Discussions of surgical AI are often clouded by the failure to distinguish between levels of autonomy. Level-based frameworks for human-automation interaction predate surgical robotics: Parasuraman and colleagues’ model of automation, developed for aviation and process control, established the underlying logic that higher levels progressively shift information acquisition, analysis, decision selection, and action implementation from the human to the machine^[8]. To orient the discussion for surgery specifically, we adopt a level-based view of surgical autonomy adapted from recent regulatory and robotics literature [Figure 1]^[4,9-11].

Research demonstrations must be distinguished from clinical reality. The FDA’s AI-Enabled Medical Devices list, which as of 2025 includes over 1,000 authorized devices, illustrates this gap^[12]. As of this writing, no systems higher than Level 3 have been cleared or authorized for clinical use in any surgical specialty, and no autonomous system of any level has been cleared or authorized specifically for ACS^[4]. The vast majority of FDA-cleared surgical robots remain at the lowest autonomy levels (predominantly Level 1), providing teleoperation with motion scaling, tremor filtering, or virtual fixtures^[4]. The highest currently cleared autonomy is Level 3, comprising three systems: TSolution One (Think Surgical) for autonomous bone milling in knee and hip arthroplasty, iSR’obot Mona Lisa 2.0 (Biobot) for transperineal prostate biopsy, and the ARTAS iX System (Venus Concept) for robotic follicular unit harvesting; AquaBeam (Procept BioRobotics) for prostate resection is classified at Level 2^[4]. Critically, all operate on rigid or predictable anatomy with well-defined geometric targets, conditions fundamentally different from the inflamed bowel, adhesions, and hemodynamic instability of emergency surgery.

Even where model performance appears strong on development datasets, translation to clinical practice remains uneven. Loftus and colleagues’ 2023 review emphasizes that many surgical decision-support models exhibit deficiencies in scientific rigor and readiness for clinical use, including limited external validation and inconsistent reporting of clinically meaningful performance characteristics^[13,14]. Parallel themes appear in broader surgical AI guidance, which stresses that surgeons must understand model limitations, dataset provenance, and operational constraints before adopting these tools^[15,16].

ACS introduces specific pitfalls that compound these concerns. Time-dependent confounding, arising from the rapid physiologic changes characteristic of ACS patients, can invalidate predictions made moments earlier. Inequitable model performance becomes a patient-safety issue when social and structural determinants influence both how patients present and how they are treated. If AI systems are deployed as opaque authorities instead of transparent tools, bias and over-reliance become direct threats to care quality^[1,17].

A relevant ACS example of Level 2 autonomy is AI-enabled, ultrasound-guided robotic vascular access, a task with clear time sensitivity during resuscitation and frequent dependence on operator skill. Brattain and colleagues described a handheld system combining ultrasound guidance and robotic actuation to facilitate femoral access. Here partial automation amplifies and standardizes operator skill; it does not replace the operator^[19]. Automated vascular access currently operates at Level 2, executing a constrained sequence under continuous operator control; more advanced machine-learning models are under investigation to expand adaptive capability toward supervised conditional autonomy^[19-21].

Level 3 partial automation is increasingly appearing in the robotic surgery literature as conditional autonomy, where the system executes defined operative steps while humans supervise and manage exceptions^[22]. A representative example is conditional autonomy for robotic appendectomy, which explicitly positions autonomy as incremental and task-scoped; independent surgery is not the near-term goal^[23]. Similarly, semi-autonomous gallbladder retraction demonstrates a realistic near-term form of automation: exposure and tissue handling performed reliably in models. This frees surgeons to focus on higher-order judgment while preserving human oversight for critical safety decisions^[24].

