Artificial intelligence-assisted colonoscopy: Transforming detection, characterization, and management of colorectal neoplasia

Abstract

Colorectal cancer remains one of the leading causes of cancer-related morbidity and mortality worldwide, and its prevention largely depends on the accurate detection and characterization of precursor lesions during colonoscopy. This narrative review provides an updated overview of the current role of artificial intelligence (AI) in colorectal neoplasia based on a structured literature search of the PubMed/MEDLINE database, focusing primarily on studies published between 2018 and 2025, including randomized controlled trials, prospective studies, systematic reviews, meta-analyses, and recent clinical practice guidelines. AI-assisted colonoscopy, particularly through computer-aided detection and computer-aided diagnosis (CADx) systems, has consistently demonstrated improvements in colonoscopic performance. Randomized controlled trials and meta-analyses have shown absolute increases in adenoma detection rate of approximately 10%-15%, accompanied by reductions in adenoma miss rates. In parallel, CADx systems have achieved diagnostic accuracies frequently exceeding 90% for the optical characterization of diminutive colorectal polyps, supporting real-time strategies such as resect and discard and, in selected cases, diagnose and leave. Emerging evidence also suggests potential applications in colorectal cancer risk stratification, therapeutic decision-making, and precision endoscopy. Despite these advances, important challenges remain, including false-positive detections, limited external validation, algorithm generalizability across diverse populations and endoscopic platforms, implementation costs, interoperability, and unresolved regulatory and medico-legal issues. Recent international guidelines recognize AI as a valuable adjunct that improves lesion detection and supports optical diagnosis, while emphasizing the need for additional evidence demonstrating long-term clinical benefits, including reductions in post-colonoscopy colorectal cancer incidence and mortality. Current evidence supports AI as an effective decision-support tool that complements rather than replaces the expertise and clinical judgment of the endoscopist. Continued multicenter validation, real-world implementation studies, and prospective evaluation of patient-centered outcomes will be essential to define its optimal role in routine colorectal cancer prevention and therapeutic endoscopy.

Key Words: Colorectal cancer; Colonoscopy; Adenoma; Artificial Intelligence; Neoplasm

Core Tip: Artificial intelligence (AI) is rapidly redefining the landscape of colorectal neoplasia, emerging as a pivotal driver of a new era in precision endoscopy. By augmenting human perception with advanced pattern recognition and real-time data processing, AI-enabled systems have demonstrated a consistent capacity to enhance adenoma detection, refine optical diagnosis, and support therapeutic decision-making with unprecedented accuracy and reproducibility. Beyond improving key quality indicators such as adenoma detection rate and diagnostic performance, AI mitigates operator-dependent variability and establishes a more standardized and objective approach to endoscopic practice. Its integration across the entire continuum of care from risk stratification and screening to lesion characterization and advanced endoscopic resection positions AI as a transformative tool capable of optimizing colorectal cancer prevention strategies. As evidence continues to accumulate, AI is no longer a complementary technology but a central component in the evolution toward more effective, individualized, and outcome-driven management of colorectal neoplasia.

INTRODUCTION

Colonoscopy remains the cornerstone of colorectal cancer (CRC) prevention because it enables the detection and removal of precursor lesions before malignant transformation. However, the effectiveness of colonoscopy is intrinsically dependent on examination quality and endoscopist performance[1-3]. Studies on screening colonoscopy have demonstrated that approximately 21% of adenomas may be missed during a single examination, with substantially higher miss rates for diminutive and non-polypoid lesions, both of which contribute to the occurrence of interval CRC despite adherence to established screening programs[4-7]. These limitations have driven the development of technologies aimed at improving lesion detection, characterization, and overall colonoscopy quality[8-12].

Recent advances in artificial intelligence (AI), particularly deep learning (DL)-based computer vision systems, have opened new opportunities to address several of these limitations. Modern convolutional neural network (CNN) algorithms are capable of analyzing endoscopic images and video streams in real time, supporting endoscopists during lesion detection through computer-aided detection (CADe) and assisting optical characterization through computer-aided diagnosis (CADx)[13-16]. Beyond improving procedural performance, AI has the potential to reduce interobserver variability, standardize optical diagnosis, optimize therapeutic decision-making, and facilitate quality assurance in colorectal endoscopy[17-21].

