From Automation to Collaboration: The Emerging Role of Artificial Intelligence in Endoscopic Reporting

For endoscopists, difficulty in recalling the exact locations or number of polyps removed after a long list of procedures is not uncommon, and discrepancies between a referring physician's report and the actual location or morphology of lesions can at times compromise treatment plans. Such discrepancies often stem not from limitations in skill or experience but from structural factors, including the pressure to complete reports quickly under heavy workloads and sustained cognitive demands. Although structured reporting has improved the quality and standardization of endoscopic documentation, it has also reduced flexibility and increased the workload, which may, in turn, contribute to burnout [1]. A review by Sekiguchi et al. provided a comprehensive overview of artificial intelligence (AI)-assisted reporting technologies as a potential solution to these long-standing challenges. Distinct from a conventional literature review, their paper serves as a practical technology review, outlining both the current state and clinical feasibility of systems already in use, including speech recognition systems such as Voice Capture (Lasis, Osaka) and VoiceRex (NTT, Tokyo), as well as report generation tools, including AR-G1 (Fujifilm, Tokyo) and Vivoly+ (Olympus, Tokyo) [2]. These platforms represent the first wave of AI-enabled endoscopic reporting, integrating speech, text, and image data to facilitate documentation. Speech recognition for medical reporting has been investigated for over three decades [3]. Recent studies have demonstrated that automated voice reporting can reduce documentation time by approximately 1 min per case, a modest improvement that nonetheless carries significant impact during the high-pressure operation of a busy endoscopy suite. However, widespread adoption of endoscopy remains limited. Factors such as latency, the delay between the command and system response, along with non-responsiveness or recognition errors (reported in nearly 10% of cases), remain major barriers [4]. The real-time correction of these errors is particularly cumbersome during endoscopic procedures involving both hands. Conversely, in radiology, where image interpretation (reading) and reporting are generally performed retrospectively and asynchronously, allowing radiologists to review images and complete reports at a convenient time after acquisition, speech-based reporting is widely adopted in certain regions [5]. With rapid advances in deep learning and large-scale language models (LLMs), speech recognition accuracy has reached human-level performance [6]. Multimodal AI systems capable of simultaneously processing speech, text, and endoscopic images raise the realistic prospect of seamless, interactive documentation. Further improvements in user interface design, particularly in synchronizing spoken findings with the corresponding time points in image sequences, could render real-time AI reporting a practical reality in endoscopy. Endoscopy platform providers are leveraging their scales to develop next-generation, integrated, data-driven AI solutions that are intended to eventually combine computer-aided diagnosis functions such as lesion detection and characterization with workflow support, including automated reporting. Another significant contribution of the review by Sekiguchi et al. was its discussion of automated quality assessment using endoscopy report data. To date, quality metrics, including completeness of imaging, withdrawal time, and adenoma detection rate, have depended largely on manual reviews. If AI can automatically extract parameters, such as imaging sites and resection timing, the administrative workload would be dramatically reduced. Importantly, minor discrepancies in these metrics rarely influence patient outcomes, making this a low-risk, high-return area for automation. Looking ahead, the application of LLMs presents new possibilities. LLMs can generate structured summaries or complete draft reports using natural human-like language. However, general-purpose models such as ChatGPT currently lack sufficient accuracy for clinical use. Domain-specific fine-tuning incorporating medical terminology, procedural contexts, and reporting conventions remains indispensable [7]. Moreover, as Sekiguchi et al. emphasized, AI implementation must be accompanied by careful consideration of ethical issues, including data privacy, factual reliability, bias, accountability, and transparency. Among these, “AI hallucination,” referring to the tendency of LLMs to generate plausible yet factually incorrect statements, represents a particularly serious concern. In a study involving AI-summarized radiology reports, more than 50% of the generated statements comprised hallucinated or omitted findings [7]. Such errors can be challenging to identify precisely because they appear convincing; in medicine, a missing malignancy or fabricated benign lesion is functionally equivalent to a clinical “oversight.” Consequently, all AI-generated content must undergo systematic human verification and be subject to traceable auditing. Recent research has explored AI-monitoring architectures [8], in which one model audits another's output. Although promising, fully autonomous reporting remains unrealistic and arguably undesirable in medicine. The key concept here is the human-in-the-loop (HITL) framework, a collaborative cycle in which AI generates proposals, humans verify them, and the model is refined based on feedback [9]. HITL serves as an “antidote” to AI overreliance. Rather than abandoning automation, HITL aims to integrate algorithmic intelligence with human judgment to ensure safety and trustworthiness. Endoscopy may be particularly well-suited to this collaborative model. Unlike radiology, where images are mechanically captured and interpretation is often separated from acquisition, endoscopists themselves generate images and retain full responsibility for their procedures and reporting. This proximity between the operator and the data allows clinicians to readily detect subtle inconsistencies or implausible AI outputs. Each finding in an endoscopic report is usually accompanied by an image, enabling direct visual cross-checking and thus high transparency. However, in radiology, LLMs are being experimentally explored as real-time quality monitors, not to replace human judgment, but to support it. In a recent study, an LLM reviewing voice-generated radiology reports detected all clinically significant errors, including speech recognition failures and human verbal slips [10]. Moreover, HITL systems, which integrate automation with human oversight, highlight how efficiency and safety can be enhanced simultaneously. Therefore, the ideal future workflow in endoscopy might proceed as follows: immediately after a procedure, AI generates a preliminary structured draft report based on voice commands, image analysis, and metadata. The endoscopist then reviews, corrects, and approves it as the final report. Beyond efficiency, human-AI collaboration offers substantial educational value. By providing standardized phrasing and highlighting overlooked findings, AI-assisted reporting can accelerate learning and promote diagnostic consistency among trainees. Ultimately, the goal is not complete automation, but balanced collaboration, a system in which AI augments human capability while preserving the clinical accountability of physicians. The timely and integrative review by Sekiguchi et al. illuminates this path of bridging technology, practice, and ethics. As AI gains the ability to draft reports, the critical question shifts from whether machines can think or write like humans to how humans can continue to stay meaningfully in the “loop”. S.K. and K.S.: study conception or design. S.K.: drafting the work or revising it critically for important intellectual content. S.K. and K.S.: final approval of the version to be published. S.K. and K.S.: agreement to be accountable for all aspects of the work, ensuring that any questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The authors have nothing to report. The authors have nothing to report. Lpixel Inc. and The Jikei University School of Medicine hold the Japanese Patent No. 7127779. S. Kamba and K. Sumiyama are the inventors of the patent. This article is linked to Sekiguchi and et al. papers. To view this article, visit https://doi.org/10.1111/den.70057.