By: Dr. Alan N. Glazier OD, FAAO, Dipl. ABO
Imagine opening your inbox tomorrow to news that an AI system has “solved” myopia. Not treated it, not slowed it. Solved it: a model has ingested every axial-length measurement, every refraction and the human genome of every single myope and produced a predictive rule so accurate that at a child’s first eye exam the doctor can tell a parent exactly how nearsighted that child will be at twenty.
As remarkable as this sounds, it also raises uncomfortable questions, most importantly “would having this information mean we understand the eye any better than we did before?” That question sits at the heart of a recent video by a mathematician who supervises PhD students
and is frustrated by how the public talks about AI in his field. His argument: We have confused answering questions with advancing knowledge, and that confusion could cost us the same people who are responsible for advancing our profession. While the subject of the video was pure mathematics, extrapolating the issues with eye care in mind, it is one of the clearest warnings yet about where clinical science, thus clinical practice of optometry might be headed. This article takes his argument and extrapolates it into optometry and vision science. The conclusion is not that AI is bad for eye care. It is that the value of research has never lived only in answers, and a profession that forgets this risks losing far more than jobs.
Solving a problem is not the same as advancing science In 1736, Leonhard Euler (pro. “oiler”) took up a question that was, by any practical measure, pointless. The city of Königsberg had seven bridges linking two islands to the riverbanks. Could you walk across all seven without crossing any bridge twice? Euler proved you could not. The answer changed nothing. What mattered was how he got there. To solve it, Euler threw away almost everything about the city, distances, streets, the shape of the river, and kept only which pieces of land connected to one another. That abstraction became what we now call graph theory, which underpins computer networks, internet routing, logistics, and countless algorithms. Had a machine simply told Euler “No,” we would have the answer and possibly none of the mathematics.
The video’s author provides another example. Picture an American flag appearing on Mars with no rocket, no new materials, no life-support systems behind it. Technically an achievement. Practically, humanity would be no closer to becoming a spacefaring civilization. The Apollo
program mattered not only because a man reached the Moon, but because of the guidance computers, materials and engineering built along the way and how those advancements could be and have been applied to many different scientific disciplines, advancing them in the process. Modern medicine has always worked the same way. A clinical question often prompts discovery, not the other way around. The struggle to answer it produces new instruments, new models of disease and new ways of thinking. Those have implications far beyond the original query.
If AI takes us straight to the answer and skips the struggle, we lose the intermediate stops on the ride that helped us get where we want to go. Sometimes we think we know where that is, but what we learn along the way shows us the intended destination isn’t really what we are searching for.
Vision science has its own seven bridges Eye care is full of discoveries that arrived sideways, while someone was chasing a different
question. A few that every optometrist now relies on:
| Original question | What came out of the struggle | Where it lives in practice today |
| How do you find faults in optical fiber and measure tissue with light echoes? | Low-coherence interferometry, adapted by engineers and clinicians at MIT into optical coherence tomography (1991) | OCT is now standard in nearly every optometric practice for glaucoma, macular and retinal disease |
| Do tumors need their own blood supply? | Judah Folkman’s angiogenesis hypothesis (1971), later the discovery of VEGF | Anti-VEGF injections for wet AMD and diabetic macular edema, a cancer question became a vision-saving therapy |
| How is the visual cortex wired? | Hubel and Wiesel’s deprivation studies in the 1960s revealed the critical period | The modern understanding of amblyopia and why early detection in children matters |
| How do astronomers see through a turbulent atmosphere? | Adaptive optics, borrowed for the eye in 1997 to image individual living cone photoreceptors | High-resolution retinal imaging and wavefront science behind aberrometry and refractive surgery |
| Why do some blind mice keep normal circadian rhythms? | Melanopsin and intrinsically photosensitive retinal ganglion cells (ipRGCs), characterized around 2002 | The science behind light and sleep, myopia-light research, and the blue-light debate in our exam rooms |
| Why are certain breeds of French dogs (Briards) born blind? | Decades of RPE65 research and gene-delivery work | Luxturna (2017), the first FDA-approved gene therapy for an inherited disease |
Notice the pattern. In almost every case, the original question was narrow, obscure and not infrequently from an entirely different field. The value came from developments along the way to answering the question: the instruments, models, knowledge and unforeseen opportunities the query forced into existence. Now imagine an AI had simply answered each of those questions. “Yes, tumors need angiogenesis.” “Visual cortex development has a sensitive window.” Correct answers, delivered in seconds. Would anyone have built the interferometer, purified the growth factor, or bred the dogs? Maybe. Maybe not. The honest answer is that we do not know, and that uncertainty is exactly the point.
Eye care is already where the flag is being planted
Optometry is not a bystander in this story. It is one of the places AI arrived first. In 2018, the FDA authorized IDx-DR, an autonomous system that screens for diabetic retinopathy without a clinician reading the image; the first autonomous AI diagnostic cleared in any field of medicine. The same year, researchers showed that deep learning could estimate age, blood pressure and cardiovascular risk from a fundus photograph alone. “Oculomics” was born, and the retina became a window onto the whole body.
These are genuine achievements, and many of them will help patients. But some of them have the shape of the flag on Mars. A model can predict cardiovascular risk from a retinal image with impressive accuracy, yet we often cannot say what in the image it is reading. The answer is clear, yet the understanding- the “method” behind it is not.