Nevertheless, despite the rapid expansion of AI-assisted colonoscopy and the growing number of randomized controlled trials, systematic reviews, and recent clinical practice guidelines, important questions remain regarding the generalizability of current algorithms, their performance across different clinical settings and endoscopic platforms, cost-effectiveness, long-term patient outcomes, and the practical challenges associated with routine implementation. As evidence continues to evolve, a balanced appraisal of both the demonstrated benefits and the current limitations of AI has become increasingly important[22-27].

To improve transparency, this narrative review was based on a structured literature search of the PubMed/MEDLINE database, focusing primarily on studies published between 2018 and 2025, while also incorporating landmark earlier publications considered fundamental to the development of the field. The search included randomized controlled trials, prospective studies, systematic reviews, meta-analyses, real-world studies, and recent international clinical practice guidelines relevant to AI-assisted colonoscopy.

Accordingly, this review provides a comprehensive and up-to-date synthesis of the current evidence regarding AI in colorectal neoplasia, with particular emphasis on CADe, CADx, emerging applications in risk stratification and therapeutic endoscopy, current barriers to clinical implementation, and future research priorities[28-37]. By integrating evidence from pivotal clinical studies, recent guidelines, and real-world experience, this review aims to provide a balanced perspective on the present and future role of AI in CRC prevention and endoscopic practice.

LITERATURE SEARCH STRATEGY

To provide a comprehensive and up-to-date overview of the role of AI in colorectal neoplasia, we performed a structured literature search in PubMed/MEDLINE focusing primarily on studies published between 2018 and 2025, while also including seminal earlier publications considered fundamental to the development of the field. The search incorporated combinations of keywords and Medical Subject Headings including AI, machine learning, deep learning, colonoscopy, computer-aided detection, computer-aided diagnosis, CADe, CADx, colorectal neoplasia, CRC, adenoma detection, and optical diagnosis. Priority was given to high-quality evidence, including randomized controlled trials, prospective studies, systematic reviews, meta-analyses, clinical practice guidelines, and landmark studies that have significantly influenced current clinical practice. References were selected according to their scientific relevance, methodological quality, and contribution to the topics addressed in this narrative review.

Given the narrative nature of this review, no formal risk-of-bias assessment or certainty-of-evidence grading was performed. Instead, priority was given to high-quality evidence, including randomized controlled trials, systematic reviews, meta-analyses, international clinical practice guidelines, and landmark studies that have substantially contributed to the current understanding of AI-assisted colonoscopy.

FUNDAMENTALS OF AI IN ENDOSCOPY

Machine learning and DL

AI in endoscopy primarily relies on machine learning (ML) and, more specifically, DL. ML algorithms learn patterns from data, while DL uses multilayered neural networks capable of automatically extracting complex features from images[8-10].

CNNs

The backbone of modern endoscopic AI systems is the CNN. CNNs process visual data through hierarchical layers[11-13]: Convolutional layers: Detect low-level features (edges, textures); Pooling layers: Reduce dimensionality; Fully connected layers: Integrate features for classification. These architectures enable AI systems to recognize subtle mucosal and vascular patterns often imperceptible to the human eye.

Training and validation

AI systems are trained using large datasets of annotated endoscopic images/videos[4,6-9]: Supervised learning with histology-confirmed labels; Internal validation (training/testing split); External validation across different populations (critical limitation). Performance is typically evaluated using[4]: Sensitivity/specificity; Area under the curve; Real-time detection latency.

AI–DRIVEN POLYP DETECTION IN COLONOSCOPY: CADE

AI-assisted colonoscopy has consistently demonstrated the ability to improve the detection of colorectal adenomas and polyps, addressing one of the principal limitations of conventional colonoscopy: Operator-dependent variability. Multiple randomized controlled trials have shown that CADe systems significantly increase adenoma detection rate (ADR) compared with conventional colonoscopy. In a landmark double-blind trial, Wang et al[14] reported an increase in ADR from 28.0% to 34.0% (P = 0.03). Similarly, Repici et al[15] demonstrated a significant improvement in ADR (54.8% vs 40.4%; P < 0.001) and adenomas per colonoscopy in a multicenter randomized trial, while Liu et al[16] observed higher ADR (39.1% vs 23.9%; P < 0.001) and polyp detection rate (PDR) (43.7% vs 27.8%; P < 0.001) with AI-assisted colonoscopy. These findings have been further supported by observational real-world studies demonstrating improved detection performance after implementation of CADe in routine clinical practice[17,18].

