“The future is already here – it’s just not very evenly distributed.”
William Gibson’s oft-quoted but still prescient words apply perfectly to today’s era of artificial intelligence in ophthalmology. Large language models (LLMs) and foundation models are no longer futuristic concepts; they are rapidly transforming how clinicians, researchers, and patients engage with eye health (1,2). Yet, as with any disruptive technology, their benefits are unevenly available, raising questions of access, trust, and responsible deployment.
Why foundation models matter in ophthalmology
Ophthalmology has always been at the forefront of medical imaging, with modalities like fundus photography, optical coherence tomography (OCT), and angiography producing vast datasets that are ideal for AI development (1,3). However, until recently most algorithms were narrow in scope – trained only for a single disease, task, or modality. But the advent of foundation models, large neural networks pre-trained on massive and diverse datasets, has changed this paradigm (4,9).
These models, such as RETFound, RetiZero, EyeFM, and multimodal agents like ChatGPT-5 for ophthalmology, are capable of generalizing across tasks and domains (3, 4, 8). Instead of building one algorithm for diabetic retinopathy screening and another for age-related macular degeneration, foundation models can be fine-tuned to perform multiple tasks – from disease detection to prognostic prediction, and even natural language report generation (9, 10).
The impact is profound: a single model can potentially support the entire spectrum of ophthalmic care, from triage in under-resourced clinics to advanced subspecialty decision-making in academic hospitals (4, 8).
The rise of LLMs in eye care
While image-based models dominate ophthalmic AI, LLMs are quietly revolutionizing the textual and conversational side of care (6, 7). In clinical workflows, residents and technicians draft thousands of imaging reports daily (Figure 1), which ophthalmologists must then review. Early pilots using ChatGPT-4o for report auditing have shown that LLMs can detect inconsistencies and errors with accuracy comparable to human experts, all while saving time and costs (6).
Beyond auditing, LLMs can serve as clinical copilots – explaining imaging findings in plain language for patients, generating structured trial eligibility reports, or assisting in guideline compliance (2, 7). Crucially, when coupled with multimodal foundation models that process images alongside text, LLMs unlock true vision–language reasoning (3, 10). For example, an AI agent could analyze an OCT scan, cross-reference it with the patient’s history, and generate a tailored recommendation for therapy escalation – all in the surgeon's natural language (Figure 2).
Barriers to adoption
As transformative as they are, LLMs and foundation models face several challenges before they can be equitably embedded in ophthalmology:
Data diversity: Current models are often trained on datasets from a few high-income countries. This risks bias when deployed in regions with different patient demographics, imaging devices, or disease prevalence (4, 9).
Trust and explainability: Clinicians remain cautious of “black-box” systems. A model that predicts progression of geographic atrophy without explaining which biomarkers it relied upon will struggle to gain true acceptance (5).
Regulatory hurdles: Unlike narrow AI tools, foundation models are general-purpose and harder to certify under existing medical device regulations (5).
Infrastructure gaps: Running multimodal LLMs requires cloud computing, robust internet, and secure data pipelines—resources many clinics in low- and middle-income countries lack (8).
Building the ecosystem for responsible AI
History has shown that AI tools succeed only when embedded in a supportive ecosystem. For LLMs and foundation models in ophthalmology, this means:
History has shown that AI tools succeed only when embedded in a supportive ecosystem. For LLMs and foundation models in ophthalmology, this means:
Multidisciplinary development: Ophthalmologists, data scientists, ethicists, and patient advocates must co-develop these systems to ensure they are clinically relevant, ethically aligned, and patient-centered (1, 2, 5).
Federated and privacy-preserving learning: To address data access and privacy challenges, federated pipelines allow models to learn from distributed datasets without moving sensitive patient information (3, 9).
Human-in-the-loop validation: LLMs should augment, not replace, ophthalmologists. Structured workflows that combine AI pre-screening with clinician oversight can maintain accountability and safety (6, 7).
Global collaboration: Initiatives such as the Global RETFound Consortium and EyeFM’s international partners are building networks to ensure models are trained and validated across diverse populations (4, 8).
The next frontier: AI equity in ophthalmology
The ultimate goal is not just technological advancement, but AI equity – ensuring that the benefits of LLMs and foundation models reach patients regardless of geography, infrastructure, or socioeconomic status (4, 8).
