Building Trustworthy Artificial Intelligence for Gastrointestinal Endoscopy

Artificial intelligence (AI) in gastrointestinal (GI) endoscopy is entering a new phase. Over the past decade, the field has focused predominantly on demonstrating technical performance, with computer‐aided detection (CADe) systems showing consistent improvements in adenoma detection and computer‐aided diagnosis (CADx) systems increasingly approaching expert‐level performance. Yet as AI moves from research settings into routine clinical practice, a fundamental question emerges: how can these technologies be implemented responsibly, safely and equitably?

In this issue of Digestive Endoscopy, El‐Sayed and colleagues provide a timely review of the ethical and legal implications of AI implementation in GI endoscopy [1]. Their work complements the recently published World Endoscopy Organization (WEO) international consensus statement [2], which represents the first global effort to establish expert consensus on the medicolegal and ethical foundations required for AI deployment in endoscopic practice. Together, these articles signal an important evolution in the field—from a focus on whether AI works to a broader consideration of how AI should work within healthcare systems.

A recurring theme across both papers is the central role of data governance. Modern AI systems depend on access to vast quantities of endoscopic images, videos, pathology data and clinical metadata. However, ownership, consent, privacy protection and secondary use of these data remain unresolved in many jurisdictions. The WEO consensus appropriately emphasizes robust data protection, transparency regarding commercial partnerships and clear institutional policies regarding data ownership and use. Such recommendations reflect a growing recognition that public trust is not simply desirable but essential for sustainable AI innovation.

The second major challenge concerns accountability. Current endoscopic AI systems remain largely assistive, with clinicians retaining ultimate responsibility for patient care. Nevertheless, emerging applications—including CADx‐guided optical diagnosis, automated reporting, AI‐derived quality indicators and future autonomous systems—blur traditional boundaries of responsibility. If a lesion is missed despite an AI alert, or conversely because an AI system failed to generate one, who is accountable? The clinician, institution, manufacturer, software developer or even regulatory body? Professional societies, regulators and legal experts must collaborate to establish practical standards before these uncertainties become barriers to adoption.

Equally important is the issue of bias and equity. The promise of AI is often presented as universal, yet AI systems are only as representative as the data used to develop them. Differences in disease prevalence, imaging equipment, endoscopic technique and patient demographics may all influence performance. Although some endoscopic tasks may appear biologically invariant, evidence from broader healthcare AI applications suggests that hidden biases can emerge unexpectedly when systems are deployed across new populations. As the authors mention, transparent reporting of dataset characteristics and subgroup performance should therefore be viewed as a minimum requirement rather than an aspirational goal.

Notably, the authors extend the discussion beyond technical and regulatory issues to consider human–AI interaction. Automation bias, algorithm aversion and potential deskilling represent underexplored threats to implementation success. The effectiveness of AI will depend not only on algorithmic performance but also on how clinicians interpret, trust and respond to AI‐generated outputs. Future implementation studies should therefore evaluate human factors alongside traditional diagnostic endpoints. Without such real‐world human interaction data, we could not deliver trustworthy AI tools.

Perhaps the most important contribution of the article is the recognition that the next wave of AI in endoscopy may differ fundamentally from what came before. Generative AI and foundation models are not limited to single tasks such as polyp detection. They may generate reports, recommend surveillance intervals, summarize procedures and potentially support broader clinical decision‐making. These capabilities challenge existing regulatory frameworks that were designed around narrowly defined intended uses. The ethical and legal questions raised by generative AI are therefore not merely extensions of previous debates but represent a qualitatively new challenge for the field. The regulatory bodies are working on developing a new legal framework for such multi‐task general AI models as medical devices. However, it will take considerable time given the uniqueness of such models in comparison with the traditional AI devices which have in principle only one intended use per each [3].

The history of endoscopy has been marked by technological innovation, from high‐definition imaging to advanced therapeutic interventions. AI may ultimately prove equally transformative. However, innovation alone is insufficient since this AI evolution brings new legal/ethical challenges that we have not seen in traditional medical device innovations. The future success of AI in GI endoscopy will depend greatly on establishing systems that are trustworthy, transparent, accountable and equitable.

The WEO consensus statement and the review by El‐Sayed and colleagues provide an important foundation for that journey. The next challenge is translating these principles into practical frameworks that ensure AI serves not only technological progress but also the interests of patients worldwide.

Funding

The author has nothing to report.

Conflicts of Interest

Y.M.—Olympus (consultation, device loan, lecture fees); Cybernet System (loyalty fees).

Linked Articles

This article is linked with El‐Sayed et al. To view this article, visit https://doi.org/10.1111/den.70198.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.