This study offers one of the first structured assessments of surgeon perspectives on intraoperative AI assistance in robotic surgery, bridging the gap between algorithm design and the support surgeons consider most useful.

Our findings indicate that surgeons are willing to employ a wide range of intraoperative AI tools, although their preferences vary depending on the type of assistance provided. The perceived usefulness of the evaluated functions followed a clear gradient. Anatomy recognition occupied the upper end of this spectrum, followed by risk-detection and vision–language model guidance. Decision-making assistance and step-recognition support were placed toward the lower end of the gradient yet both remained within the range of functions considered useful by most respondents. Overall, these results indicate that surgeons value all intraoperative AI functions assessed, while favoring tools that enhance anatomical recognition over those oriented toward procedural segmentation or decision-making.

Beyond the perceived usefulness of individual functions, our results reveal a gap between surgeons’ attitudes and their current use of AI applications. More than 86% of respondents agreed that intraoperative AI assistance could positively impact surgical performance, particularly during early stages of training, and 75.5% expressed confidence in relying on clinically validated tools. By contrast, 79.2% reported never using intraoperative AI tools during surgery. This mismatch suggests that current applications may still lack the availability, accessibility, integration, and reliability required for routine clinical use. Recent studies have attempted to explain this gap, consistently citing persistent barriers related to workflow integration, training requirements, system costs, and the need for robust oversight and validation in real-world settings [13,14,15].

From a translational perspective, these findings allow preliminary recommendations regarding how preferred AI functions could be integrated into intraoperative workflows. Overall, surgeons appear to envision intraoperative AI primarily as an assistive, context-aware system rather than a decision-maker. Highly ranked functions such as anatomy recognition and risk detection emerge as the most clinically relevant entry points for implementation, as they directly support safety and anatomical understanding during critical surgical steps. In contrast, lower-ranked functions, including surgical step recognition and decision-making assistance, seem to be perceived as complementary features that may be best implemented as optional or on-demand tools, thereby limiting unnecessary interruptions and cognitive load. The consistently high ratings for vision–language model assistance further indicate interest in flexible user-driven interaction, reinforcing a model of intraoperative AI that prioritizes safety, usability, and surgeon control. Taken together, these findings support the development of a modular, integrated intraoperative AI assistance system in which individual components can be selectively activated based on surgical context, experience level, and clinical need.

Experience-based stratification showed that surgeons across all experience levels consistently recognized the usefulness of intraoperative AI, in contrast to prior work by de Jong et al., which reported expertise-dependent variation in how AI outputs are interpreted, particularly in anatomy-segmentation tasks [10]. In our study, no significant difference in attitudes, confidence, or perceived usefulness was observed across experience groups. The uniformly low rates of intraoperative AI use across all groups further suggest that limited access and availability, rather than experience-dependent factors, may be the primary barrier to adoption. Although the survey question specifically referred to intraoperative surgical AI tools, the possibility that some respondents interpreted this broadly cannot be entirely excluded.

Most survey studies on AI in surgery evaluate surgeons’ expectations in domains such as diagnosis or perioperative management rather than examining intraoperative implementation in the context of robotic surgery [15,16,17,18]. Nevertheless, they consistently report positive attitudes toward AI, a pattern that aligns with the optimism observed in our cohort despite the differing clinical scope. In another survey study, Pecqueux et al. found that most participants rated their own knowledge as average or rudimentary and acknowledged limited use of AI tools in their clinical environment [17]. This combination of favorable attitudes, average self-reported knowledge, and restricted implementation is consistent with our findings.

Several limitations of this study should be acknowledged. First, although the sample was expanded through a snowball distribution strategy, recruitment was initiated at a single minimally invasive techniques course, which may have introduced selection bias favoring surgeons with prior interest in implementing new technologies. Even so, the cohort encompassed the full spectrum of training levels, ensuring representation from residents to senior experienced consultants, and represented 19 countries across 5 continents, reflecting broad international participation. A further limitation is that respondents with limited knowledge on AI may have held inaccurate or incomplete conceptualizations of intraoperative AI, which could have influenced how they interpreted the presented intraoperative functions, as suggested in previous surveys [16]. Moreover, the questionnaire was developed specifically for this study, and its measurement validity was not formally assessed. Finally, AI functions were assessed using static images, which may not reflect how surgeons interpret such tools during live procedures. Further research should assess these functions in real-time surgical settings to better define their practical clinical value and generalizability.

Irrespective of these limitations, our study helps fill a gap repeatedly highlighted in prior work: the need to understand how surgeons want intraoperative AI to be implemented and which functions they consider most valuable for real-time use [10, 16]. Defining these requirements provides a foundation for future development, guiding the creation of AI tools that can be effectively integrated into the operative field.

In conclusion, surgeons across experience levels highly value intraoperative AI assistance, particularly for anatomy recognition. Current limited adoption likely reflects barriers of availability and implementation rather than lack of interest, and defining these preferences provides a foundation for developing clinically meaningful AI tools.