T he conceptual origins of artificial intelligence (AI) can be traced to the mid-20th century, when Alan Turing first questioned whether machines could emulate human reasoning in his seminal 1950 paper Computing Machinery and Intelligence . This foundational idea was formalized shortly thereafter at the Dartmouth Conference, where John McCarthy and associates introduced the term “artificial intelligence” and proposed that aspects of cognition could be encoded computationally. Although early approaches relied on rule-based systems, meaningful progress remained limited until advances in computational power and data availability enabled the development of modern machine learning and deep learning.
Within ophthalmology, these advances have translated rapidly into clinically relevant applications, largely due to the field’s reliance on high-resolution imaging and quantitative analysis. Early landmark studies demonstrated that deep learning systems could achieve diagnostic performance comparable to that of specialists across multiple retinal diseases. In parallel, additional works extended these approaches to optical coherence tomography, emphasizing not only disease detection but also clinically actionable decision-making. More recently, Lee and associates have highlighted the broader implications of AI in ophthalmology, particularly its role in large-scale data integration, telemedicine, and workflow optimization.
The broader impact of AI in medicine has been underscored by high-profile studies across multiple specialties in which AI augments rather than replaces human expertise, improving diagnostic accuracy and efficiency. Similarly, a recent systematic review showed that deep learning systems can perform at levels comparable to health care professionals in image-based diagnosis. Against this backdrop, ophthalmology has emerged as a potentially leading domain for clinical AI implementation. Yet, as these technologies move from research settings into routine care, their adoption has been accompanied by a series of concerns that extend beyond performance metrics alone. Importantly, the discussion should not be confined to image-based algorithms, as the emergence of large language models and generative AI introduces many of the same concerns, along with additional challenges related to hallucination (meaning the generation of inaccurate or unsupported information), privacy, reliability, and oversight that are highly relevant to clinical care.
Among these, bias remains one of the most consequential. AI systems are often perceived as inherently objective, yet their outputs are fundamentally shaped by the data on which they are trained. Imbalances in training data sets may lead to systematic disparities in performance, particularly across different demographic or clinical subgroups. The work of Ting and associates sought to address this through multiethnic validation, yet subsequent analyses have demonstrated that residual variability persists. , The concern is not simply that bias exists, but that it may be obscured by the apparent precision of algorithmic outputs, creating an illusion of objectivity that can mask underlying inequities. Bias in clinical AI is also more complex than demographic imbalance in training data sets alone because even models that have undergone multipopulation validation and achieved regulatory approval may still demonstrate variable performance when deployed locally, owing to differences in patient populations, care pathways, workflow, and data acquisition across health systems. An additional strategy to mitigate bias may be the development of AI systems that allow, but do not require, clinicians to provide one or more clinical diagnostic possibilities as contextual input. Importantly, such systems should also permit clinicians to update these possibilities after an initial AI output is generated, enabling an iterative process that more closely mirrors real-world clinical decision-making.
Closely related is the issue of interpretability. Many high-performing AI systems function as “black boxes,” producing accurate predictions without transparent reasoning. In their landmark study, De Fauw and associates demonstrated the feasibility of deep learning systems capable of both diagnosing retinal disease and recommending referral decisions from optical coherence tomography data. However, for AI systems intended for clinical medicine, it is not sufficient for outputs to consist only of diagnoses or referral recommendations. Such systems should also provide clinically meaningful explanations of how relevant inputs contributed to each recommendation, allowing clinicians to understand, verify, and appropriately challenge the output. Without this level of explainability, even highly accurate systems may encounter resistance and may be difficult to integrate responsibly into clinical care.
The reliance of AI on large-scale data sets introduces further complexity in the form of data privacy and ownership. The aggregation of imaging and clinical data across institutions is essential for developing robust models, yet it raises important questions regarding consent, governance, and data stewardship. Patients are often unaware of how their data are used beyond direct clinical care, and the increasing involvement of commercial entities adds another dimension to these concerns. Balancing the need for data accessibility with the imperative to protect patient privacy remains a central challenge in the responsible development of AI, whereas transparent governance and appropriate safeguards are essential for its eventual clinical deployment.