A meaningful share of the autonomous-execution demonstrations cited in this section rest on preprints not yet peer-reviewed; we flag the most consequential of these inline. Some of the most ambitious autonomous surgery demonstrations to date have emerged from ACS-relevant procedures. The Smart Tissue Autonomous Robot (STAR) framework demonstrated Level 3 autonomous laparoscopic intestinal anastomosis in phantom, ex vivo, and animal (porcine) models, combining fluorescent imaging, 3D visual tracking, and force sensing to achieve dexterity comparable to expert surgeons; these are ACS-relevant preclinical demonstrations, not ACS clinical evidence, and no acute care patients were studied^[5,25]. This work suggests that, with further validation in contaminated and inflamed bowel typical of acute abdomen cases, autonomous suturing could eventually support ACS teams managing perforated viscus, strangulated hernias, or traumatic small bowel injuries^[26]. The Hierarchical Surgical Robot Transformer (SRT-H) extended this line of work, combining language-conditioned planning with low-level control to perform a full surgical step in ex vivo cholecystectomy models^[6]. Unlike prior systems that replayed fixed trajectories, SRT-H could adapt its behavior in response to intraoperative variation; whether this adaptability extends to ACS-type anatomical variation and hemodynamic instability, rather than the elective ex vivo model in which it was demonstrated, remains untested. A Level 3 autonomous suturing benchmark on the da Vinci Research Kit, spanning needle pickup, tissue insertion, and knot tying, has further demonstrated task-level capability^[7,27,28].

For ACS, technical feasibility cannot be defined solely by whether a model can perform a task. It must also encompass whether the model can quantify its own confidence and act safely when that confidence is low. Modern end-to-end systems in autonomous driving provide a useful analogy. A perception stack identifies objects and free space, a planning stack outputs an intended trajectory, and a low-level controller executes while continuously checking for deviations. A similar multi-layered architecture is needed for ACS autonomy. Before executing a maneuver, the system should communicate what it is about to do, an associated confidence estimate, and an indication of whether current visual and force patterns fall within or outside its training distribution. When the system detects an out-of-distribution scenario (dense adhesions, unexpected bleeding, distorted anatomy), it should automatically alert the supervising surgeon and trigger a predefined fallback: pause, retract to a safe pose (a predefined low-risk instrument configuration), pack the wound, or hand back full control. This framing echoes two of the concrete safety research problems Amodei and colleagues identified for autonomous systems generally: robustness to distributional shift, and safe exploration under uncertainty [preprint]^[42].

As autonomy increases, evaluation must extend beyond accuracy metrics to encompass human factors, failure-mode characterization, training and credentialing requirements, and real-world monitoring. The IDEAL (Idea, Development, Exploration, Assessment, and Long-term monitoring) Robotics framework offers a staged approach for development, comparative evaluation, and long-term surveillance that is directly applicable as ACS systems progress from prototypes to trials and post-market use^[18].

The task-specific readiness assessment in Table 1 complements this framework by mapping current evidence, uncertainty sources, failure modes, and proposed fallback behaviors to individual ACS subtasks. It identifies where bounded autonomy is most tractable and where the gap to safe deployment remains widest. Moving from qualitative readiness assessment to clinical trials will require defining quantitative performance targets for each task: acceptable complication rates relative to expert surgeon benchmarks, maximum allowable time for the surgeon to resume full control, and noninferiority margins that account for ACS patient heterogeneity. Together, the IDEAL framework and the task-readiness matrix offer a structured approach to sequencing investment and regulatory engagement across the autonomy spectrum.

For ACS, a demonstrable safety property is a pre-registered, quantitative, task-specific acceptance criterion, a higher bar than a general assurance that a system performed well in testing. No such criterion has yet been validated in a human study of a Level 3 system: a 2023 scoping review found all 129 intraoperative AI applications then published were still at IDEAL stage 0, with no human clinical investigations^[43], and a 2026 Society of American Gastrointestinal and Endoscopic Surgeons (SAGES) white paper names the absence of clinically meaningful performance metrics as one of the field’s central unmet regulatory challenges^[44]. What exists instead are endpoints from adjacent, better-established work. At the preclinical tier, the STAR intestinal-anastomosis platform reports the most concrete quality endpoints likely to carry into clinical evaluation: suture spacing and bite depth, anastomotic leak pressure, needle-placement corrections, completion time, and lumen patency, measured in phantom and in vivo porcine models with no acute care or human patients studied^[5,25]. At the clinical tier, the first-in-human Versius trial, though a teleoperated rather than autonomous platform, shows what an early conditional-autonomy study’s safety endpoints would likely look like: unplanned conversion, return to the operating room, and 30- and 90-day readmission, reported in a Stage 2a interim analysis built toward IDEAL-D Stage 2b^[45]. Building on this evidence base, we propose candidate targets here as illustrative starting points, each still pending task-specific validation: a force- or tissue-tension limit set below the ex vivo damage threshold for the specific tissue and task; a complication or leak rate noninferior to an expert-surgeon benchmark, with a margin pre-specified along the lines the CORE-MD consortium recommends for high-risk device evaluation^[46]; and, for out-of-distribution scenarios, a validated detection rate paired with a required rate of successful fallback execution for each task in Table 1. Takeover latency deserves its own, narrower target: the time from an out-of-distribution trigger to restored surgeon control is a different clock than the communication delay studied in teleoperation, where performance measurably degrades above roughly 300 ms and becomes difficult above 600 to 700 ms^[47]. Deriving any of these values is itself a research task, properly sequenced through the IDEAL framework’s staged evaluation.

