For this narrative review, we searched PubMed/MEDLINE, Embase, Web of Science, Scopus, IEEE Xplore, and Google Scholar for articles published through April 2026. Search terms combined spine-related keywords, including “spine surgery”, “spinal disorders”, “lumbar disc herniation”, “lumbar spinal stenosis”, “vertebral fracture”, “spinal tumor”, “spinal deformity”, “spinal cord injury” and “spine rehabilitation” with AI-related terms, including “artificial intelligence”, “machine learning”, “deep learning”, “foundation model”, “large language model”, “multimodal model”, “clinical decision support”, “surgical navigation”, “robotic surgery”, “augmented reality”, “brain-computer interface” and “exoskeleton”. Additional searches included translational terms such as “multi-omics”, “drug-target prediction”, “virtual screening”, “absorption, distribution, metabolism, excretion, and toxicity (ADMET) prediction”, and “organ-on-a-chip”.

We included peer-reviewed studies evaluating AI applications across the spine-care continuum, including diagnosis, image segmentation and grading, surgical planning, navigation, robotics, intraoperative support, postoperative risk prediction, complication surveillance, rehabilitation, and translational research. Priority was given to clinical studies, external validation studies, multicenter evaluations, systematic reviews, meta-analyses, and methodologically relevant studies. Editorials, commentaries, conference abstracts, non-English articles, studies unrelated to spine care, and technical studies without a plausible clinical or translational connection were excluded. Isolated case reports were generally excluded, although landmark proof-of-concept studies in emerging fields were retained when more mature evidence was unavailable.

Database records and additional references identified through reference-list screening were deduplicated and assessed by title, abstract, and full text. Studies were retained when they informed the main clinical or translational themes of the Review. Direct evidence from spine-specific datasets was distinguished from indirect evidence from other surgical, radiological, rehabilitation, or biomedical fields. Evidence from non-spine settings was used only to support technical analogy or future research directions and was not treated as proof of clinical effectiveness in spine surgery. The selected literature was organized into five categories: diagnostic and multimodal data integration; surgical planning, simulation, navigation, and intraoperative support; postoperative prognosis and rehabilitation; molecular and translational research; and safety, governance, and implementation. To improve transparency, the clinical readiness of each AI application was assessed using a qualitative framework adapted from technology readiness concepts used in healthcare AI evaluation. The assessment considered technical validation, external validation, clinical workflow integration, and evidence of patient-centered outcomes. Because most spine AI applications remain at different stages of translation and lack uniform prospective evaluation criteria, readiness categories were used as descriptive summaries of evidence maturity rather than formal regulatory classifications. The literature search and selection process is summarized in Figure 3. Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 and Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension (PRISMA-S) informed the reporting of the search process, although this article was conducted as a narrative review rather than a systematic review or meta-analysis. Because the search was iterative, Figure 3 is presented as a schematic of the identification and screening process and does not provide numerical counts for each intermediate stage. A total of 147 references were cited in the final narrative synthesis.

Direct spine-specific evidence supports several narrowly defined imaging tasks, but integrated multimodal diagnosis remains less mature. Large multimodal models have been tested for lumbar foraminal stenosis, text-guided methods have linked magnetic resonance imaging (MRI) segmentation with abnormality identification, and external validation studies have assessed the generalizability of automated degeneration grading^[16-18]. Overall, automated spine-image analysis currently performs best in anatomical segmentation and the grading of specific abnormalities. These studies represent an initial move beyond isolated lesion detection, but they do not establish reliable integration of imaging outputs with structured reports and the broader clinical record. This distinction is clinically important because a useful lumbar imaging assessment requires not only lesion detection, but also accurate localization and characterization and explicit consideration of whether the imaging findings correspond with the patient’s symptoms and clinical history^[19]. Broader experience in radiology further suggests that successful clinical translation depends on close collaboration among imaging specialists, clinicians, technical teams, and other implementation stakeholders^[20].

Reliability also depends on data quality and compatibility. External validations of SpineNet in independent lumbar MRI cohorts show why performance should be assessed beyond the development dataset^[21,22]. Variation in imaging acquisition and dataset characteristics, together with missing sequences and examinations obtained at different points in the clinical course, can complicate registration, segmentation, longitudinal comparison, and downstream analysis. Deformable image-registration methods provide one technical approach to image alignment^[23]. Evidence from outside spine illustrates other possible strategies: domain adaptation has been used to address cross-device variation in thyroid ultrasound^[24], while multi-sequence MRI studies in brain and abdominal imaging show how complementary sequences can be integrated for segmentation^[25,26]. The applicability of these approaches to spine imaging remains to be established. General-purpose frameworks such as nnU-Net and an MRI study of dorsal root ganglion segmentation demonstrate the feasibility of automatically delineating selected anatomical structures^[27,28]. However, performance may not generalize across anatomical targets, patient populations, or imaging protocols. The preprocessing, registration, and segmentation pipeline should therefore be validated in the intended clinical setting before its outputs are integrated with other clinical information.

