Glass Health Review 2026: An AI Scribe With a Differential Diagnosis Engine (Read the Terms First)

Glass Health is unusual among the tools in our directory. Most AI medical scribes write the note and stop there. Glass writes the note and also reasons about the patient. It produces a differential diagnosis, drafts an assessment and plan, and answers clinical questions with citations. It has a free tier and posts its prices publicly, which is rare in clinical decision support.

Over the past year Glass has also added two things that are easy to miss on a pricing page: advertising on its two cheapest plans, and terms of service that let it train its AI models on the patient information you enter. Neither one is a scandal, and both are disclosed. But anyone recording patient visits should read them before signing up.

This review covers what Glass does, what each plan costs, the independent evidence, and the four documents a practice should read first. Everything below comes from Glass's own pricing page, Terms of Service (last updated July 1, 2026), Privacy Policy (last updated July 5, 2026), Business Associate Agreement and Advertising Policy, all checked on September 23, 2026.

The short version

• What it is: an ambient scribe combined with a differential-diagnosis and evidence engine. Few tools in our directory do both. The others that come close are Suki, Tali AI and Abridge.
• Price: Lite is free, Starter is $20/month, Pro is $90/month and Max is $200/month. EHR integration comes only with Max.
• The catch on cheap plans: Lite and Starter are ad-supported. Pro and Max are ad-free.
• The catch in the contract: the individual Terms of Service authorize Glass, as your business associate, to use patient information to train and fine-tune its models. The same terms let Glass sell data once it has been de-identified. The Business Associate Agreement is opt-in.
• Evidence: two independent studies found Glass beating ChatGPT on specific tasks. We found no prospective study of patient outcomes.
• Best fit: primary-care physicians and nurse practitioners who want a reasoning partner as well as a note-taker, especially if they buy through a practice agreement and not an individual click-through sign-up.

Who makes it

Glass Health launched in 2021. Its founders are Dereck Paul, then a medical student at UC San Francisco, and Graham Ramsey, an engineer who came from Modern Fertility. TechCrunch reported in September 2023 that the company had raised $6.5 million in total. Breyer Capital led a $1.5 million pre-seed round in 2022, and Glass went through Y Combinator's Winter 2023 batch. Initialized Capital is also listed among its backers. At that time Glass reported more than 59,000 registered users.

The product began as a way to type in a patient summary and get back a ranked differential. Ambient scribing, EHR integration, a developer API and, most recently, a consumer app for patients were all added later.

What it actually does

Ambient scribing. You record the visit and Glass drafts the note using templates you can customize. Language options include English, several bilingual modes such as English + Spanish, and a range of other languages.

Differential diagnosis. Glass builds a differential from the live encounter, from typed clinical context, or from uploaded records. It groups the results into Most Likely, Expanded and Can't Miss tiers. The Can't Miss tier is the most clinically interesting feature. It is a structured prompt to consider the dangerous diagnoses that anchoring can hide.

Assessment and plan drafting. Glass links its reasoning to a draft A&P that includes citations to guidelines and literature, so you can check its sources.

Clinical Q&A. You can ask free-text questions and get cited answers. This makes Glass a competitor to reference tools as well as to scribes.

EHR integration (Max only). Glass lists athenahealth, Epic, eClinicalWorks and Elation. Integration runs through SMART on FHIR and is described as "an assisted implementation," meaning it is a project and not a switch you flip. Glass has also written guides for Epic, athena and eCW. On the three cheaper plans you copy and paste into your EHR.

Pricing

These are the prices on Glass's clinician pricing page as of September 23, 2026:

• Glass Lite: free. Glass describes it as "ad-supported." It includes limited scribing and limited clinical decision support.
• Glass Starter: $20/month. Standard scribing and CDS, plus customization. It is still ad-supported.
• Glass Pro: $90/month. Extended scribing and CDS, plus Glass's "Advanced Model." This is the first ad-free tier, and Glass marks it as the recommended plan.
• Glass Max: $200/month. Maximum usage, "Max Mode," premium AI models and EHR integration. It is ad-free.

Glass's upgrade flow also offers annual billing at about 10% off, which works out to $18, $81 and $180 per month for Starter, Pro and Max. Glass does not publish exact usage caps for "limited," "standard," "extended" and "maximum," so run your real daily volume on a trial before you commit.

For comparison, Freed Core is $79/month for unlimited scribing and OrbDoc is $199/month. Neither includes a diagnosis engine. Our scribe pricing guide covers the whole market, and the free and low-cost scribe roundup covers the free tiers.