A related limitation is the issue of generalizability. AI systems frequently demonstrate strong performance in controlled environments but may encounter reduced accuracy in real-world settings. Variability in imaging devices, acquisition protocols, and patient populations can all influence model performance. While studies such as those by Ting and associates have emphasized the importance of external validation, prospective real-world evaluations remain limited. The potential for “silent failure,” in which performance degrades without obvious indication, underscores the need for continuous monitoring and validation following deployment.
Beyond these technical considerations lies a more nuanced concern regarding the evolution of clinical judgment. As AI systems assume an increasing role in diagnostic interpretation, there is apprehension that reliance on automated outputs may diminish independent clinical reasoning. However, this perspective may overlook the potential for AI to enhance, rather than erode, clinical expertise. As described by Topol, AI offers the opportunity to reduce cognitive burden and allow clinicians to focus on higher-level decision-making. In this context, the role of the ophthalmologist may shift from primary interpreter to integrator of complex data, including algorithmic insights. Ensuring that clinicians retain the ability to critically evaluate these outputs will be essential to preserving the integrity of clinical practice. Although AI has the potential to democratize specialized ophthalmic expertise in low- and middle-income settings, its dependence on costly infrastructure, technical expertise, and computational resources may just as easily deepen existing disparities in access to high-quality care.
Finally, the regulatory and medicolegal landscape surrounding AI remains incompletely defined. Unlike traditional medical devices, AI systems may evolve over time, challenging existing regulatory frameworks. Questions of accountability are similarly unresolved. When an AI system contributes to a clinical decision, it is not always clear where responsibility lies in the event of error. Current guidelines emphasize the need for standardization and ongoing validation, yet the rapid pace of technological advancement continues to outstrip the development of regulatory guidelines. Equally important are patient acceptance and the financial realities of implementation, as adoption will depend not only on stronger evidence regarding how patients perceive AI in their care but also on whether reimbursements, cost savings, or workflow gains create a sustainable model for routine use.
These concerns suggest that the apprehension surrounding AI in ophthalmology is not rooted in the technology itself but in the broader implications of its integration into clinical practice. Ophthalmology has traditionally been at the forefront of innovation, and AI represents a natural extension of this trajectory. The question, therefore, is not whether these technologies should be adopted, but how they can be implemented responsibly. The challenge lies in ensuring that integration remains aligned with the fundamental values of medicine. An alternative perspective is that some of the concerns surrounding AI may themselves be overstated, reflecting a degree of cognitive and professional bias rather than intrinsic limitations of the technology. Clinicians have historically been cautious in adopting disruptive innovations, and skepticism toward AI may, in part, mirror earlier resistance to technologies such as optical coherence tomography or automated perimetry. In this context, the emphasis on bias, opacity, and generalizability may disproportionately highlight the limitations of AI relative to those of human decision-making. Human clinicians are themselves subject to variability, fatigue, and cognitive bias, factors that are well documented in the medical literature yet often underacknowledged in comparisons with algorithmic performance. Indeed, studies such as those by Ting and Liu and associates have demonstrated that AI systems can achieve performance comparable to, and in some cases exceeding, that of human experts in image-based diagnosis. , Finally, ethical and trustworthy AI must include far more than transparency and explainability alone, requiring attention to governance, fairness, robustness, accountability, and the explicit quantification of epistemic and aleatoric uncertainty so that clinicians can better judge when algorithmic outputs are reliable and when caution is warranted.
In this sense, what we are really afraid of is not AI, but the possibility of losing clarity in how decisions are made and responsibility is assigned. Addressing these concerns will require not only technological solutions but also thoughtful engagement from the ophthalmic community. With careful stewardship, AI can evolve from a source of uncertainty into a trusted component of clinical care.
CRediT authorship contribution statement
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