Imagine an under-resourced clinic in rural Africa with access to only a basic fundus camera and internet connection. A cloud-based foundation model like EyeFM could analyze images, generate a diagnostic report in the local language, and suggest referral pathways – all within minutes (8).
For patients, equity also means empowerment. AI-powered portals, driven by LLMs, could translate complex imaging results into understandable terms, track disease progression, and facilitate participation in clinical trials (6, 7).
For researchers and pharmaceutical companies, equitable deployment expands the data landscape, bringing underrepresented populations into global studies. This not only enhances fairness but also improves the robustness and generalizability of AI systems (3, 4, 10).
Conclusion
The convergence of LLMs and foundation models marks a pivotal moment for ophthalmology. We are moving from single-task algorithms to general-purpose AI agents capable of bridging images, language, and clinical reasoning (3, 4, 9, 10). But the promise of these technologies will only be realized if we address the pressing challenge of AI equity (4, 8).
The future of AI in ophthalmology is not just about model performance; it is about access, trust, and fairness. By investing in global collaborations, infrastructure, and inclusive datasets, we can ensure that models like RETFound, RetiZero, and EyeFM do not simply make care more efficient for privileged settings, but instead democratize eye health for all (1, 2, 4).
More on recent developments in this field can be found in our 2025 edition of the book “AI in Ophthalmology:” https://link.springer.com/book/10.1007/978-3-031-83756-2 (Chinese edition, 2024) and on International AI in Ophthalmology Society website (https://iaisoc.com/).
Q1. What is the main advantage of foundation models in ophthalmology compared to traditional AI algorithms?
A. They eliminate the need for ophthalmologists entirely
B. They can generalize across multiple tasks and modalities ✅
C. They are only useful for diabetic retinopathy
D. They require no data for training
Q2. How are LLMs being used in ophthalmic clinical workflows?
A. Designing new imaging devices
B. Auditing imaging reports and generating structured outputs ✅
C. Performing surgical procedures
D. Manufacturing intraocular lenses
Q3. What is the major concern with current ophthalmic AI datasets?
A. They contain no OCT scans
B. They are mostly from high-income countries, risking bias ✅
C. They are too small to be useful
D. They are entirely synthetic
Q4. What is identified as the next frontier for AI in ophthalmology?
A. Improving OCT resolution
B. Expanding cataract surgery access
C. Achieving AI equity and global fairness ✅
D. Reducing hospital staff training
Q5. Why is a supportive ecosystem necessary for deploying LLMs and foundation models?
A. To reduce internet costs
B. To ensure multidisciplinary collaboration and safety ✅
C. To replace ophthalmologists entirely
D. To avoid using clinical data
References
- K Jin, T Yu, A Grzybowski, “Multimodal artificial intelligence in ophthalmology: Applications, challenges, and future directions,” Surv Ophthalmol, S0039-6257(25)00120-1 (2025).
- K Jin, A Grzybowski, “Advancements in artificial intelligence for the diagnosis and management of anterior segment diseases,” Curr Opin Ophthalmol, 4:335 (2025).
- D Shi et al., “A multimodal visual–language foundation model for computational ophthalmology,” NPJ Digit Med, 8:381 (2025).
- YC Tham et al. (Global RETFound Consortium), “Building the world’s first truly global medical foundation model,” Nat Med, 31:1452 (2025).
- A Grzybowski, K Jin, H Wu, “Challenges of artificial intelligence in medicine and dermatology,” Clin Dermatol, 42:47 (2024).
- Z Su et al., “Assessment of large language models in cataract care information provision: A quantitative comparison,” Ophthalmol Ther, 13:1321 (2024).
- D Kang et al., “Evaluating the efficacy of large language models in guiding treatment decisions for pediatric myopia: An observational study,” Ophthalmol Ther, 14:705 (2025).
- Y Wu et al., “An eyecare foundation model for clinical assistance: a randomized controlled trial,” Nat Med, 31:1675 (2025).
- Y Zhou et al., “A foundation model for generalizable disease detection from retinal images,” Nature, 622:156 (2023).
- M Wang et al., “Enhancing diagnostic accuracy in rare and common fundus diseases with a knowledge-rich vision-language model,” Nat Commun, 16:5528 (2025).