Clinical records add another layer of complexity. The significance of an imaging finding depends on its spinal level, timing, disease course, and previous treatment, and the same finding may be described differently across reports and notes. Knowledge-graph-enhanced clinical models illustrate how structured concepts can be linked to electronic health record data^[29]. Explainability work outside spine likewise illustrates the need to link model outputs to source features^[30]. A language model could help organize radiology reports, progress notes, and earlier examinations, and retrieve the text supporting a summary. but such use requires task-specific testing for calibration and for unsupported or clinically inconsistent outputs^[31,32]. Multimodal methods may then relate imaging findings to clinical text and structured patient data. Some architectures perform this in stages, first aligning local imaging features with the relevant report text and then combining this evidence with structured clinical information or question-guided representations^[33,34]. Within a clinician-supervised system, a foundation-model component could relate the outputs of specialist tools to the clinical record, standardize summaries, and flag uncertain or conflicting evidence for review. Lesion detection would still rely on dedicated imaging models, while interpretation and clinical decisions would remain with the clinical team.

Among common spinal disorders, lumbar disc disease has received particular attention in AI-assisted imaging. For this condition, clinically useful analysis requires more than labeling a disc as normal or abnormal. The involved level must be identified correctly, and the extent and pattern of degeneration or herniation must be described, including any effect on the dural sac or adjacent neural structures^[35]. Recent pipelines combine several of these tasks, including disc localization or segmentation with grading and abnormality classification^[36-39]. Graph-based and boundary-constrained methods can also model the anatomical relationships between vertebrae and intervertebral discs to improve segmentation consistency^[40]. Automated segmentation and measurement of the dural sac may provide additional quantitative information about canal narrowing^[41]. Together, these methods provide a more detailed description of lumbar disc abnormalities than a single classification label.

Vertebral fracture and lumbar spinal stenosis pose different imaging tasks. Mild osteoporotic fractures can be difficult to distinguish from chronic or degenerative vertebral deformity, particularly when vertebral height loss is limited^[42]. Recent fracture-detection pipelines commonly use a staged design in which vertebrae are first localized or segmented and then classified at the vertebral level^[43,44]. Lumbar stenosis is less a problem of detecting a discrete lesion than of grading anatomical narrowing consistently. Deep-learning models have been developed to grade central canal stenosis across multiple lumbar levels, while segmentation-based methods can quantify dural sac narrowing^[45]. Comparison across studies remains difficult because the anatomical targets, grading systems, datasets, and reference standards differ substantially^[46].

Rare spinal conditions are difficult to assess not simply because they are uncommon, but because their significance often emerges only when imaging is interpreted within a broader clinical history. Spinal tumors, congenital malformations, and syndromic deformities are heterogeneous, and relevant information may be distributed across institutions and different stages of follow-up. In syndromic deformity, systemic disease can directly alter surveillance and operative risk: FBN1-related disorders require attention to aortic disease and prior cardiovascular treatment^[47,48]. This heterogeneity is not limited to spinal morphology; genetic background and systemic comorbidities can also change the clinical significance of the same deformity. FBN1 variants have been linked to both Marfan syndrome and nonsyndromic scoliosis^[49]. In Turner syndrome, cardiovascular risk may directly affect the timing and planning of deformity surgery^[50]. A single spinal examination therefore provides only part of the information needed for multidisciplinary review.

Spinal tumors are a reasonable setting in which to evaluate imaging-based AI because radiology already plays a central role in lesion detection, localization, and treatment planning. Direct evidence from spine-specific studies, however, remains limited. AI methods have been used to localize poorly visualized tumors in radiotherapy settings, but their reliability in spinal oncology has not yet been established^[51]. In metastatic spinal disease, vertebral destruction, marrow replacement, and epidural or paraspinal extension help define the extent of disease and narrow the differential diagnosis. The pattern of involvement may sometimes suggest a particular tumor type or primary site, but imaging alone is rarely conclusive. These findings need to be interpreted alongside the clinical history, systemic staging studies, and, when indicated, histopathology^[52]. A multimodal system could make multidisciplinary review easier by bringing this information together and drawing attention to missing or conflicting evidence^[53].