Beyond individual trials, contemporary systematic reviews and meta-analyses consistently demonstrate that CADe increases ADR by approximately 10%-15% points, improves PDR, and significantly reduces adenoma miss rates, particularly for diminutive (< 5 mm), flat, and sessile lesions[19-23]. Collectively, these findings indicate that AI can function as an effective second observer during colonoscopy, improving mucosal inspection and reducing operator-related variability without substantially prolonging examination time.

Importantly, recent international clinical practice guidelines have provided a more balanced interpretation of the available evidence. The American Gastroenterological Association (AGA) Living Clinical Practice Guideline, the BMJ Living Clinical Practice Guideline, and the European Society of Gastrointestinal Endoscopy (ESGE) Position Statement acknowledge that current evidence consistently supports the use of CADe to improve adenoma detection during colonoscopy. However, they also emphasize that the certainty of evidence regarding patient-important long-term outcomes including reductions in post-colonoscopy CRC (PCCRC) incidence and CRC mortality remains limited. Consequently, these guidelines recommend continued prospective evaluation, broader multicenter validation, and assessment of implementation across diverse healthcare settings before universal adoption can be unequivocally recommended[33-35]. Table 1 summarizes the principal randomized controlled trials evaluating AI-assisted detection of colorectal neoplasia during diagnostic colonoscopy.

Table 1 Comparative table of the effectiveness of artificial intelligence in the detection of colorectal neoplasms.

Ref.	Design	ADR (AI vs control)	P value	PDR (AI vs control)	P value
Wang et al[14]	Double-blind RCT	34% vs 28%	0.03	Not reported
Repici et al[15]	Multicenter RCT	54.8% vs 40.4%	< 0.001	Not reported
Liu et al[16]	Prospective RCT	39.1% vs 23.9%	< 0.001	43.7% vs 27.8%	< 0.001

ADR: Adenoma detection rate; AI: Artificial intelligence; PDR: Polyp detection rate; RCT: Randomized controlled trial.

Clinical rationale

Missed adenomas remain one of the principal causes of PCCRC. Human factors including fatigue, distraction, variable withdrawal technique, and differences in endoscopist experience contribute substantially to missed lesions and interobserver variability, providing the biological and clinical rationale for AI-assisted detection systems[4].

Evidence from randomized controlled trials

High-quality randomized controlled trials and contemporary meta-analyses consistently demonstrate that CADe systems: Increase ADR by approximately 10%-15% points; Improve PDR. Reduce adenoma miss rates, particularly for: Diminutive polyps (< 5 mm); Flat and sessile lesions. Improve detection consistency across endoscopists with different levels of experience[19-23]. Nevertheless, current evidence has not yet demonstrated that these improvements translate into reductions in CRC incidence, PCCRC, or CRC mortality, underscoring the need for long-term outcome studies[33-35].

Real-time performance

Modern CADe systems are capable of real-time image analysis with detection latencies typically below 50 ms, providing visual alerts such as bounding boxes or highlighting suspected lesions without significantly increasing withdrawal time[15-18]. These characteristics allow seamless integration into routine colonoscopy while maintaining procedural efficiency.

Limitations

Despite the consistent improvement in lesion detection, several important limitations should be recognized[13-18,33-35]: Increased detection of diminutive lesions, with potential implications for overdiagnosis and healthcare resource utilization. False-positive detections caused by mucosal folds, bubbles, stool debris, or light reflections. Variable performance across different AI platforms and endoscopic systems. Limited external validation across diverse patient populations and clinical environments. Uncertain impact on long-term patient-centered outcomes, including PCCRC incidence and CRC mortality. Current evidence supports AI as a decision-support tool that complements rather than replaces the expertise and clinical judgment of the endoscopist.

