In Utah, an AI system can now renew an existing prescription. In the United States, another can screen a retinal photograph for diabetic retinopathy without a clinician reading it. In some clinics, an AI scribe now drafts the consultation note. These are very different responsibilities. What evidence should we expect before handing each one over?
The Utah agreement is between the state's Office of Artificial Intelligence Policy and Doctronic, a company that sells what it calls an AI doctor [1]. The system takes a history from the patient and can renew prescriptions for 192 medicines used for long-term conditions such as hypertension, diabetes and depression. It renews an existing prescription and cannot change it. Physicians hired by the company review its output for the first 250 patients before it acts, and review the next 1,000 retrospectively. A request goes to a physician automatically when, for example, the patient reports a problem with the drug or a change in their health [2].
In April, the Utah Medical Licensing Board published a letter saying it learned of the agreement only after the program had launched. It recommended that the program "be immediately suspended pending further discussion" [1]. As far as I can find, no national regulator has assessed the system.
For an Australian optometrist this is becoming a practical question. AI may arrive in imaging software, in a screening pathway or as a scribe that offers to write the notes. Most optometrists I know read about AI in eye care, if they read about it at all. They do not read the radiology or cardiology literature, and I did not either until I started this series. This issue looks at the rest of medicine first, with the consulting room in view.
What does a regulatory authorisation tell us?
The FDA keeps a public list of the AI-enabled devices it has authorised. The version updated on 4 September 2026 holds 1,614 devices, with decisions up to 29 June 2026 [3]. Authorisations rose from 6 in 2015 to 335 in 2025.
Data as a table
| Year of decision | Devices |
|---|---|
| 2015 | 6 |
| 2016 | 18 |
| 2017 | 27 |
| 2018 | 65 |
| 2019 | 80 |
| 2020 | 114 |
| 2021 | 129 |
| 2022 | 162 |
| 2023 | 226 |
| 2024 | 235 |
| 2025 | 335 |
| 2026, January to June | 181 |
The list has limits. The FDA says it is not comprehensive, and it finds devices mainly by the AI-related terms in their published summaries [3]. A product appears again each time an updated version is authorised, so the entries count authorisations rather than products.
Most of the devices work with medical images. Radiology led the review of three in every four. Ophthalmology led 10, and seven of those are diabetic retinopathy screening systems. The eye was early. In April 2018 the FDA authorised IDx-DR, which gives a diabetic retinopathy screening result without a clinician interpreting the photograph [4].
Data as a table
| Lead review panel | Devices | Percent of 1,614 |
|---|---|---|
| Radiology | 1230 | 76.2 |
| Cardiovascular | 154 | 9.5 |
| Neurology | 73 | 4.5 |
| Anaesthesiology | 30 | 1.9 |
| Gastroenterology and urology | 27 | 1.7 |
| Haematology | 22 | 1.4 |
| Pathology | 10 | 0.6 |
| Ophthalmology | 10 | 0.6 |
| Ten other panels | 58 | 3.6 |
An authorisation tells us that a regulator accepted the case a manufacturer put forward. It does not tell us whether patients benefited. Lin and colleagues read the FDA's public summaries of all 691 AI devices cleared to July 2023. Only 6 reported data from a randomised trial, and 3 reported patient outcomes. Most were cleared because they were substantially equivalent to a device already on the market [5]. The authors note that public summaries may not contain everything a manufacturer submitted.
What can the newest models do?
Most devices on that list do one narrow task. The newest models are general. They read text and images, take a history and reason through a case.
They have improved quickly. Google's Med-PaLM was the first model to pass questions in the style of the United States medical licensing examination. Its successor, Med-PaLM 2, scored up to 86.5 percent on them, and physicians preferred its answers to other physicians' answers on eight of nine measures [6].
In October 2026, Microsoft researchers reported in Science on an AI system that works through a diagnosis one step at a time, asking for history and choosing tests. On difficult teaching cases published in a medical journal it reached the correct diagnosis in 83.2 percent of cases. Generalist physicians given the same cases reached 36.1 percent. The authors note that these results came from a controlled benchmark [7].
Google's Articulate Medical Intelligence Explorer (AMIE) took histories by text chat from actors playing patients, in a randomised, double-blind comparison with 20 primary care physicians. AMIE had greater diagnostic accuracy, and specialist physicians rated it better on 30 of 32 measures. The authors caution that consulting by text chat is unfamiliar in clinical practice [8]. In its first study with real patients, AMIE interviewed 98 patients before urgent primary care visits while a physician watched every conversation. No conversation had to be stopped for safety, and the supervisors noted one hallucination [9]. That study measured safety, conversation quality and acceptance, not diagnoses or outcomes.
