Digital health · Clinical decision support
Glass Health
An AI tool that reads a one paragraph patient summary and returns a ranked differential diagnosis, now sold alongside an ambient scribe. It has no FDA clearance, and its founders have said from the start that it does not need one.
One line in, ten diagnoses out
A clinician types a sentence: a 71 year old man with a prior heart attack, short of breath on stairs for a few weeks. That compressed sentence has a name in medicine, the problem representation, and it is what a resident says when handing a patient to an attending.
Glass hands back a list. “Glass analyzes the patient summary and recommends five to 10 diagnoses that the clinician may want to consider and further investigate,” co-founder Dereck Paul told TechCrunch in 2023. The product sorts them into tiers, including a bucket for the serious possibilities a tired clinician might skip past.
That is a small interface with a large regulatory shadow behind it. A list a doctor reviews is one thing. An answer a doctor follows is another. The distance between those two sentences is where this company lives.
The line the product sits on
Congress drew that line in 2016. Section 3060 of the 21st Century Cures Act amended section 520 of the Federal Food, Drug, and Cosmetic Act to pull certain decision support software out of the definition of a device. Four criteria sit at section 520(o)(1)(E). The software must not analyze a medical image or a signal from a diagnostic device. It must display or analyze medical information about a patient. It must support or provide recommendations to a health care professional about prevention, diagnosis or treatment. And, the criterion that does the real work, it must let that professional independently review the basis for the recommendation, so that it is not the intent that the clinician rely primarily on it.
FDA issued a final guidance interpreting those criteria in September 2022, then a revised one on January 6, 2026 under the same docket, FDA-2017-D-6569, superseding it. The revision loosened one point: FDA now says it will exercise enforcement discretion for a tool that gives a single recommendation where only one option is clinically appropriate. It kept the rest. Software meant for time critical decisions still cannot be non device clinical decision support, and recommendations are still expected to rest on what FDA calls well understood and accepted sources. Covington and Burling, reading the revision for clients, noted that it is silent on artificial intelligence, as the 2022 version was.
Nothing in that framework involves an FDA review. A developer reads the four criteria and decides. No public record shows FDA has ever evaluated Glass.
A resident and an engineer
Dereck Paul went to medical school at UCSF and did internal medicine residency at Brigham and Women’s Hospital. What struck him was not the medicine. “I was just totally shocked by how analog things were,” he told NPR in 2023.
He started Glass Health in 2021 with Graham Ramsey, an early engineer at Modern Fertility who had built a telemedicine electronic health record at PlushCare. The first product was not AI. It was a notebook, a place for a physician to store and share their own approaches to diagnosing and treating conditions. It found users on Twitter, among residents and students, and produced a $1.5 million pre-seed round led by Breyer Capital in 2022.
The pivot
Paul has said he ignored ChatGPT when it appeared. Then younger doctors kept telling him it was good at clinical questions. Glass entered Y Combinator’s Winter 2023 batch and shipped Glass AI that January. Paul told NPR the company had about 500 users in December 2022, 2,000 monthly active users in January and 4,800 in February. By September 2023 TechCrunch reported more than 59,000 signups and a $5 million seed round led by Initialized Capital.
The technical answer to the obvious objection was retrieval. Rather than trust a general model’s memory, Glass wired it to a database of clinical guidelines written and maintained by academic physicians, about 20 to 30 of them, working part time the way they would for a medical journal.