Congenital and syndromic spinal deformities are typically evaluated longitudinally because growth can alter curve severity, spinal balance, symptoms, and treatment risks. In children who undergo repeated imaging, three-dimensional ultrasound with automated landmark detection has been investigated as a means of measuring spinal morphology without repeated exposure to ionizing radiation^[54]. These measurements still require interpretation in relation to vertebral development and changes in curve severity over time^[55,56]. Studies in adolescent idiopathic scoliosis have used radiographic and clinical data to predict curve progression^[57,58], providing a spine-specific precedent for longitudinal modeling. Time-dependent transformer models developed in oncology offer a broader methodological example for analyzing complex changes over time^[59]. But their applicability to congenital or syndromic spinal deformity remains unproven. Such approaches could help compare serial findings and flag unexpected patterns of progression, although this role has not been established in routine care. For the foreseeable future, the main contribution of multimodal AI in rare spinal disorders is likely to be the organization of longitudinal and cross-specialty information for clinical review.

Accurate preoperative assessment and procedurespecific planning remain central to spine surgery. Many of the decisions, however, still depend on the experience of the surgeon - in particular for anatomically complex or uncommon presentations. Patient-specific finite-element analysis can make biomechanical assumptions explicit, but conventional workflows depend on labor-intensive segmentation, meshing, and parameter assignment. Automated lumbar-spine pipelines now link deep-learning segmentation with finite-element modeling to reduce this burden^[60,61]. Deep-learning methods can also segment vertebrae, intervertebral discs, and the spinal canal from lumbar MRI^[62], while statistical shape models and related geometric methods support patient-specific mesh generation^[63]. Estimates of bone quality derived from Hounsfield units, particularly when combined with patient-level factors, may make biomechanical simulations more representative of the individual patient^[64]. These advances increase efficiency, but they do not by themselves establish that a simulated plan improves patient outcomes.

The clinical value of these tools lies less in producing a single “optimal” plan than in comparing reasonable surgical options under explicit assumptions. Risk calculators, registry-based models, and biomechanical simulations can help surgeons evaluate fixation levels, screw trajectories, implant choice, correction targets, and mechanical risk^[65-68]. Multiscale modeling may add local information about screw size, angulation, fusion level, and bone-implant stress^[69,70]. Orthopedic digital twins can be viewed as an emerging extension of patient-specific simulation, linking multiple data streams to support surgical simulation and prognostic modeling; however, the current evidence remains early and heterogeneous^[71]. In this setting, multimodal coordination could present anatomy, simulated options, model assumptions, estimated risks, and uncertainty in one reviewable interface. It should not select the procedure, determine fusion levels, or replace surgeon judgment.

Surgical simulation allows surgeons to examine patient-specific anatomy and rehearse difficult procedural steps before entering the operating room. Virtual reality, augmented reality, and mixed reality support different parts of this process: virtual reality provides a fully simulated environment, augmented reality overlays planned trajectories or structures on the operative view, and mixed reality anchors interactive three-dimensional models in the user’s physical space^[72-75]. Some platforms also incorporate haptic interfaces and deformable or synthetic tissue models that reproduce selected aspects of bone and soft-tissue behavior^[76,77]. In spine training, haptic systems have been used for pedicle screw placement and drilling, including models designed to distinguish cortical from cancellous bone^[78]. Multilayered simulators can also reproduce selected aspects of soft-tissue manipulation, pressure on neural structures, and intraoperative bleeding^[77,78]. Most published evidence for these systems comes from education and preoperative rehearsal; intraoperative applications have developed mainly through separate augmented-reality and image-guidance platforms^[79,80].

In the operating room, navigation and robotic systems represent a major pathway through which AI-assisted technologies are being incorporated into spine surgery. Navigation platforms integrate preoperative computed tomography (CT)/MRI data, intraoperative imaging, and registration algorithms to provide real-time localization of instruments relative to patient anatomy. Robotic systems extend this workflow by translating a preoperatively defined trajectory into controlled instrument guidance. The strongest clinical evidence still concerns pedicle screw placement: compared with conventional fluoroscopic or freehand techniques, robot-assisted navigation has demonstrated improved screw-placement accuracy in comparative studies and meta-analyses, although the magnitude of benefit varies among platforms, procedures, and outcome definitions^[81,82]. AI is increasingly embedded within these platforms. Machine learning enhances image segmentation, registration, and anatomical recognition from CT and fluoroscopy, enabling more accurate and efficient navigation^[83-85]. In parallel, deep learning-based path planning and reinforcement learning frameworks are being developed to enable real-time trajectory optimization, allowing robotic systems to adjust guidance as anatomy or positioning changes during surgery^[86-88].

