Deeper Learning,Human–Machine Interactions and Lessons Learned from the Early Adoption of Artificial Intelligence in Gastroenterology

Introduction

Artificial intelligence (AI) and deep learning technologies are poised to transform gastroenterology, one of the medical domains at the forefront of AI integration. In recent years, gastrointestinal (GI) endoscopy has seen some of the first real-time AI applications in clinical medicine, particularly in colonoscopy, where computer-aided detection (CADe) algorithms for colorectal polyp detection have obtained regulatory approvals. Early studies, including multiple randomized, controlled trials (RCTs), have shown that AI assistance can significantly improve endoscopic lesion detection performance. Beyond colonoscopy, AI systems are being explored in capsule endoscopy for faster image reading, in pathology for automated analysis of biopsy slides, and in clinical decision support to predict patient outcomes from electronic health data. These developments promise to enhance diagnostic accuracy, efficiency, and personalized care in gastroenterology.

However, early adoption of AI in GI unveils complex challenges and opportunities related to human–machine interaction. Gastroenterologists must learn to collaborate with AI tools to trust them as useful assistants while maintaining vigilance against errors. There are reports of both “automation bias,” where clinicians may over-rely on AI outputs, and “algorithmic aversion,” where early errors cause mistrust and rejection of AI recommendations. Understanding how physicians perceive AI systems (as authoritative colleagues or mere instruments) is crucial for designing workflows that foster successful design and implementation of these tools. Indeed, the theoretical insights from human–robot interaction research become relevant: As Kate Darling, a prominent researcher who studies the field of human–robot interactions, observed that humans are naturally prone to anthropomorphize technology and project agency onto machines that mimic human behavior. This tendency can influence how medical AI systems are accepted and utilized in practice. Concepts such as anthropomorphism (the attribution of human traits, emotions, or intentions to nonhuman entities), perceived moral agency of AI, and the broader sociotechnical context in which AI is deployed all shape the outcomes of human–AI collaboration in health care.

Early international experiences offer valuable lessons. Different health care systems have adopted AI with varying results and levels of enthusiasm. In Japan, national insurance began reimbursing AI-assisted endoscopy in 2024, reflecting institutional support for adoption. In contrast, the United States has proceeded cautiously, with the FDA requiring robust evidence for each device and reimbursement pathways still evolving. Across these settings, the interplay between technology and clinical practice helps determine whether AI improves outcomes. Implementation requires addressing ethical and regulatory issues (data privacy, liability, and bias) and ensuring that AI is introduced as a tool to support clinicians and patients rather than as a disruptive force.

This article provides a structured review of the early adoption of AI in gastroenterology, focusing on colonoscopy. We emphasize human–AI collaboration in each domain and discuss the lessons learned regarding acceptance, trust, and co-learning between clinicians and AI systems. We also integrate perspectives from Kate Darling’s work on human–robot interaction to analyze how anthropomorphic perceptions and questions of agency apply to medical AI. Through a synthesis of peer-reviewed studies, case reports, and thought-leadership insights, we aim to inform a path forward for the responsible and effective integration of “deeper learning” AI technologies in gastroenterology practices and beyond.