The newest version of AMIE also reads what a patient shares during the chat. In a randomised, blinded study of 105 simulated telehealth consultations, the cases included skin photographs, electrocardiograms and clinical documents. Eighteen specialist physicians judged its diagnoses more accurate than those of primary care physicians. They rated it better on 29 of 32 measures, including 7 of 9 that assessed how it reasoned about the images and documents [10]. The authors describe the study as exploratory.
Google has also released MedGemma, a family of medical models that read images and text. They are open, so anyone can download them and build on them. Their performance is reported in a technical report that has not been peer reviewed [11].
Eye care is included. RETFound, built at Moorfields Eye Hospital and University College London, learned from 1.6 million unlabelled retinal images. Adapted to each task, it outperformed comparison models at detecting sight-threatening eye disease and at predicting heart failure and myocardial infarction from the retina [12]. These were tests on stored images. On examination questions, GPT-4 scored 69 percent on a mock ophthalmology examination. Expert ophthalmologists scored a median of 76 percent and ophthalmology trainees 59 percent [13].
Conversational tools are now being built for primary eye care. In Singapore, Meta-EyeFM answers typed questions about a fundus photograph by passing it to one of eight models trained on retinal images. On 91 stored photographs without cataract, it identified 90 of 102 diagnoses (88.2 percent). Four optometrists identified between 71.6 and 81.4 percent, and five ophthalmologists between 74.5 and 84.3 percent. In a separate test, GPT-4o and Gemini were less accurate at detecting the same diseases [14]. The authors write that the model needs further improvement before it is used for screening.
What happens in routine care?
None of these results is a patient outcome. Penda Health runs primary care clinics in Nairobi. In early 2025, it gave half its clinicians, chosen at random, an assistant built on GPT-4o that checked their notes during visits and flagged possible errors. Across 39,849 visits in 15 clinics, independent physicians found 16 percent fewer diagnostic errors and 13 percent fewer treatment errors with the assistant. The share of patients who still felt unwell about a week later did not differ significantly: 3.8 percent against 4.3 percent [15]. OpenAI funded the study, which has not been peer reviewed.
A randomised trial in 16 Penda clinics then made a patient outcome its primary endpoint. Clinical officers were randomised to work with or without the assistant. In 2,000 notes reviewed by physicians, an appropriate diagnosis and treatment plan were recorded more often with the assistant. Treatment failed within 14 days in 2.2 percent of patients with the assistant and 2.0 percent without. The difference was not significant, and there was no safety signal [16].
What happens when AI is tested in a screening program?
Eye care has randomised trials of AI too. Wolf and colleagues randomised 164 young people with diabetes. Every one of those offered an AI eye examination in their own diabetes clinic completed it within six months. Of those given a scripted referral to an eye care provider and education, 22 percent did [17]. That is a process of care. It did not measure vision.
In the Mammography Screening with Artificial Intelligence (MASAI) trial, Gommers and colleagues randomised 105,934 Swedish women [18]. In one group, an AI system sorted mammograms between single and double reading by radiologists and marked suspicious areas. The other group had standard double reading. The primary outcome was the rate of interval cancers, those diagnosed between screening rounds. Some were present but not seen at screening, and some developed after it.
Interval cancers occurred in 1.55 per 1,000 women with AI-supported screening and 1.76 per 1,000 with standard reading. The ratio was 0.88 (95 percent confidence interval 0.65 to 1.18). The trial showed that AI-supported screening was no worse on this outcome, not that it reduced interval cancers. Sensitivity was 80.5 percent against 73.8 percent, specificity was 98.5 percent in both groups, and radiologists read fewer screens. The trial does not show whether fewer women died of breast cancer.
Would you trust it to write your notes?
An ambient AI scribe records the consultation and drafts the clinical note. Lukac and colleagues randomised 238 outpatient physicians to one of two scribes or to usual care [19]. With one scribe, time spent writing notes fell by 9.5 percent. With the other, it did not change significantly. The physicians reported clinically significant errors in the drafts "occasionally".
A better working day for the clinician is worth measuring. It answers a different question from whether the patient received better care. Every draft still has to be read before it is signed.
Where this leaves eye care
The general models can now pass examinations, read images and work through hard cases. In Kenya, an assistant improved clinicians' notes and reduced errors. What has not yet been shown in a trial is that patients do better when clinicians use them. The one such trial I found did not show it. Many retinal models have been published, few have been through a regulator and fewer still have been tested in a randomised trial. The next issues of this series follow that gap from the laboratory to the consulting room.
Conflict of interest, funding and how this was written
I own OptoGuide Pty Ltd, which sells OptoGuide, a clinical decision support product for optometrists. Its clinical recommendations come from fixed rules, and no AI model runs when a clinician uses it.
No one funded this issue. I have no financial relationship with any maker of an AI system named here. I checked every figure against the paper or the register it is cited to, and the errors are mine.
Issue 2 begins in the eye, with a glaucoma model whose sensitivity fell from 95.6 percent in validation to 65.0 percent in two Australian general practices.