AI IN COLORECTAL LESION CHARACTERIZATION: CADX

AI-assisted optical diagnosis has emerged as one of the most clinically relevant advances in colorectal endoscopy because it enhances real-time lesion characterization and directly informs therapeutic decision-making during colonoscopy. By integrating surface pattern, vascular architecture, and microstructural features, CADx systems can support the distinction between neoplastic and non-neoplastic lesions with high diagnostic performance, thereby refining the choice between cold snare polypectomy, conventional endoscopic mucosal resection (EMR), endoscopic submucosal dissection (ESD), or referral for surgery when features suggest advanced pathology. In the prospective study by Byrne et al[23] a DL model differentiated diminutive adenomatous from hyperplastic polyps with 94% accuracy, 98% sensitivity, and a 97% negative predictive value against histology, supporting the feasibility of real-time optical diagnosis strategies such as “resect and discard” and, in selected rectosigmoid lesions, “diagnose and leave”. Similar findings were reported by Kudo et al[24], whose prospective clinical study showed that AI-based characterization of diminutive polyps achieved performance compatible with real-time clinical use and met thresholds relevant to pathology-sparing strategies. Importantly, when AI systems are compared with expert endoscopists, the differences are often not statistically significant, indicating that current CADx platforms tend to perform at least at an expert-equivalent level rather than consistently surpassing expert human assessment in standard optical diagnosis settings. At the same time, more advanced platforms based on endocytoscopy have shown even higher diagnostic performance: In the Endo-BRAIN study, Mori et al[25] reported 98.0% accuracy with stained endocytoscopic images and 96.0% accuracy with endocytoscopic narrow-band imaging, with excellent sensitivity and specificity when histopathology was used as the reference standard. These findings are reinforced by diagnostic meta-analytic data showing that CADx systems for diminutive colorectal polyps achieve high pooled sensitivity and specificity overall, supporting their potential role in improving the standardization of optical diagnosis, reduce interobserver variability, and strengthen real-time therapeutic decision-making in precision colonoscopy. Table 2 presents a comparison of the most representative studies on the application of AI for the characterization of colorectal neoplasms[26,27].

Table 2 Comparative table of the use of artificial intelligence in the characterization of colorectal neoplasms.

Ref.	Design	AI vs comparator	Diagnostic accuracy (AI vs human)	P value	Sensitivity (%)	Specificity (%)
Byrne et al[23]	Prospective real-time	AI vs expert endoscopists	94% vs 91%	0.16	98	83
Kudo et al[24]	Multicenter diagnostic	AI vs histology	98.0% (stained)/96.0% (NBI)	Not reported	96.9	94-100
Mori et al[25]	Prospective	AI vs expert endoscopists	92% vs 80%-85%	> 0.05	98	83

AI: Artificial intelligence; NBI: Narrow-band imaging.

Optical diagnosis and histology prediction

CADx systems aim to predict histology in real time using[23,24]: Narrow-band imaging; Blue light imaging; Optical enhancement. They classify lesions into: Adenomatous vs hyperplastic; Neoplastic vs non-neoplastic.

Evidence and diagnostic accuracy

High-quality studies demonstrate: High-quality prospective studies and contemporary meta-analyses consistently demonstrate that CADx systems achieve high diagnostic performance for the optical characterization of diminutive colorectal polyps, with overall diagnostic accuracies frequently exceeding 90%. When compared with expert endoscopists, AI-assisted optical diagnosis generally demonstrates comparable diagnostic performance, while reducing interobserver variability and improving diagnostic consistency. Meta-analytic evidence further supports high pooled sensitivity, specificity, and negative predictive value, particularly when CADx is applied to diminutive colorectal lesions using image-enhanced endoscopy[24-28].

Major findings include: Diagnostic accuracy frequently exceeding 90%. Diagnostic performance generally comparable to expert endoscopists. High sensitivity, specificity, and negative predictive value for the differentiation of neoplastic and non-neoplastic diminutive polyps. Improved diagnostic consistency and reduced interobserver variability.

Clinical implications

The principal clinical application of CADx is to support real-time optical diagnosis, allowing more individualized management of diminutive colorectal polyps while reducing unnecessary histopathological evaluation.

According to the American Society for Gastrointestinal Endoscopy Preservation and Incorporation of Valuable Endoscopic Innovations (ASGE PIVI) recommendations, implementation of a diagnose-and-leave strategy for diminutive rectosigmoid polyps requires an optical diagnosis technology capable of achieving a negative predictive value of at least 90% for adenomatous histology when used with high confidence. Likewise, implementation of a resect-and-discard strategy requires at least 90% agreement between surveillance interval recommendations based on optical diagnosis and those based on histopathology.

When these performance thresholds are achieved, CADx may facilitate: Diagnose-and-leave management of diminutive rectosigmoid hyperplastic polyps. Resect-and-discard management of diminutive adenomas. Reduction in pathology costs. Shorter time to post-procedure management decisions. Greater standardization of optical diagnosis during routine colonoscopy.