AI in Colonoscopy: Polyp Detection and Diagnosis

AI-Augmented Polyp Detection, or CADe: One of the earliest success stories of AI in GI is real-time polyp detection during colonoscopy. Multiple AI systems have been developed to alert endoscopists to possible polyps in the video feed, typically by highlighting regions of interest. In clinical trials, these CADe tools have shown improvements in adenoma detection rate (ADR). For example, an RCT in Italy reported an increase in ADR from 44% to 53% in both expert and nonexpert endoscopists. ¹ Since then, multiple RCTs have been performed, and a recent meta-analysis of 28 RCTs showed similar results. ² However, real, pragmatic studies showed mixed results from CADe, pointing to the possible influence of real-world human factors. ³ Notably, the effectiveness of AI is intertwined with the quality of the human–AI interaction. The endoscopist must still carefully visualize the mucosa and respond to AI alerts. Studies have documented that if clinicians dismiss AI prompts as false alarms, true lesions can be inadvertently ignored. Conversely, there is concern that overreliance on AI might cause an eventual degradation in technique. For instance, an endoscopist might unconsciously pay less attention to peripheral vision or spend less time on mucosal visualization if they assume the AI “will catch everything.” ⁴ Eye-tracking research supports this, showing altered gaze patterns when CADe is in use. ⁵ It is not known how these “adaptations” impact behavior in the long term, and subsequently, assuming that a study showing a change in outcome ignores that long-term use of AI will have an impact on the endoscopist’s abilities and behavior that today is not being studies with these shorter-term outcome studies. Studies suggest that AI works best as a collaborative tool. The endoscopist and AI effectively form a team, and optimal outcomes depend on the endoscopist’s ability to integrate the AI’s alerts into their workflow without reducing their own vigilance. For this collaboration to work, we need to understand the factors that impact the outcomes in both the short term and long term and how to manage factors in a way to enhance the collaboration and optimize outcomes in the long term. For example, cognitive biases, a systematic pattern of deviation from rationality in judgment wherein individuals create their own subjective reality from their perception of the input, may dictate behaviors and lead to counterproductive use of AI. ⁶ Many other factors such as complementarity, amount of information, and timing of information provided have been proposed as other factors that impact the way humans use and act on AI inputs and subsequently patient outcomes. ⁷ Complementarity refers to situations in which the performance of a human with AI assistance exceeds the performance of an unassisted human or the AI in isolation is exemplified in polyps characterization discussed below (Fig. 1).

OPEN IN VIEWER

Fig. 1.

Image depicting Human AI collaboration.

Polyp Characterization (computer-aided diagnosis [CADx]): Beyond detection, AI is being explored for real-time polyp characterization (CADx). The goal is for AI to predict histology (adenomatous vs. hyperplastic polyps) from endoscopic images, enabling an instant “optical biopsy.” Studies have shown that AI models can classify polyps with high accuracy, especially in left-sided diminutive polyps. ⁸ While this could support a “resect and discard” strategy, where diminutive left-sided polyps diagnosed by AI as non-neoplastic could be safely removed and discarded without pathological review. This may result in workflow and economic benefits. In practice, CADx has seen slower adoption than CADe. One reason is clinician caution: There is reluctance among endoscopists to rely on AI for a definitive diagnosis due to medico-legal liability concerns. ⁹ Mischaracterizing even a tiny polyp could have consequences, and endoscopists hesitate to change management without pathological confirmation, despite promising AI performance. Thus, early CADx studies note that while AI can achieve optical diagnosis with standards similar to experts, getting clinicians to trust and act on those predictions remains a challenge. Interestingly, there is evidence that having AI as a “second-reader” can improve physicians’ confidence in their own optical diagnoses. In addition, the diagnostic accuracy was the highest when there was concordance between the physician and AI. This highlights a theme of human–AI collaboration: That collaboration leads to higher accuracy, and thus together, humans and AI could reach thresholds that are acceptable for clinical decision making, ¹⁰ thereby demonstrating a real-world example of complementarity.

Finaly, human reliance on AI can augment our capabilities, yet deskilling can occur and withdrawal of AI support once deskilling has occurred can have negative consequences. For example, in a recent study from Poland, the ability of endoscopist to detect polyps was diminished after their exposure to AI assistance during colonoscopy. What is even more interesting is that males seemed more affected than females raising the possibility that gender differences might play a role in how our brain and skills are impacted by AI. ¹¹

Discussion

The early adoption of AI in gastroenterology offers a wealth of insights into how advanced algorithms can support clinicians. A central theme in these experiences is that human–AI collaboration works best when the technology is implemented as a partner to, rather than replacement of, the clinician. The concept of “deeper learning” in our title alludes not only to deep learning algorithms but also to the deeper understanding that humans and machines gain from working together over time. In this discussion, we synthesize key lessons learned, regarding concepts such as acceptance, trust, and the evolving human–AI relationship. We also present thoughts on why data in RCTs might be different than that from real-world deployments. Finally, we highlight the influence of anthropomorphism and perceptions of AI’s agency on the sociotechnical integration, drawing on Kate Darling’s theoretical insights.

