers carry inline citations to identified sources, and the independent systematic review confirms those citations are real, finding across eleven studies that the platform avoided fabricated citations. A verified claim is worth considerably more than an asserted one.

Two things hold this below a higher grade, and the first is the more interesting. The same review documents interpretive errors despite accurate citations, which is a distinct and under discussed failure mode: verifiable provenance creates an appearance of checkability that the synthesis layer has not necessarily earned, and a clinician who spot checks the citation rather than the reasoning will not catch it.

Second, the review found the platform often reinforced rather than altered clinical decisions, which is the confirmatory pattern automation bias predicts and the opposite of what a second opinion is for. Nothing located states what the system does when evidence is absent, thin or conflicting.

Proprietary models trained on peer reviewed literature claimed in third party coverage; no model card or technical documentation verified.

The published policy set is real and the one question that matters is missing from it, which is a choice about scope rather than a publishing gap. What exists is substantive: encryption in transit and at rest, conversations private by default, explicit user controls over sharing, a data classification policy, and a retention and disposal policy with deletion on customer request.

A named retention and disposal policy with a deletion route is more than most of this category publishes, and default private rather than default shared is the correct orientation for a product a clinician uses about a real patient. What the trust centre does not state is whether user submitted content is used to train or improve models, and that is the central stewardship question for a clinical artificial intelligence product.

The distinction this index applies is that an absence where no document exists is a publishing gap, while an absence inside documents that do exist and cover adjacent ground is a decision about what to answer. A trust centre carrying a classification policy and a disposal policy has already engaged with the lifecycle; the training question is the next line and it is not there. No model or model family, hosting arrangement or sub processor list was located either. Ask whether prompts and uploaded content train or improve models, whether that is severable, and for a sub processor list.

The independent literature on this vendor is deeper than on almost anything else in this index, and the company did not commission it. A systematic review following PRISMA identified eleven studies published between 2024 and 2026 evaluating OpenEvidence, eight as the primary platform and three as a comparator, searching MEDLINE, Scopus, Web of Science and Google Scholar.

Its findings are mixed and specific: the platform generates evidence supported responses and avoided fabricated citations, performance was strongest in structured guideline based contexts, accuracy varied in complex clinical scenarios, and the platform often reinforced rather than altered clinical decisions.

Named limitations include interpretive errors despite accurate citations, dependence on the available retrieval context, and the difficulty of comparing point in time evaluations of a continuously updated system. Separately, a November 2025 preprint, explicitly not peer reviewed, tested both product modes against 100 subspecialty board style questions using two independent evaluators, with the quick mode scoring 34 percent and the deep mode 41 percent.

Those figures are best read in context rather than as a headline, since the same dataset produced 14 to 46 percent across eleven large language models in the originating work, so they sit inside the established band rather than beneath it. What the vendor publishes itself is the weakest part of the picture: a 100 percent USMLE score, which is precisely the benchmark the subspecialty literature exists to argue is no longer informative. No clinical outcome study, error rate or evaluation output is published by the company.

PHI handling is documented: encryption in transit and at rest, private by default conversations, explicit user controls over sharing, a Data Classification Policy, and a Data Retention and Disposal Policy with deletion on customer request. Held back from A because the trust center does not state whether user submitted content or PHI is used to train or improve models, which is the central stewardship question for a clinical AI product.

The security page states that covered entities as defined by HIPAA may choose to transmit protected health information, that such information is stored, processed and transmitted under a standard business associate agreement or customer specific agreements, and that the company complies with the HIPAA Privacy, Security and Breach Notification Rules. That was announced publicly in April 2025 and remains among the clearest such disclosures in this index. Two findings temper it.

The vendor's own material contradicts itself on the central question. The about page carries a disclaimer stating that user questions and other inputs are not covered by HIPAA and that it is the responsibility of the user to ensure questions do not contain protected health information, which is the opposite of what the security page invites, and the site wide description of the product as an experimental technology demonstrator persists into a 2026 dated page.

A buyer reading two pages of the same site receives two incompatible answers. More useful to a buyer, an individually accepted agreement does not settle the institutional question. MaineHealth publishes a notice instructing its clinicians not to submit protected health information, stating that it holds no agreement with OpenEvidence, that an individually accepted one does not cover the sharing of health system data with external parties, and that it attempted to negotiate and had not reached acceptable terms.

The general point is worth carrying: an agreement accepted by an individual clinician is not the same instrument as a covered entity agreement, and a product adopted bottom up on personal devices routes around the institution that holds the compliance obligation.

Upgraded from C on verification of a full public trust center hosted on Secureframe at trust.openevidence.com, plus a dedicated security page. SOC 2 Type II certification for the Security trust services category, verified as an independent audit of controls operating over an extended period.

Documented controls include AES-256 at rest, TLS 1.2 with SHA256 certificates in transit, a Secure Development Policy, change approval by an independent reviewer, continuous control monitoring, a tested incident response plan with tabletop exercises and lessons learned documentation, formal risk assessments, and cybersecurity insurance. Depth and independence of disclosure both meet the A anchor.

No FDA clearance, none claimed and none apparently required. The clinical decision support exclusion position is strong on the same basis as other reference products in this category: the basis for an answer is an inline citation to identified peer reviewed literature the clinician can open and read, so the recommendation is independently reviewable. The more consequential regulatory fact is a jurisdictional one, and it is the first of its kind recorded in this index.