What it became
By 2026 the product is broader: ambient scribing that drafts the note, differential diagnosis, plan drafting, and a developer API selling the same machinery to other software companies. Pricing is published, unusual here: free, then $20, $90 and $200 a month, with EHR workflows for Epic, eClinicalWorks, athenahealth and Elation at the top tier.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| FDA device records | No 510(k) and no PMA listing Glass Health as applicant, openFDA, queried Sep 23, 2026. Per NPR in April 2023, the company said it was not planning to seek clearance. | Public record |
| Non device CDS status | Meeting the four criteria at section 520(o)(1)(E) is a determination a developer makes about its own product. FDA does not clear or confirm non device status, and no public record shows any FDA evaluation of Glass. | Public record |
| Company regulatory claim | A 2026 Glass Health guide says the tool is designed to meet the exemption criteria and the clinician makes all clinical decisions. No third party attestation is offered. | Company-stated |
| Radiology appropriateness | Journal of the American College of Radiology, 2024. Across 1,075 scenarios from 11 ACR panels, Glass AI scored 2.32 against ChatGPT at 2.08 on a 0 to 3 scale, P = .002, and beat it in 10 of 11 panels. | Peer reviewed |
| Neuroimaging selection | Same journal, 2023. Across 147 conditions Glass AI scored 1.83 and ChatGPT 1.75, no significant difference. An experienced neuroradiologist scored 2.20 and significantly outperformed both. | Peer reviewed |
| Hard diagnostic cases | The Neurologist, 2024. On four NEJM clinicopathologic conference cases, GLASS AI got 1 of 4, ChatGPT-4 got 0 of 4, a rule based diagnostic generator got 4 of 4. A 2026 follow up in Diagnosis on two rare diseases found the correct answer came from the generator, not from expert opinion or either AI. | Peer reviewed |
| User perception | Medical Teacher, 2024. Of 202 first year medical students, 73 responded: 96 percent said Glass AI increased confidence in the diagnosis, 43 percent thought it lacked sufficient explanation, 68 percent expressed risk concerns for the physician workforce. | Peer reviewed |
| Internal benchmark scorecard | The developer API page puts Glass 5.5 Deep Reasoning at 88.6 percent overall and 84 percent on NEJM clinicopathologic conferences, ahead of named frontier models. Methodology and scoring are not published, and no outside party has reproduced it. | Company-stated |
| Outcomes evidence | No registered trial found, and no published study of the product’s effect on diagnosis, treatment or patient outcomes in live care. | Not found |
| Security attestation | HIPAA deployment under a business associate agreement is described by the company. No public SOC 2 report or HITRUST certification found. | Not found |
Read plainly: the independent evidence is real, published, and modest. Four peer reviewed papers exist, none funded or authored by the company, and they point two ways. On structured appropriateness tasks Glass beat a general chatbot. On hard diagnostic cases, and against a human specialist, it did not. The company’s own scorecard reports far stronger numbers, on a newer model and an undisclosed method. Both can be true. Only one has been checked by someone else.
What to watch
- Whether Glass publishes its benchmark methodology, or submits its model to an evaluation it did not design.
- Whether ambient scribing pulls the company toward documentation, where time savings are now well evidenced, and away from the diagnostic claims that carry regulatory weight.
- Any named health system deployment. The 2024 East Coast integration and the 2023 waitlist of 15 institutions were never identified.
- FDA’s planned revision of its device software functions policy, flagged as the place generative AI decision support may finally be addressed.
- A funding event. Nothing announced since September 2023, and no Form D ever filed.
In their words
“I was just totally shocked by how analog things were”
Dereck Paul, co-founder and CEO, NPR, 2023 · Independent
“We need these folks not in burnt-out states, trying to complete documentation”
Dereck Paul, NPR, 2023 · Independent
“Glass analyzes the patient summary and recommends five to 10 diagnoses that the clinician may want to consider and further investigate.”
Dereck Paul, TechCrunch, 2023 · Independent
“You have to supervise it, the way we supervise medical students and residents, which means that you can’t be lazy about it”
Dereck Paul, NPR, 2023 · Independent
“AI clinical decision support is a new product category in healthcare.”
Dereck Paul, Initialized Capital, 2024 · Interview
“Anything that involves decision-making about a patient’s care is something that has to be treated with extreme caution for the time being.”
Pearse Keane, University College London, NPR, 2023 · Independent
“I would express considerable caution using this in a clinical scenario for any reason, at the current stage”
Marc Succi, MD, Massachusetts General Hospital, NPR, 2023 · Independent
“Most AI-powered CDS tools, including Glass Health, are designed to meet these exemption criteria.”
Glass Health clinical decision support guide, 2026 · Company-stated
Related companies
Sources
- Public recordClinical Decision Support Software guidance, docket FDA-2017-D-6569
- Public recordopenFDA 510(k) and PMA databases, applicant query
- Peer reviewedThe Application of Large Language Models for Radiologic Decision Making
- Peer reviewedUse of Large Language Models to Predict Neuroimaging
- Peer reviewedNeurological Diagnosis: Artificial Intelligence Compared With Diagnostic Generator
- Peer reviewedComparison of AI, diagnostic generator, and expert opinion in difficult neurologic diagnosis
- Peer reviewedMedical students’ perceptions of an AI assisted diagnosing program
- IndependentDoctors are drowning in paperwork. Some companies claim AI can help
- IndependentGlass Health is building an AI for suggesting medical diagnoses
- InterviewFounder Spotlight: Glass Health CEO Dereck Paul
- Independent5 Key Takeaways from FDA’s Revised Clinical Decision Support Software Guidance
- CompanyGlass Health Raises $5M for AI-Powered Clinical Decision Support Platform
- CompanyDeveloper API page with benchmark scorecard
- CompanyOpenEvidence vs Glass Health page with published pricing
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