Continued advances in navigation robotics are increasingly centered on the integration of multimodal data streams. AI-enabled planning systems may combine anatomical segmentation, deformity measurements, bone quality assessment, implant selection, and biomechanical simulation before surgery, then transfer these outputs into intraoperative navigation and robotic execution^[89,90]. To translate these capabilities into clinical practice, emerging platforms incorporating machine vision, 3D reconstruction, automated registration, and real-time workflow monitoring aim to reduce dependence on manual interpretation and improve procedural consistency^[91]. Key limitations persist: registration error, variability across institutions, equipment costs, learning curves, and insufficient prospective validation of patient-centered outcomes. Addressing these challenges, a future clinical orchestration system could coordinate planning algorithms, navigation data, robotic status, and uncertainty information in a unified interface while preserving surgeon control over every critical decision.

Postoperative risk is shaped by interacting baseline, procedural, and time-varying factors. Most published models have also been developed retrospectively, so their performance may decline in hospitals with different patient populations or postoperative care pathways. More refined outcome prediction could help clinicians explain risk more clearly, identify patients who may need closer surveillance, and adjust recovery plans before complications or delayed recovery become apparent. Multimodal AI may be useful in this setting because postoperative risk is usually shaped by several interacting factors rather than by a single measurement. For example, appropriately designed survival models can integrate continuous variables such as Cobb angle, categorical or ordinal factors such as the American Society of Anesthesiologists physical status classification (ASA) grade, and time-dependent intraoperative or postoperative physiological signals within a shared prognostic framework^[92]. As new clinical information becomes available, these models could update the estimated risk rather than rely on a single assessment made soon after surgery. External validation and calibration would still be required before such estimates could guide care^[93]. A recent comparative study in percutaneous kyphoplasty found that DeepSeek R1 performed similarly to conventional machine-learning models and modestly outperformed surgeon-only judgment in predicting bone-cement leakage, but it was less reliable in predicting subsequent vertebral fracture^[94]. Performance on one postoperative outcome therefore provides little assurance that the same model will perform well on another. Each outcome should be evaluated separately in the population and clinical setting in which the model is intended to be used.

Prediction and surveillance answer different questions: prediction estimates who may develop a complication, whereas surveillance asks whether current recovery is departing from the expected course. Continuous sensors and signal classifiers can detect physiological events^[95,96]. However, their clinical value depends on prospectively validated thresholds, false-positive rates, and actionability. Multimodal models may place postoperative changes in context and distinguish expected fluctuations from early wound, thromboembolic, cardiopulmonary, infectious, or delayed-recovery signals^[97]. Rehabilitation-specific reviews suggest that longitudinal summaries of strength, balance, pain, adherence, and function may support treatment review, while emphasizing inconsistent clinical effects and the need for real-world validation^[98-100]. Such summaries may prompt reassessment, but investigation and treatment must remain grounded in the patient’s examination and broader clinical context.

Multi-omics and computational molecular analyses belong to an outer-loop translational layer rather than the routine spine-care pathway. Their near-term purpose is to generate testable mechanisms, candidate biomarkers, and provisional endotypes for laboratory validation and future trial enrichment. Spine-specific studies have identified candidate pathways and network modules in intervertebral disc degeneration and adult degenerative scoliosis^[9,101]. Mechanistic work on circRNA CDR1as after SCI goes a step further by linking a defined molecular signal to mothers against decapentaplegic homolog (SMAD)-related fibrosis in a disease-relevant model^[102]. Broader biomedical and computational studies provide methodological precedents for integrating heterogeneous molecular data and modeling cellular, temporal, and subgroup variation^[103-111], but they do not establish causality or clinical utility in spine care. Molecular findings should therefore remain outside routine diagnosis and surgical decision-making until they are confirmed in relevant spinal tissue, supported by perturbation experiments, and prospectively associated with meaningful phenotypes or outcomes. Within the proposed system, a foundation-model component could organize and trace this evidence, but it should not convert an association into a mechanism.