Recent international guidelines from the AGA, BMJ, and ESGE recognize the potential clinical value of CADx in supporting optical diagnosis. However, they also emphasize that broader prospective validation, consistent performance across different endoscopic platforms, and additional real-world implementation studies are still required before these strategies can be universally adopted[33-35].

Challenges

Despite the promising diagnostic performance of CADx, several important limitations remain: Variability in regulatory approval across different countries and healthcare systems. Not all commercially available CADx systems have consistently demonstrated performance meeting the ASGE PIVI thresholds under routine clinical conditions. Limited external validation across diverse patient populations, endoscopic platforms, and non-expert clinical settings. Performance variability related to image quality, bowel preparation, lesion morphology, and endoscopist experience. Additional prospective multicenter studies are needed to confirm long-term clinical effectiveness and facilitate widespread implementation[4,33-35].

AI FOR RISK STRATIFICATION AND SCREENING

AI is progressively extending its role beyond real-time endoscopic image analysis, emerging as a promising tool for CRC risk stratification and screening. Recent AI-based predictive models have integrated multidimensional data including demographic characteristics, clinical variables, family history, lifestyle factors, laboratory parameters, and, in selected studies, genetic information to estimate individual CRC risk and support the development of more personalized screening strategies[29-31]. These approaches have the potential to optimize screening intervals and improve resource allocation by identifying individuals who may benefit from earlier or more intensive screening.

AI has also demonstrated promising applications in non-invasive CRC screening modalities. ML algorithms have been investigated to enhance the diagnostic performance of fecal immunochemical test (FIT)-based screening by integrating quantitative FIT values with clinical risk factors, thereby improving risk prediction and prioritization for colonoscopy referral. Similarly, AI-assisted image analysis in computed tomography colonography has shown the potential to improve automated polyp detection, reduce perceptual errors, and assist radiologists in lesion identification, although its widespread implementation remains under evaluation[29-31].

Despite these encouraging developments, most AI applications for CRC risk stratification and non-invasive screening remain at the stage of clinical validation. Current evidence is heterogeneous, with considerable variability in study design, populations, predictive models, and outcome measures. Consequently, additional prospective multicenter studies and external validation across diverse healthcare settings are required before these approaches can be routinely incorporated into population-based CRC screening programs.

IMPACT OF AI ON THERAPEUTIC ENDOSCOPY

AI is increasingly expanding its role in therapeutic endoscopy by supporting clinical decision-making before and during endoscopic resection. The most mature application of AI in this setting is the real-time characterization of colorectal lesions, which assists endoscopists in selecting the most appropriate therapeutic approach, including cold snare polypectomy, EMR, ESD, or referral for surgical management when endoscopic features suggest deep submucosal invasion[19-28]. By integrating high-resolution image analysis with advanced pattern recognition, AI-assisted systems may improve the consistency of optical diagnosis and facilitate more standardized therapeutic decision-making.

Beyond lesion characterization, emerging studies have explored the potential of AI to predict clinically relevant pathological features, including the depth of submucosal invasion and the presence of submucosal fibrosis, which may influence procedural planning and lesion selection for advanced endoscopic resection. Additional investigational applications include automated lesion segmentation, delineation of lesion borders, and identification of residual neoplastic tissue during EMR and ESD. Although these developments are encouraging, most remain in the early stages of clinical validation, and evidence demonstrating improvements in complete (R0) resection rates, reductions in non-curative resections, or long-term oncological outcomes is currently limited.

Accordingly, recent international guidelines recognize AI as a promising adjunct for therapeutic decision support while emphasizing that further prospective multicenter studies are needed to validate these emerging applications before they can be routinely incorporated into advanced therapeutic endoscopy. At present, AI should be regarded as a decision-support technology that complements rather than replaces the expertise and procedural judgment of the therapeutic endoscopist.

LIMITATIONS AND CHALLENGES

Despite the remarkable advances achieved by AI in colorectal endoscopy, several important challenges continue to limit its widespread implementation in routine clinical practice. One of the principal concerns relates to the occurrence of false-positive detections. Although current CADe systems generally maintain high sensitivity, they may incorrectly identify normal mucosal folds, bubbles, residual stool, or light reflections as suspected lesions, potentially increasing unnecessary endoscopist attention and contributing to operator fatigue[30].