Eye care has a ready-made version. An algorithm trained on one population, one camera and one clinic workflow can perform beautifully in validation and stumble in the field. Field studies of retinal AI have already shown how image quality, lighting and clinic routines can erode performance that looked excellent on paper. Recognizing that a pigmented fundus, an unusual optic disc or a rare dystrophy sits outside a model’s experience is not a calculation. It is clinical judgment, and judgment is built over years, by people who have done the work themselves.
The 17-year-old in the waiting room
The most unsettling part of the video is not about answers at all. It is about a teenager. Picture a 17-year-old fascinated by how the eye works. Maybe she has worn glasses since grade school and remembers the first time the world came into focus. She is considering optometry, or vision research. Ahead of her lie four years of undergrad, four years of Optometry school, perhaps a residency or a PhD, and real debt. Then she opens her phone. AI outperforms eye doctors at detecting disease. AI predicts heart attacks from a photo of the retina. AI refracts patients without a clinician. Every week, another headline.
Would she still choose this path? Many would. But the video’s author makes a point every clinical educator will recognize: training is not a weekend hobby. It demands years of commitment under real uncertainty. A large part of what sustains people through it is the belief that they will one day contribute something, diagnose what others missed, understand what no one understood, help a patient no algorithm was built for.
As a PhD supervisor, he describes months of work with no visible progress, and the doubt that comes with it. Any optometry resident who has wrestled with a puzzling visual field, or any vision scientist whose experiment has failed for the fifth time, knows that feeling. What carries you through is the possibility of a breakthrough. Now imagine discovering, in the middle of that struggle, that a model has already produced your answer. Then the next one. How many times before a talented student concludes she is running a race she’s already lost? The danger is not that students will dislike vision science. It is that they will be persuaded there is no future in contributing to it.
This is bigger than jobs
It’s tempting to dismiss these worries as clinicians protecting turf. The video’s author was angered by exactly that reaction from mathematicians’ critics, and the same dismissal will surely be aimed at optometry. Of course people worry about their livelihoods, that is expected. But the larger risk is to clinical progress itself, and it works in at least four ways.
The decline would be silent at first. Clinicians and scientists do not appear out of nowhere. They come from an ecosystem of students, residents, faculty, and mentors built over generations. If fewer talented students enter it over the next 20 years, nothing visibly breaks. Clinics stay open. Journals keep publishing. AI produces more findings than ever. Only slowly does the pool of people who can independently understand, test and extend that work begin to shrink. AI depends on the experts it might displace. Every retinal-imaging model learned from images graded by human experts. Its “ground truth” is a clinical consensus. Fewer deep experts mean weaker labels, weaker validation, and fewer people able to tell when a model has quietly drifted. The tools would erode the very foundation they were built on.
New disease does not come pre-labeled. Models learn from what has already been described. Much of clinical history begins with an attentive clinician noticing something that did not fit, as when physicians in the early 1950s connected the epidemic of blindness in premature infants to the oxygen they were given in incubators. A system trained on yesterday’s knowledge is unable to anticipate what happens tomorrow and make the adjustments the human brain is capable of. We could end up knowing more and understanding less. The video’s starkest warning is that of a future with more knowledge than any generation before us, with fewer and fewer people capable of understanding it. In medicine, that is not an abstraction. It means treatment guidelines no one can fully interrogate, devices no one can independently audit, and a profession dependent on systems it can no longer challenge.
What the profession can do about it
None of this is an argument for rejecting AI. Autonomous screening already reaches patients who would never have seen an eye doctor, and that is a real benefit to society. The task is to adopt these tools without hollowing out the human expertise and experience that makes them
trustworthy.
• Change the story we tell students. The message should not be “AI will do your job” or
“AI will never replace you.” It should be that the field needs people who can question,
validate and extend what machines produce. That is harder and more interesting work
with a tool our predecessors could not have imagined to support their work, not easier
and less interesting.
• Value understanding, not only answers. When an AI result makes headlines, journals,
schools and trade media can ask the follow-up question: what new mechanism,
instrument or insight came out of it? What else can we apply that solution to?
• Keep students doing the hard thinking. Optometric education should still require
students to reason through a case before they see the algorithm’s verdict. Challenge is not
a bug in training; it is how clinical reasoning and judgment forms.
• Protect the research pipeline. Residency and clinician-scientist tracks, vision-science
PhD funding, and mentorship deserve deliberate support now, while the effects of a
shrinking pipeline remain invisible.
• Put clinicians in the loop of AI development. A “clinician in the loop” is going to
complete it.
• Demand explainability where it matters. A prediction we cannot interrogate is a flag
on Mars. Funding research into why a model works is how AI results turn into novel
vision science solutions instead of just new outputs.
The answer was never the point
Euler’s bridges gave us graph theory. A cancer researcher’s hunch about blood vessels gave us anti-VEGF therapy. A puzzle about blind mice that still kept time gave us an entirely new class of retinal cell. None of those outcomes was the answer to the question that started it.
Eye care has transformed how humanity sees because generations of clinicians and scientists did the slow work of understanding, building on those who came before, and because new students kept choosing to join them. AI can make that work faster, broader and more accessible. It cannot replace the people who know what the answers mean.
Those are the two things optometry cannot afford to lose: the discoveries made along the way, and the next generation willing to go looking for them.
Reference: “You Truly Don’t Understand What Just Happened to Mathematics” – Professor
Giordano Scarciotti – https://www.youtube.com/watch?v=KqiDqJMxg6o