Human–AI interaction: Trust and acceptance: One of the most striking findings across studies is how much the effectiveness of AI depends on human factors. A highly accurate polyp detector is of little value if an endoscopist ignores its alerts, whereas even a less-than-perfect AI can be useful if integrated thoughtfully with physician oversight. Early adopters have encountered both extremes. Some clinicians exhibit automation bias, placing unwarranted trust in AI outputs. For example, accepting an AI’s classification of a polyp without question, or assuming a negative AI screen means “all clear.” This can be dangerous if the AI is wrong, as overreliance may erode the clinician’s own vigilance. At the other end, algorithmic aversion has been observed: A physician experiences one or two false alarms or mistakes from the AI and consequently becomes overly critical, quickly discounting the AI’s input altogether. Even more extreme are physicians who refuse to “turn on AI” machines as “they do not need them” or they find them “annoying.” Such mistrust can nullify any benefit of the tool, since the user stops heeding its advice even when it is correct or categorically refuses to use it.

The balance between these is delicate. Training and experience are crucial to calibrate trust. While it might be tempting to turn on a machine that highlights polyps and say no training is needed, the long-term outcome we are hoping AI will deliver will likely require training, understanding AI–human collaboration, and the change in human perception and behavior over time. This echoes what Kate Darling describes in human–robot interactions: We form mental models of a machine’s competence and intent, sometimes even emotional bonds, which then guide our behavior towards it. If the AI is framed as a helpful colleague, a “second pair of eyes,” physicians may be more likely to value its input and collaborate, whereas if it is seen as an untrustworthy black box or a threat to their authority or even seen as a potential replacement, resistance will ensue. ¹²

A strategy to build appropriate trust is to ensure transparency and feedback. AI systems that are transparent and understandable will likely gain more clinician confidence than opaque ones. Additionally, incorporating user feedback loops, like allowing doctors to correct AI errors, thereby improving future performance, helps foster a sense of teamwork. This can create a virtuous cycle of co-learning: The AI updates with real-world data, and the human learns the AI’s strengths and weaknesses. In practice, most GI AI devices are not self-learning on the fly due to regulatory constraints, but periodic updates based on aggregated data are possible. Furthermore, as clinicians use these tools, their own skills may sharpen; for example, endoscopists might become more adept at noticing subtle lesions after being alerted by AI, effectively training their eyes to AI’s level of attentiveness. Thus, rather than de-skilling physicians or trainees, a common fear, AI can also upskill them in some respects. The determinants of this outcome require research to help guide future AI developments and deployments.

Anthropomorphism and Perceived Agency: Kate Darling’s work highlights how humans tend to anthropomorphize technology, attributing human-like qualities or agency to machines, especially if they exhibit social cues or perform tasks traditionally done by humans. In gastroenterology, most AI tools are not embodied robots but rather software on screens; yet even here, anthropomorphic effects can emerge. For instance, an endoscopist might begin to personify the polyp detection AI as an “invisible assistant” watching their back. If the AI has a name or a voice prompt (some systems beep or say “polyp detected”), it could further anthropomorphize the experience. Anthropomorphic framing can cut both ways. On one hand, it might increase trust, the physician feels as if a colleague pointed something out. There is anecdotal evidence of endoscopists playfully referring to AI suggestions as coming from a trainee or a “second observer,” which implicitly gives the AI a quasi-human role in the room. On the other hand, treating the AI as an agent could diffuse the physician’s sense of responsibility, a dangerous mentality if it leads to complacency. Notably, machines have no moral agency in the clinical context; the accountability for decisions and patient outcomes remains squarely with human professionals and the institutions deploying the AI. However, if clinicians subconsciously assign some moral agency to the AI such as blaming the AI for a miss it complicates the ethical landscape.

