Artificial intelligence and robotics in Ophthalmology – Understanding the spectrum, respecting the limits

Abstract

Artificial intelligence (AI) in Ophthalmology has moved decisively from theory to practice. From patient education and screening algorithms to workflow automation and early forays into robotic surgery, AI is now embedded, quietly but pervasively, into contemporary eye care. The critical question is no longer whether AI will shape Ophthalmology, but where along the spectrum of autonomy it is safe, justified, and beneficial to deploy it, and at what cost. At the lowest end of this spectrum lie tightly scoped, patient-facing AI tools designed to educate, reassure, and support patients between clinic visits. Examples include WhatsApp-based health coaches and symptom-guidance systems that operate within strict guardrails, avoid clinical decision-making, and redirect users to human care when uncertainty arises. These systems represent a controlled, non-damaging application of AI, where the risk of harm is minimal and the benefit, improved access to information and reassurance, is tangible. Accountability remains clearly human. Such tools should be viewed not as endpoints, but as entry points. They demonstrate what is possible when AI is deliberately constrained. Importantly, they also establish a reference baseline: AI that informs without deciding, supports without directing, and reassures without replacing clinicians. As we move higher along the spectrum, patient-facing AI becomes more sophisticated. Systems begin to generate summaries, risk estimates, progression forecasts, and triage suggestions.[1,2] While still not autonomous, these outputs increasingly influence patient perception, expectations, and anxiety. A patient arriving with an AI-generated “printout” of risk scores subtly alters the clinical interaction, even if the physician retains final authority.[1] At this level, the requirement for human-in-the-loop oversight increases, particularly as disease severity and irreversibility of harm rise. Crucially, AI in Ophthalmology is not yet autonomous. This distinction matters. The higher the disease severity and the greater the potential for irreversible harm, the stronger the ethical, clinical, and legal requirement for human judgment. Stable chronic disease may tolerate more AI mediation; acute, sight-threatening pathology does not. Any responsible framework for AI deployment must scale human oversight in proportion to clinical risk. The discussion becomes even more complex as AI converges with robotics, particularly in cataract and anterior segment surgery. Robotic-assisted systems promise improved consistency, precision, and ergonomics, and many believe robotics in Ophthalmology is inevitable, just as it has been in other surgical disciplines. That may well be true. However, inevitability should not be confused with readiness. Several unresolved concerns demand attention. First, safety and efficacy. To date, there is insufficient high-quality randomized controlled trial (RCT) evidence demonstrating the superiority of AI-driven or robotic systems over experienced human surgeons. Non-inferiority is not enough when the baseline standard of care already delivers excellent outcomes. Superiority trials, large, multicentric, and independently conducted, are essential, particularly if these technologies are to be adopted widely. Second, cost-effectiveness and cost-utility. Technologies that offer marginal gains at disproportionate cost struggle to justify reimbursement. The experience with femtosecond laser-assisted cataract surgery (FLACS), which failed to achieve broad insurance coverage in many regions due to unfavourable cost-benefit profiles, is instructive. Robotic cataract surgery will face similar scrutiny, including quality-adjusted life year (QALY) analyses. Third, inequity. Advanced technologies rarely distribute evenly. High-income regions adopt first; resource-limited settings lag. If AI or robotics demonstrably improves outcomes, global disparities in visual health may widen rather than narrow. Fourth, sustainability. Large-scale AI systems and robotic platforms demand substantial computational power, energy, and infrastructure. In an era increasingly conscious of environmental cost, this cannot be ignored. Fifth, accountability. When systems fail, responsibility defaults to the clinician. A surgeon may abandon a robotic console and convert to manual surgery mid-procedure, yet outcomes may still be poor. Ethical and legal responsibility in such hybrid human–machine failures remains poorly defined. Sixth, regulation. Regulatory frameworks vary widely across countries, and oversight often lags behind innovation. The recent step-back by the U.S. Food and Drug Administration from prescriptive pre-market oversight of certain AI-enabled medical software may accelerate deployment,[3] but acceleration is not validation. Incremental software and hardware updates, introduced as minor modifications, risk bypassing rigorous clinical evaluation altogether. A sobering counterbalance to prevailing enthusiasm is provided by the Assessing the Real Impact of AI in Healthcare (ARISE) report, a joint clinician-led initiative by Stanford and Harvard Medical School.[4] The report highlights that while clinical AI is already widely deployed, reliable real-world impact remains uneven, performance is brittle in uncertainty-heavy scenarios, and human–AI workflow design, not model capability, remains the dominant determinant of safety. In short, capability has outpaced validation. This tension reflects what has been described as the AI paradox: systems that demonstrate impressive technical performance may fail to translate into commensurate real-world clinical benefit. Improved detection rates, higher accuracy metrics, or earlier identification of disease do not automatically lead to better outcomes, lower costs, or reduced inequity. Without careful integration into clinical pathways, AI risks amplifying diagnostic activity without improving care delivery, an experience not unfamiliar to ophthalmology. Compounding this challenge is the emergence of AI-integrated medicine preceding evidence-based medicine (EBM). Unlike traditional interventions that progress from controlled trials to adoption, AI tools are often embedded into workflows first, refined iteratively, and evaluated retrospectively. This inversion of the evidence pipeline creates discomfort among clinicians trained in EBM, yet reflects the practical realities of rapidly evolving digital technologies. Taken together, these developments underscore a central reality: AI and robotics in ophthalmology exist along a continuum, not as discrete categories. The outer boundaries of this continuum, how much authority machines should hold, how uncertainty should be communicated, and where accountability should reside, are still being actively tested. In an environment of regulatory relaxation and rapid deployment, physicians cannot remain passive end-users. As primary stakeholders, clinicians must participate in defining and refining these boundaries before they are implicitly set by technology alone. None of this argues against AI or robotics. Rather, it argues for measured progress. Robotics in ophthalmology may well become routine, but not yet, and not without robust, independently generated evidence across diverse settings. The real question is not whether we need AI and robotics, but where we need them, why we need them, and at what cost. Early signals are promising, but skepticism is not obstructionism; it is stewardship. As clinicians, we must resist both reflexive rejection and uncritical adoption. AI will increasingly speak to patients before clinicians do. Regulatory retreat will accelerate deployment. Patients will arrive armed with machine-generated numbers. The urgent question is not whether this will happen, but whether we are equipped to interpret it responsibly. If the next decade of Ophthalmology is to be augmented rather than overwhelmed by AI, clinicians must develop a new literacy, one that blends medical knowledge with statistical reasoning, calibrated skepticism, and ethical judgment. Otherwise, we risk mistaking numerical precision for clinical wisdom and confidence scores for care. Financial support and sponsorship Nil. Conflicts of interest There are no conflicts of interest to declare.

Ähnliche Arbeiten