Once a mechanism has adequate biological support, computational methods can help prioritize targets and candidates for experimental testing. Machine-learning-guided screening, generative design, and ADMET prediction can narrow chemical space and remove candidates with obvious liabilities^[10,112-116], but favorable rankings are not evidence of therapeutic effect. Candidates should advance only after demonstrating target engagement, activity in disease-relevant spinal tissue, acceptable exposure and toxicity, and reproducible benefit in appropriate models. Work on osteogenic pathways illustrates the type of mechanistic confirmation required^[117]. Organ-on-chip and perfused microenvironment systems may provide intermediate test platforms, although current examples remain general biomedical models rather than validated platforms for disc disease or spinal tumors^[118-120]. The output of this outer-loop layer should therefore be a transparent, experimentally supported target, candidate, biomarker, or trial-enrichment hypothesis - not a treatment recommendation. A foundation-model component may support evidence retrieval, cross-study comparison, and auditability, while decisions to advance a candidate remain with the translational research team.

Brain-computer interfaces (BCIs) address the loss of reliable communication between motor intention and movement after SCI. In a single-participant proof-of-concept study, a brain-spine interface decoded cortical activity and linked it to spinal stimulation to support voluntary standing and walking^[121]. Because this evidence comes from a single highly selected participant receiving intensive experimental support, it demonstrates technical feasibility rather than an approach ready for routine rehabilitation. Even so, the findings suggest that intention-guided neuromodulation may restore selected components of functional mobility under closely supervised conditions. Broader clinical use will require robust decoding across sessions, safer and simpler hardware, standardized outcome measures, and evidence from larger studies. One important direction is to make the link between neural decoding and stimulation more adaptive. Rather than relying on fixed stimulation patterns, future systems may need to respond to changes in motor intention, fatigue, posture, task demands, residual motor output, and signal quality^[122]. AI methods may improve signal classification, reduce the need for recalibration, and help track signal quality across repeated rehabilitation sessions. These real-time functions would be handled by task-specific decoders and device controllers rather than by an LLM.

The role of LLMs and related foundation models therefore should be positioned outside the real-time control loop. General BCI literature and an exploratory ChatGPT-BCI/virtual-reality proposal suggest possible higher-level roles in documentation, education, or information integration^[123,124], but neither validates an LLM orchestration layer in SCI rehabilitation. Reviews of steady-state visual evoked potential (SSVEP) and lower-limb motor-imagery systems instead identify patient-specific signal processing, noise, and repeated recalibration as core technical challenges^[125,126]. A clinician-supervised foundation-model layer would sit outside the real-time control loop, allowing device performance to be reviewed alongside pain, fatigue, and functional progress. This combined view may help the rehabilitation team notice when improved walking is accompanied by worsening pain or when declining signal quality precedes poorer performance. Translating BCIs into everyday clinical practice remains constrained by signal instability, repeated recalibration, latency, and the maintenance demands of implanted hardware. Gains in decoding accuracy or stimulation precision become clinically meaningful only when they reduce the burden on patients and rehabilitation teams and produce safer, more durable improvements in daily function. This proposed use of foundation models has not yet been validated in routine SCI rehabilitation. Future studies should therefore assess safety, usability, maintenance requirements, and the durability of functional benefits in real rehabilitation settings. Decisions about stimulation settings and rehabilitation progression would remain with the clinical team.

Robotic exoskeletons allow people with SCI to practice repeated overground walking with adjustable mechanical support. Therapists can vary assistance according to residual motor function, balance, fatigue, and safety needs^[127]. Current evidence suggests that exoskeleton-assisted training may improve walking capacity and selected functional outcomes in some patients with SCI, especially incomplete SCI. The size and durability of benefit vary^[7,128]. Translation to daily life is less certain. Use may be limited by donning and doffing time, fatigue, discomfort, environmental barriers, device access, and the need for trained assistance^[127].

Exoskeleton sessions generate clinically relevant data. These include kinematics, assistance level, walking distance, heart-rate response, pain, fatigue, adherence, the Walking Index for Spinal Cord Injury II (WISCI II), the 10-meter walk test, the 6-min walk test, and the Spinal Cord Independence Measure III (SCIM III)^[129,130]. A foundation-model system could summarize these data across sessions. It could help clinicians identify improvement, plateau, compensation, declining tolerance, or mismatch between clinic performance and daily function. Future trials should measure home and community use, caregiver burden, safety events, sustained adherence, and patient-reported value. Treatment decisions should remain with the rehabilitation team, consistent with expert guidance on exoskeleton prescription and therapist-in-the-loop control^[131,132].