Ads in a clinical tool: what Glass's policy says

The words "Ad-Sponsorships" appear in the feature lists for Lite and Starter. Glass's Advertising Policy says ads may include "display placements, sponsored modules, promoted links, branded resources, and similar materials." It then makes three commitments:

"Glass Health's clinical AI outputs are generated with no input from the advertising system. Our clinical AI has no access to information about which ads are being served, which advertisers are active, or whether advertising is present at all."

"All Advertising on the Platform is clearly identified as paid or sponsored content and is visually distinct from clinical output."

"Is protected health information shared with advertisers? No. … The advertising system does not have access to any identifiable patient health information."

These are the right commitments to make, and we have no evidence that Glass breaks them. Still, keep in mind what a policy can and cannot promise:

• The obvious advertisers in a clinical tool are drug and device makers. Even if the AI's answer is untouched, a sponsored module for a branded drug that sits next to a differential diagnosis is still in your line of sight while you decide.
• Glass can change the policy by posting a new version. The policy says so directly. Screenshot the version you relied on.
• If you would rather not have the question at all, $90/month removes it. For a solo practice that uses Glass every day, ad-free Pro is the plan we would actually recommend.

The terms most reviews skip

Four documents govern what happens to your patients' data. We read all four.

1. The individual Terms of Service authorize model training on PHI

Section 11 of the Terms of Service (July 1, 2026) covers the clinician service. Section 11.1 gives Glass a "worldwide, royalty-free, fully paid-up, sublicensable, transferable" license to what you enter "for any legal purpose." It lists these purposes, among others:

"(b) improve or modify the Clinician Service (including by training Glass Health's artificial-intelligence and machine-learning models); (c) aggregate or de-identify data, which Glass Health may use, retain, disclose, sell, and otherwise process for any lawful purpose without restriction"

Section 11.2 then applies this to protected health information directly. By accepting, you as the covered entity authorize Glass, "as Customer's business associate, to use and disclose such PHI … including for the operation, maintenance, support, security, evaluation, improvement, training, fine-tuning, and development of the Clinician Service, Glass Health's artificial-intelligence and machine-learning models, and Glass Health's other products and services."

The terms describe these uses as the "proper management and administration" of Glass and as "data aggregation" under HIPAA, citing 45 CFR 164.504(e). That regulation lets a business associate agreement permit PHI use "if necessary … for the proper management and administration of the business associate." It also allows data aggregation "relating to the health care operations of the covered entity." Whether training a vendor's commercial models fits inside those two categories is a question for your compliance counsel, not for a directory. It is a broader reading than you will see in most scribe contracts. Section 11.2 also has you represent that you have "obtained all consents, authorizations, and waivers, and made all required notices" needed for these uses. That puts the responsibility on your practice.

2. Deleting your data doesn't remove it from trained models

The Privacy Policy (July 5, 2026) says that trained models are "the sole and exclusive property of Glass Health." It also says your deletion rights "do not extend to, and we are not able to remove, extract, or 'unlearn' Your Data from, models or other materials that have already been trained or derived." This is true of any vendor that trains on customer data, and Glass deserves credit for stating it plainly. It does mean that deleting your account later cannot undo training that has already happened.

3. The BAA is opt-in, so opt in

Glass publishes a HIPAA Business Associate Agreement. It defines the covered entity as "the entity that executed the Services Agreement and opted in to this BAA." The Privacy Policy adds that when you enter patient information, Glass "may act as a 'business associate'… A BAA or other Separate Agreement may or may not be in place."

In practice: don't record a single patient visit until you have confirmed the BAA is executed on your account, and keep a copy. Our HIPAA basics for AI tools explains why the BAA, and not a "HIPAA compliant" badge, is what counts.

4. The organization contract is narrower, so negotiate it

Glass has a separate Healthcare Organization Terms of Service for practices that buy seats through an order form. In that contract the data license runs "solely to" a list of purposes. Model training is described in terms of non-identifying "Usage Data" and de-identified "Aggregated Data." We did not find the explicit authorization to train on PHI from Section 11.2 of the individual terms. If you have more than one or two clinicians, buy through the organization agreement and ask for a written no-training-on-PHI clause. That is standard practice for enterprise AI purchases, and Glass already has the contract structure for it.