The goal of robotic autonomy in ACS is not to imitate surgical performance but to amplify expert-level technical consistency. In settings where surgical expertise varies, whether due to training differences, fatigue, case volume, or geographic isolation, structured autonomy could reduce variability in technically sensitive subtasks by embedding learned patterns from experienced surgeons into systems that execute bounded tasks with precision under supervision.

The earliest autonomous capabilities to reach ACS practice will likely function as skill equalizers: a camera-holding arm that maintains exposure as reliably at 3 a.m. in a rural trauma center as it does during a daytime case at an academic medical center, or a suturing assistant that achieves consistent bite spacing regardless of operator fatigue. They narrow the gap between ideal and real-world operative performance without replacing surgical judgment. Full autonomous capability, the kind that could independently navigate a hostile abdomen, remains the distant end of the spectrum and is, at present, speculative^[48,49]. But bounded autonomy that raises the floor of technical performance across diverse care settings represents a meaningful and achievable near-term contribution.

When AI systems transition from recommending to acting, ethical risk increases, because errors become immediate physical harms. They gain the ability to harm the patient on their own, whereas previously they only misinformed the surgeon. In ACS, physiologic instability and time pressure intensify this shift. The ethical bar therefore rises from “keep a clinician in the loop” to specifying meaningful human control in operational terms: what the surgeon must be able to perceive (system mode, intent, uncertainty), how quickly takeover can occur, and what the system does when uncertainty or patient status changes (pause, safe-state, request confirmation). Operationally, this entails specifying a maximum acceptable takeover time, the minimum information continuously visible to the supervising surgeon, and explicit authority to override or terminate system action. It also requires assigning responsibility for case selection and contraindications, documenting the autonomy mode used in each case, and predefining accountability when failures arise from software updates, sensor errors, or inappropriate deployment. Autonomy level becomes a patient-safety variable rather than a marketing descriptor, and low-to-moderate autonomy is more defensible in ACS unless reliable interruption and fallback have been proven^[2,17].

Systems that act reshape accountability in ways that matter uniquely in ACS. Supervision is often intermittent during resuscitation. Competing priorities demand divided attention and the patient may deteriorate faster than a human can diagnose the automation’s failure mode. This creates a predictable responsibility gap unless institutions define accountability structures for case selection and contraindications, supervisory competence and credentialing, and escalation rules. A useful starting apportionment distinguishes three loci of responsibility for intraoperative harm: manufacturer or algorithm-developer liability for device defects, inadequate validation, or unsafe autonomous behavior within the system’s cleared envelope; institutional liability for credentialing decisions, case selection, and deployment outside validated indications; and supervising-surgeon responsibility for intraoperative oversight and timely takeover, consistent with an emerging framework for risk and liability in AI-enabled surgical systems^[50]. From a litigation-aware standpoint, ethical deployment requires a device-like evidentiary posture: version control, explicit documentation of the autonomy mode in use, and high-fidelity audit trails sufficient for root-cause analysis and transparent disclosure after harm, the foundation on which fair apportionment across these three loci actually depends.

The patient-surgeon relationship in emergency surgery is already fragile, often formed minutes before incision. Autonomy can either stabilize trust through reliability and standardization or erode it through perceived delegation and blurred responsibility^[51]. Ethical communication should be proportional to autonomy level and realistic for emergency contexts: patients or surrogates should hear, in plain language, what the system will do, what the surgeon will do, how safety and takeover work, and why this approach is appropriate for the case.