OpenEvidence terminated access across the European Union and the United Kingdom at the end of April 2026, citing mounting regulatory uncertainty regarding the treatment of AI systems in those markets, including the EU Artificial Intelligence Act. It was first surfaced by trade press in late April 2026 and covered independently by several outlets thereafter, and the platform is now United States centred, with verification built around the US National Provider Identifier.

Under the AI Act a clinical decision support system may be classified as high risk, which carries obligations around transparency, clinical validation, ongoing monitoring and bias evaluation. Stated as structure rather than accusation, the obligations declined are close to the same disclosures this record finds absent on its transparency and governance axes, and a buyer weighing regulatory durability should register the choice.

Keep the comparison honest in both directions: the leading subscription alternative remains available in Europe, but its own generative layer is likewise bounded to individual subscribers in the United States and Canada.

There is one genuine published governance artefact, and it is an unusual one: a public advertising policy governing the conflict of interest the business model creates. Analysis of that policy reports it states advertisements shall not be considered an endorsement, that advertisers cannot influence answers, and that the information system and the advertisement display system are fully unconnected.

A named medical advisory board is published rather than anonymous, with affiliations at major academic medical centres. Both sit above pure assertion. Everything this axis actually asks for is absent: no model card, no bias disclosure, no subgroup or demographic analysis, no error taxonomy and no published output of any internal evaluation. Two structural exposures belong on the record. The direction of influence is asymmetric and only one direction is policed.

The policy addresses whether advertisers can shape the answer. It does not address the reverse, and reporting on the privacy policy describes engagement with particular clinical topics and device data being used to tailor advertising to a clinician's specialty and interests. So advertisers may not shape the answer, but the questions a clinician asks shape the advertising that clinician sees, and that second direction is where the commercial value sits.

And the editorial separation is self policed: no external audit, no published compliance reporting and no third party review of the boundary was located. The existence of the policy is credited; the absence of evidence it is enforced is what holds the grade.

Two passes located no model card, no technical documentation, no evaluation methodology, no accuracy or error rate, no published limitations and no warranty, indemnity or remediation commitment. The only description of the models available is a claim carried in third party coverage that they are proprietary and trained on peer reviewed literature, and a characterisation a company has not published itself is not a disclosure by that company: it cannot be relied on, cannot be dated, and cannot be held against them.

This index has recorded the opposite behaviour as creditable, where a vendor publicly corrected a flattering third party description of its own technology, and the pattern here is the more common one of letting a generous characterisation stand. The gap is more notable than usual given the product's prominence and its use pattern, since clinicians consult it during care and act on what it returns without a reviewer between.

What would fill it is ordinary and available: an accuracy evaluation against a defined question set, a statement of the conditions under which an answer should not be relied upon, and a published position on what the system does when the literature is thin or conflicting. Ask for an evaluation on your own specialty's questions, the limitations documentation, and what the system does when evidence is contested.

Epic embedded enterprise deployments documented at Mount Sinai (seven hospitals) and Cedars-Sinai, the latter with patient context aware queries from the EHR.

Delivery is broad and the infrastructure is named, which is more than most vendors in this category disclose. Web plus native iOS and Android applications, and enterprise deployment embedded in Epic, documented at Mount Sinai across seven hospitals and at Cedars-Sinai with patient context aware queries launched from the chart.

Scale on the AI product itself is the largest in the category: more than 757,000 verified clinicians reported, use claimed across more than 10,000 hospitals and medical centres, and approximately 20 million clinical consultations per month as of January 2026. Hosting is disclosed specifically and by name, on Google Cloud Platform and Vercel, both identified on the security page, which is a level of infrastructure candour almost nothing else in this category offers.

Two things hold it below a higher grade. No data residency statement of any kind was located, no region named and no residency commitment offered, and that matters more here than at most vendors because the same security page invites covered entities to transmit protected health information.

And the geographic footprint contracted in 2026 rather than expanded, with access across the European Union and United Kingdom terminated at the end of April 2026, so this is now effectively a United States product with primary verification built around the US National Provider Identifier. A vendor whose availability map got smaller during the review period is a different procurement risk from one whose map is simply small.

Free individual tier for verified clinicians published; enterprise per seat structure reported without amounts.

Breadth is established through named content partnerships rather than asserted, and the list is specific: official AI partnerships with the New England Journal of Medicine, with JAMA and its eleven specialty journals, with Cochrane Systematic Reviews and with the National Comprehensive Cancer Network for treatment algorithms, plus content agreements and collaborations named across the American College of Cardiology, American Diabetes Association, American College of Emergency Physicians, American Academy of Family Physicians, American Academy of Orthopaedic Surgeons, American Academy of Otolaryngology, American Academy of Pediatrics, NORD, the American College of Obstetricians and Gynecologists, the American Urological Association and the American Academy of Ophthalmology.

More than 300 medical journals in total, alongside FDA and CDC sources. That spans cardiology, endocrinology, emergency medicine, primary care, orthopaedics, otolaryngology, paediatrics, rare disease, obstetrics and gynaecology, urology, ophthalmology and oncology, and it is specialty society endorsement at greater scale than anywhere else in this index. Held below a higher grade on three limits.

Independent evaluation found accuracy fell on subspecialty complexity relative to general clinical questions, so breadth of coverage is established while depth at the subspecialty end is independently questioned rather than demonstrated. The content model is United States centred, citing American guidelines and FDA approvals. And availability narrowed in 2026 with the European Union and United Kingdom withdrawal, so a coverage claim that reads global is in practice national.