Glass is non-device CDS, and its terms say so

Glass does not claim FDA clearance. Its Terms of Service instead follow the four statutory criteria for non-device clinical decision support almost word for word, the same framework we covered in the FDA's 2026 CDS guidance. By signing up you acknowledge that:

• the service "is not intended for use with inputs that include images, patterns, or signals from a signal acquisition system," so don't paste in ECG tracings or images for interpretation;
• the output is "not intended to be primarily relied upon" for diagnosis or treatment decisions;
• you must "disclose the use of generative artificial intelligence in such clinical diagnosis or treatment decisions where required by Law," which ties back to the state AI disclosure laws; and
• it "is not intended to support time-critical decision-making."

The last point matters because Glass also markets itself for emergency and hospital medicine. You can use it in the ED, but under its own terms the Can't Miss list is a way to cross-check your thinking after the fact. It is not a resuscitation aid.

What the evidence says

We searched Europe PMC for peer-reviewed studies that name Glass Health. Two independent studies tested it directly:

• Radiology imaging selection (J Am Coll Radiol, 2024). Zaki et al. tested "Glass AI" and ChatGPT on 1,075 clinical scenarios from 11 American College of Radiology panels and scored their choices against the ACR Appropriateness Criteria. Glass scored 2.32 vs. 2.08 on a 0–3 scale (P = .002) and outperformed ChatGPT in 10 of 11 panels (PMID 38224925).
• Endodontic diagnosis (Niger J Clin Pract, 2026). Alnassar ran 16 published case reports through four chatbots three times each. Glass placed second on accuracy (mean 2.35) behind Gemini 2.5 Pro (2.81) and ahead of ChatGPT 4.5 (1.96). Its consistency across repeated runs was only 43.8% (PMID 42241195).

Both studies used vignettes, both tested earlier versions of the product, and neither measured patient outcomes, diagnostic errors in real clinics or time saved. We found no prospective clinical study of Glass's differential diagnosis in practice. The endodontic study's consistency figure is the most useful finding for a clinician: the same case can produce a different answer on another run. Treat the differential as a checklist to compare against your own thinking, not a ranking.

For the scribe half, the general research on ambient documentation applies, including its known failure modes. Our summary of what the research says about AI scribe accuracy is a good primer.

Pros and cons

Pros

• Scribing and diagnostic reasoning in one tool, at a price small practices can afford
• Public prices and a real free tier
• Cited outputs you can check, plus a Can't Miss tier built to counter anchoring
• EHR integration with Epic, athenahealth, eCW and Elation (on Max)
• Unusually candid legal documents that answer the hard questions directly

Cons

• Ads on Lite and Starter
• Individual terms authorize training and fine-tuning on PHI, and de-identified data can be sold
• The BAA is opt-in and not automatic
• EHR integration only on the $200 Max plan, and usage caps aren't published
• No prospective clinical evidence, and low run-to-run consistency in one independent study
• Its own terms say it is not for time-critical decisions

Who should use it

A good fit: family medicine physicians and nurse practitioners in primary care who see undifferentiated complaints and would value a second set of eyes on the differential. Small clinics that will sign the organization agreement and negotiate the data terms also fit well.

Probably not a fit: practices whose compliance policy prohibits vendor training on PHI and who cannot get a contract exception; anyone who wants a pure scribe (a scribe-only tool like Freed is cheaper per feature); and behavioral health practices, where a dedicated therapy-note tool and our therapy AI comparison are a better place to start.

A 15-minute pre-purchase checklist

1. Trial on Lite with de-identified test cases, not real patients, so you can judge the differential quality before any PHI is involved.
2. Opt in to the BAA and save the executed copy before your first real visit.
3. Decide on the training question. If your policy forbids it, go straight to the organization agreement and ask for a no-training clause in writing.
4. Choose Pro or higher for clinical use if ads next to clinical output are a problem for you.
5. Check patient consent. Recording laws and state AI-disclosure rules still apply. See our consent and recording guide.
6. Measure a week of real volume against the plan's usage tier before you pay annually.

For side-by-side options, see clinical decision support tools for family medicine and for nurse practitioners, or browse every AI scribe for small clinics.

---

This review is for informational purposes only and is not legal, compliance or medical advice. Prices, plan features and contract terms were taken from Glass Health's website and legal documents on September 23, 2026, and may change without notice. Glass Health did not review or sponsor this article, and MedAI Directory has no affiliate relationship with Glass Health. Verify current terms directly with the vendor, and have your compliance counsel review any BAA and data-use provisions before you enter patient information.