Darling points out that emotional attachment or trust toward machines is “not inherently dangerous, but it is powerful.” Context is key. In a positive context, viewing the AI as a partner could encourage clinicians to embrace it in their workflow, leading to better outcomes. In a negative context, anthropomorphism could lead physicians to delegate liability and complacency. Thus, designers of medical AI should be mindful of how interface and branding create impressions

The major lesson from early AI projects in gastroenterology is that technical success alone does not guarantee clinical impact the context of deployment matters immensely. The early lessons in GI mirror those in other fields: AI works best as augmentation, not automation, at the current state of technology. As AI algorithms inevitably become more powerful and perhaps inch towards greater autonomy, the partnership model will need continuous re-evaluation.

In summary, the early chapters of AI in gastroenterology teach us that achieving the potential of AI requires as much attention to human and systemic factors as to algorithmic performance. When gastroenterologists and AI systems are complementary, the result is a powerful synergy. But this relationship must be cultivated through training, experience, and ethical guardrails. The lessons learned so far are guiding us to treat AI not only as a mysterious black box or a human replacement but also as a sophisticated tool that thrives in partnership with skilled clinicians. By internalizing these lessons, the GI field can continue to innovate with AI in a way that genuinely improves patient care and outcomes.

Conclusion

The early adoption of artificial intelligence in gastroenterology has been a journey of both promising achievements and instructive challenges. On the technical front, AI has demonstrated clear benefits. We have learned that successful integration of AI is not merely a software upgrade but a transformation in workflow and mindset. Gastroenterologists and their teams must adapt to working alongside intelligent machines, developing a calibrated trust, neither blindly deferential nor unduly dismissive. Early experiences underscore the need for training clinicians in the use of AI tools, as well as training the tools to better serve the clinicians’ needs.

A recurring insight is that AI in gastroenterology works best as an augmentative ally. In colonoscopy, AI serves as a vigilant observer, but the physician remains the pilot in command. It aligns with the notion that while AI can replicate or even exceed human pattern recognition in narrow tasks, it lacks the broader contextual understanding, intuition, and ethical reasoning that human clinicians contribute. Moreover, as Kate Darling’s work reminds us, technologies reflect the values of the systems in which they operate. The lesson for us is to intentionally shape the adoption of GI AI in ways that reinforce empathy, patient-centered care, and physician empowerment, rather than undermining them. For example, if AI can take over the mundane aspects of a procedure, the physician might have more time to communicate with the patient—a positive reallocation of effort. Conversely, we must guard against scenarios where overreliance on AI diminishes practitioner skills or patients feel alienated by automation.

In conclusion, the story of AI’s early foray into gastroenterology is one of augmented intelligence, combining the strengths of human and machine. We have seen that AI can indeed help us “learn deeper,” not only through deep learning algorithms but also through the deeper insights gained about our own practices when we collaborate with these technologies. Gastroenterology stands as a model for other fields in how to thoughtfully integrate AI: Start with evidence-backed applications, keep the clinician in the loop, observe the human–machine dynamics, and refine the approach continually. The interdisciplinary nature of this effort, drawing from computer science, clinical expertise, psychology, and even fields like human–robot interaction, will be integral for a well-rounded successful development. As these systems mature, patients should begin to experience tangible improvements: More accurate detections, more personalized treatments, and possibly safer, more efficient care. The lessons learned from early adoption remind us that achieving these benefits requires more than just clever algorithms; it demands a commitment to ethical deployment, ongoing education, and treating AI as a partner in our pursuit of better digestive health for all. With these lessons in mind, the gastroenterology community can be well prepared to navigate the next stages of the AI revolution, ensuring that technology truly serves to enhance (and never diminish) the art and science of healing. It can also serve as a model for other disciplines on the benefits and pitfalls of AI adoption and how to plan their implementation.

Authors’ Contributions

H.F. led article drafting, literature review, and primary writing. T.K. conceived the study, provided senior oversight, and critically revised the article.

Author Statement

Each author contributed substantially to the work, participated in drafting or revising the article, approved the final version, and agrees to be accountable for all aspects of the work.

Ethics Statement

This work does not involve human or animal subjects and does not require IRB approval.