Another important limitation is the potential for algorithmic bias. Many currently available AI systems have been developed using datasets derived from selected patient populations, expert centers, or specific geographic regions. Consequently, their diagnostic performance may not be fully generalizable across different ethnic populations, healthcare settings, bowel preparation quality, lesion prevalence, or endoscopist experience[30,31].

Closely related to this issue is the limited availability of robust external validation. Although numerous randomized controlled trials have demonstrated promising results, relatively few AI systems have undergone large multicenter prospective validation across diverse clinical environments and different endoscopic manufacturers. Furthermore, considerable heterogeneity exists among currently available AI platforms regarding algorithm architecture, training datasets, performance metrics, and clinical endpoints, making direct comparisons between studies challenging[31].

Technical interoperability also remains an important consideration. Not all commercially available AI platforms are fully compatible with every endoscopic processor or imaging technology, and integration into existing endoscopy units may require additional hardware, software, or infrastructure, creating practical barriers for widespread implementation[30,31].

Economic considerations represent another major challenge. While AI-assisted colonoscopy has the potential to improve quality indicators and may ultimately prove cost-effective through improved CRC prevention, the initial costs associated with hardware acquisition, software licensing, maintenance, and personnel training may limit adoption, particularly in resource-constrained healthcare systems[31].

In addition, regulatory and legal aspects continue to evolve. Regulatory approval requirements differ across countries and jurisdictions, and many AI systems continue to undergo post-marketing evaluation as evidence accumulates. Questions regarding accountability, liability, data privacy, cybersecurity, and transparency of “black-box” algorithms remain incompletely resolved and represent important considerations before widespread clinical implementation[31,32].

Finally, it should be emphasized that AI is currently intended to complement not replace the expertise and clinical judgment of the endoscopist. Appropriate interpretation of AI-generated outputs within the clinical context remains essential to ensure safe and effective patient care[30,31].

An additional consideration is the potential for automation bias, whereby excessive reliance on AI-generated outputs could inadvertently reduce endoscopists’ vigilance or independent clinical judgment[31,32]. Therefore, AI should be regarded as a decision support tool rather than a substitute for expert endoscopic assessment.

FUTURE PERSPECTIVES

Future research is expected to focus on strengthening the clinical evidence supporting AI-assisted colonoscopy through large-scale multicenter prospective studies, broader external validation, and evaluation across diverse patient populations, healthcare systems, and endoscopic platforms. Particular emphasis should be placed on determining whether the improvements observed in surrogate quality indicators, such as ADR, ultimately translate into meaningful patient-centered outcomes, including reductions in PCCRC incidence, CRC mortality, and healthcare costs[33-37].

In parallel, continued technological development is expected to improve the integration of CADe and CADx into routine clinical workflows while enhancing interoperability among different endoscopic imaging platforms. Advances in explainable AI may further improve the transparency and interpretability of DL algorithms, potentially increasing clinician confidence and facilitating broader clinical acceptance.

Additional priorities include the standardization of AI performance metrics, harmonization of regulatory pathways, evaluation of cost-effectiveness, and assessment of implementation in real-world clinical practice. As emphasized by recent international clinical practice guidelines, future investigations should move beyond demonstrating improvements in lesion detection alone and establish the long-term clinical value of AI through robust evidence demonstrating sustained benefits for CRC prevention and patient outcomes. Until such evidence becomes available, AI should continue to be regarded as a decision-support technology that complements rather than replaces the expertise and clinical judgment of the endoscopist.

CONCLUSION

AI has emerged as an important technological advancement in colorectal endoscopy, with accumulating evidence demonstrating its ability to improve adenoma detection, assist real-time optical diagnosis, and support therapeutic decision-making. Current data indicate that AI can enhance several quality indicators of colonoscopy and contribute to greater consistency in lesion detection and characterization. Nevertheless, important challenges remain, including the need for broader external validation, demonstration of long-term clinical benefits, integration across different healthcare settings, and resolution of regulatory and medico-legal issues. At present, AI should be regarded as a decision-support tool that complements rather than replaces the expertise, clinical judgment, and procedural skills of the endoscopist. Continued technological refinement and high-quality prospective research will determine the extent to which AI can be effectively integrated into routine CRC prevention and therapeutic endoscopy.