On Friday, October 2 at Academy 2026 in Anaheim, Todd Peabody, OD, MBA, and Chris Clark, OD, PhD, presented “AI in Eye Care: Current Realities and the Frontier.” Here are 5 takeaways from that presentation.

1. AI is well suited to image-based eye care.

Artificial intelligence (AI), particularly deep-learning systems, can analyze the high-resolution imaging data that is routinely generated in optometric practice. Fundus photographs, optical coherence tomography (OCT), OCT angiography, corneal topography, and meibography provide structured visual data that can support AI-assisted screening, detection, diagnosis, and longitudinal management.

Dr. Clark noted that when AI assessments conflict with clinical findings, clinicians' judgment, bedside evaluation, and professional accountability take absolute precedence over automated algorithms, decision-support tools, or AI recommendations. “The clinician should systematically check the AI’s input data for errors, evaluate what contextual or physical nuances the algorithm may have missed, and order targeted diagnostic testing or peer consults if a critical disparity remains,” he said. “Finally, the clinician must document the specific AI recommendation alongside the clinical rationale for overriding or adopting it, and report recurring false outputs to institutional patient safety or clinical informatics teams.”

2. Data quality determines clinical reliability.

AI systems learn patterns from their training data, including unintended patterns. One example is a pneumonia-detection model that identified differences among hospitals and departments rather than relying solely on disease-related findings. Similar confounding could occur in eye care if an algorithm learns associations related to imaging devices, clinical sites, patient demographics, or image-acquisition protocols.

Generative AI introduces an additional concern because it can produce plausible but incorrect information when its source material is inaccurate or incomplete. It is recommended to restrict the content that is available to the model, use systems designed for health care, and to independently check outputs.

3. Administrative applications may offer the most immediate benefits.

Practices do not have to purchase new diagnostic equipment to begin using AI. Current applications include patient intake, symptom triage, scheduling, telephone support, documentation, coding, staff training, patient education, and professional communications. Ambient scribes may reduce after-hours documentation, while automated intake systems can convert patient-entered information into structured histories and potential risks before the examination begins.

These systems should supplement established workflows. Clinicians and staff remain responsible for reviewing generated documentation and communications before they are entered into the record or sent to patients.

According to Dr. Peabody, ambient scribes generally yield the fastest return on investment for small practices by significantly reducing after-hours charting, lowering clinician burnout, and freeing up time to increase patient capacity without requiring expensive hardware. Closely following are automated intake systems, which generate quick savings by reducing front-desk data entry, streamlining scheduling, and highlighting critical risks before an exam begins. “Because both applications integrate into existing computers or mobile devices, small practices can capture immediate administrative efficiency gains without heavy capital expenditure, provided clinicians continue to review and verify all AI-generated documentation,” explained Dr. Peabody.

4. Clinical AI is moving from screening toward prediction and management.

Autonomous diabetic retinopathy screening is one example of AI’s ability to extend standardized disease detection into primary care and underserved settings. Emerging models may go even further by combining structural, functional, and vascular biomarkers to estimate glaucoma progression, classify disease risk, and help individualize follow-up intervals.

AI is also being used for myopia screening and axial length estimation, pediatric vision assessment, dry eye disease, orthokeratology, and simultaneous evaluation of multiple retinal conditions. These tools could provide more reproducible measurements and identify patterns that are difficult to recognize through observation alone.

Additionally, machine-learning models may help quantify meibomian gland morphology, automate tear break-up time measurements, classify dry eye subtypes, and generate reproducible severity scores. For instance, in orthokeratology, models could assist with predicting landing zone angles, selecting toric parameters and base curves, and estimating the risk of lens decentration or induced higher-order aberrations.

These applications may reduce measurement variability and support clinical decision-making, but their value depends on performance across different devices, patient populations, and practice environments.

The presenters stressed that clinician experience should always supersede AI influence because every patient is unique. “A predictive model does not need to achieve a single, universal accuracy percentage to influence clinical care; rather, its required performance depends on the risk level of the decision and its proven impact on patient outcomes,” says Dr. Clark.

For low-risk choices like adjusting dry eye follow-ups, moderate predictive power may suffice, but high-risk decisions—such as delaying glaucoma monitoring—demand exceptionally high sensitivity and negative predictive value to ensure progressive disease is not missed. “Before influencing treatment or scheduling, a model must demonstrate that its predictions yield equal or superior clinical outcomes compared to standard care, while maintaining consistent performance across diverse patient populations and different diagnostic devices,” addsDr. Clark.

5. Clinical oversight, transparency, privacy, and equity remain essential.

AI should function as decision support rather than a substitute for clinical judgment. Before adopting a diagnostic system, clinicians should ask several due diligence questions:

• Validation: Has the model been validated specifically in my clinical environment?

• Evidence: Are peer-reviewed trials and real-world evidence from similar settings available?

• Reference Standard: Does the system utilize a high-quality “ground truth” for its training?

• Transparency & Liability: Is the vendor transparent about lifecycle monitoring and liability assumption?

The presenters’ central analogy was AI is like a global positioning system: The technology can suggest a route, but the clinician remains responsible for the destination and the patient’s care. OM