Three insights into Chat GPT and AI large language models in optometric practice

How artificial intelligence (AI) can be safely incorporated into optometric practice was discussed during a peer discussion on the Saturday of 100% Optical 2025 (1 March).

The session took place in the Optical Academy, and was run by optometrist Kishan Devraj, an honorary clinical research fellow at Moorfields Eye Hospital.

Devraj’s PhD, completed at UCL Institute of Ophthalmology, focused on harnessing smartphone technology to assess the lifestyle impact and quality of life with eye disease.

OT noted three key insights on AI language models that were highlighted by Devraj during the discussion.

1 More regulatory frameworks are needed

Citing a 2024 study from the University of Cambridge and the University of Birmingham, Devraj noted that GPT-4 can exceed non-specialised doctors in diagnosis for patients – although oversight by a clinician is still needed.

“[The AI] was very close to that of specialist ophthalmologists. For a given scenario, a patient with certain symptoms, it could suggest whether it was an emergency that needed same day specialist care, or if it was something that a GP could handle,” he said.

All similar studies are highlighting the need for clinical oversight, he emphasised.

Devraj also noted that Moorfields Eye Hospital has utilised data from smartphones to assess how eye disease can affect patients’ quality of life.

This technological progress comes with many questions and limitations, including whether it could lead to a reduced demand for trained optometrists, as well as legal issues such as who has accountability if the AI gets something wrong, Devraj said.

Devraj explained that AI technology often struggles with rare conditions, and that even a 1% misdiagnosis rate could lead to significant issues.

He also noted that accessibility is a problem when it comes to incorporating AI in healthcare settings.

He emphasised that, because of these significant potential pitfalls, updated regulatory frameworks are needed.

Failure to upskill at the same rate that regulation and technology advances means there will be a risk of the profession deskilling, Devraj believes.

“But if there is regulation around how we upskill, and we do upskill, and we can perform more eye treatments, then it positions us very, very well for some of these technologies,” he added.

2 Combining language models with clinical guidance makes the AI more reliable

“What you get out of language models depends on what you put in,” Devraj said.

As part of the peer discussion, he asked attendees to scan a QR code and paste a prompt into Chat GPT.

He then explained the methodology behind his app, Ask fellow optoms, which allows clinicians to post questions that can be answered both by optometry peers and by an AI language model.

The app’s AI is trained purely on General Optical Council and College of Optometrists guidance, Devraj said.

Clinicians should not rely purely on language models, he said – but combining them with clinical opinion enables their output to become more powerful.

He emphasised that Ask fellow optoms should be viewed as “purely educational,” and should not be used in the management of patients.

3 A full-scale rollout of AI is not imminent

Devraj emphasised his belief that, when it comes to large-scale AI language models, “it will be many years before we see anything in practice.”

There are “big processes involved” with any large-scale technological development, he said.

He used the example of the lightbulb, which was developed years before it became a staple in people’s homes, partly because public approval takes time to develop and be maintained.

“We had to have a system in place,” Devraj said. “We needed networks. We needed a grid. We needed regulation. We needed public approval, because people still wanted to use candles.”

The purpose of the 100% Optical session was “to show you what's out there and what’s possible, and where things could be the next five or 10 years,” he added.