If AI refraction and retinal-scan reads produce the prescription and the screening, is the in-person optometrist the gatekeeper or just the legal signer?

The shift

Two of the core technical tasks of a routine eye exam — measuring refractive error and interpreting fundus images — flip from scarce trained-clinician time to abundant, fast, and near-free. Online and autorefractor-driven refraction can output a glasses prescription without a clinician measuring anything, and AI retinal screening reads fundus photos at a level that is at or near specialist agreement for the common referable conditions (diabetic retinopathy, and increasingly AMD, glaucoma suspects, vessel signs). The measurement and the read stop being what a human is scarce-and-needed for.

The axioms

• Producing an accurate glasses prescription requires a trained clinician performing the refraction in person — the technical exam is the optometrist's core value.
• The exam is worth booking because a person must operate the equipment and measure the eye.
• The routine eye exam is bundled with, and paid for by, dispensing glasses and contacts — the retail margin funds the clinical visit.
• Reading a fundus image for disease requires scarce trained interpretation.
• A licensed, accountable human must own any diagnosis and any referral for disease.
• Some findings only surface through a physical exam — slit-lamp, dilation, pressures, hands, and a clinician looking at more than a retinal photo.
• Catching incidental and systemic disease in the eye depends on a person examining the eye during the glasses visit.
• Trust and continuity of care come from a relationship with a person over time.

Invalid axioms

1. Producing an accurate refraction is skilled clinical work and the optometrist's core value. Autorefractors and online/remote refraction produce a usable glasses Rx without a clinician doing the measurement; for straightforward eyes the technical accuracy is good enough. The habit-trap: the profession and the fee schedule still treat "the refraction" as the billable core of the visit, and patients still book "an eye exam" meaning "get my number," when the number is now the abundant part. The remaining value migrated to what the machine can't sign off on — and the pricing hasn't moved with it.
2. Reading a fundus image requires scarce trained interpretation. AI screening reads the common referable retinal conditions at or near specialist agreement and never gets tired or backlogged, so the read itself is no longer the bottleneck. The habit-trap: screening capacity is still rationed as if each image needs a scarce human pass, when the scarce act has moved to adjudicating the flagged cases and owning the misses — not doing the first read.
3. The exam is worth booking because a person must operate the equipment. The equipment increasingly operates itself and can be self-serve or remote. The habit-trap: the whole retail-optometry footprint — a clinician's chair time gating a glasses purchase — is built on the assumption that measurement requires an appointment, which is exactly the step being removed.

Unchanged axioms

1. A licensed, accountable human must own the diagnosis and the referral. An AI screen can flag; it can't be liable for a missed melanoma, a swollen disc, or a retinal detachment. When the read is wrong, someone with a license and a name answers for it, and that doesn't transfer to a model no matter how good the read gets. This is legal and institutional, not a capability gap.
2. A fundus photo is not a full eye exam, and some disease only shows on physical examination. Slit-lamp findings, anterior-segment and lid pathology, intraocular pressure in context, peripheral retina, binocular vision, and the decision to dilate — much of this is not in a single retinal image the AI reads. Narrow-band screening catches what it was trained to catch and is blind to the rest; a camera pointed at the macula does not see the peripheral tear. The scarce thing is a clinician looking at the whole eye, not the read of one photo.
3. Judgment on the ambiguous or atypical case stays human. The patient who doesn't fit the pattern — conflicting signs, a subtle asymmetry, a systemic complaint the eye is hinting at — is exactly where pattern-matching against typical images degrades, and where an accountable person has to make the call under uncertainty.
4. Trust and continuity are human. Getting a worried patient to come back, disclose symptoms, and follow through on a referral runs on a relationship, not on the quality of a screening output.

(Calibration flag: axiom 2 is the one moving fastest. "The AI only reads one macula-centered photo" is a description of current deployed screening, not a fixed limit — wider-field imaging, multi-modal capture, and models trained across more of the eye are closing this gap. If the screening captures and reads more of the eye, part of what STILL HOLDS here migrates toward INVALID, and the physical-exam moat narrows to what genuinely needs hands and a slit-lamp.)

New axioms

1. Patients get a glasses Rx without anyone ever looking at the eye — and the exam that would have caught disease never happens. The eye exam's quiet public-health function was that catching diabetic retinopathy, glaucoma, tumors, and systemic disease came bundled with people showing up to renew their glasses. Unbundle the refraction and that incidental screening opportunity disappears for whoever now skips the visit. Nobody owns closing that gap.
2. Who is accountable when AI screening misses pathology? If the read was done by a model and the referral wasn't made, liability sits — where? With the vendor, the supervising optometrist who signed a batch of AI reads without seeing the patient, the platform, or no one clearly. Malpractice frameworks assume the party who read the image is the party who examined the patient; screening-at-scale breaks that, and "the licensed signer who rubber-stamps machine output" is a role the law hasn't defined.
3. The exam-bundled-with-glasses business model breaks. The clinical visit was cross-subsidized by dispensing margin. When refraction goes self-serve and glasses go direct-to-consumer, the disease-detection visit loses the retail revenue that funded it — right as disease detection becomes the only part that still needs the optometrist. The scarce, valuable activity is stranded from its historical funding source.
4. Signing off on machine output without independent examination is a distinct role nobody has scoped. If the optometrist's job becomes adjudicating AI flags and countersigning refractions, that is a real, defensible verification-and-accountability function — but it is being treated as a downgrade or an afterthought rather than defined, staffed, priced, and regulated as the actual work.

Where it breaks

"The exam is bundled with and paid for by dispensing glasses" (invalid) collides head-on with "someone must catch and own the disease in the eye" (still holds). The disease-catching function only ever got funded because it rode along with a glasses sale; unbundle the refraction and route it online, and the one part of the visit that genuinely still needs a clinician — the physical exam, the adjudication of flags, the accountable referral — loses the revenue that paid for it, at the exact moment it becomes the whole remaining value. The field is set up to lose its screening reach precisely by succeeding at automating refraction.

A second collision: "reading the fundus image requires a scarce human" (invalid) meets "who is accountable when the AI misses pathology" (new). Once the read is automated and abundant, the reasonable move is to screen far more people — but no one has decided whether the optometrist countersigning those AI reads is a gatekeeper who examined the patient or a signer who never saw them, and the liability lands differently depending on an answer nobody has written down.

ai-axioms.vercel.app/axioms/healthcare/optometry-ai-axioms