When moving an AI tool into spine care, the starting question should be whether it improves practice, not whether its architecture is novel. A model is worth using only if, compared with current care, it improves the accuracy or consistency of clinical decisions, shortens review time, reduces risk, or leads to better patient outcomes. Evaluation should proceed from retrospective development and internal testing to external testing across institutions with different patient populations, scanners, imaging protocols, surgical practices, and rehabilitation pathways. Prospective silent deployment can then assess performance, calibration, failure modes, and workflow fit without allowing the system to influence care. Early live evaluation and, where justified, interventional trials are needed before outputs are used to guide clinical decisions^[15,133]. Reporting and appraisal should match the study design: prediction-model studies should follow Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD)+AI and be assessed with Prediction model Risk Of Bias Assessment Tool (PROBAST)+AI; imaging studies should follow Checklist for Artificial Intelligence in Medical Imaging (CLAIM); early live evaluations of decision-support systems should follow Developmental and Exploratory Clinical Investigations of Decision support systems driven by Artificial Intelligence (DECIDE-AI); and AI clinical trial protocols and reports should follow Standard Protocol Items: Recommendations for Interventional Trials (SPIRIT)-AI and Consolidated Standards of Reporting Trials (CONSORT)-AI, respectively^[134-138].

For multimodal AI, integration itself should be treated as a testable claim. A system combining imaging, clinical records, surgical variables, rehabilitation measures, or wearable data should be compared with strong single-modality models, validated specialist tools, and existing clinician-led workflows^[139]. Component analyses should determine whether retrieval, specialist modules, uncertainty estimation, and rule-based safety checks add measurable value or merely increase complexity. Studies should report calibration, performance across clinically relevant subgroups, failure patterns, safety during silent deployment, and effects on workload and decision-making. The same scrutiny applies to musculoskeletal digital twins: current work remains centered largely on patient-specific modeling and simulation, with limited evidence that these systems improve routine clinical care^[140]. Table 1 summarizes the current evidence and clinical readiness of AI applications across the spine-care pathway.

Responsible AI use depends on how data are collected, protected, and reused^[141]. Spine care generates sensitive longitudinal information, including imaging, surgical records, rehabilitation outcomes, patient-reported measures, and data from wearable or home-monitoring devices. These data may be reused for model development, external validation, updating, and post-deployment monitoring, making clear arrangements for consent, access control, cybersecurity, retention, and ownership essential. Federated learning may reduce the transfer of raw patient data between centers, but it does not remove the risks of re-identification, information leakage, security breaches, or uneven data quality^[143]. Patients should receive a practical explanation of how their data will be used, who may access them, and whether the data may be retained for future model updates or new applications^[144].

Average performance can conceal important failures. A model may perform well overall yet remain unreliable in rare spinal disorders, underrepresented populations, patients with atypical anatomy or complex comorbidities, and hospitals with limited resources^[145,146]. Subgroup evaluation should therefore be accompanied by continued monitoring after deployment, clear routes for reporting failures, and restrictions on use when performance is uncertain. A system that performs poorly in a defined population should be recalibrated, updated, restricted, or withdrawn from that use case. Responsibility for harm must also be agreed before deployment, because an adverse event may involve decisions made by the clinician, hospital, developer, and device vendor^[147]. Unclear accountability may encourage either excessive reliance on the model or reluctance to use it at all. Practical constraints also matter. If reliable computing, connectivity, technical support, and time for workflow redesign are available only in well-resourced centers, AI may widen existing differences in access to and quality of spine care.

The risks associated with AI-enabled systems partly depend on the degree of their integration with clinical operations. A language model used to summarize records or retrieve evidence has a different role from a navigation platform, robotic system, or neuromodulation device that interacts more directly with patient care. These functions should be evaluated separately. In high-risk settings, the broader clinical system may present information and outputs from validated specialist modules, but changes to an operative trajectory, instrument movement, or treatment delivery should remain under explicit clinician control^[142]. Knowing when not to provide an answer is equally important. If the available data are incomplete, poor in quality, internally inconsistent, or outside the intended population, deferral may be safer than a confident recommendation. The system should make uncertainty visible and indicate why a result may be unreliable. High-risk outputs should also be traceable to the evidence or specialist-module result on which they are based. A fluent explanation is not a substitute for adequate support and may, in some circumstances, give an unsupported conclusion undue credibility.

Clinical use also requires an auditable record of how an output was produced and handled. Relevant logs should include the system version, the information used, retrieved evidence, generated outputs, and any clinical review or override. These records allow institutions to investigate incidents, identify changes in performance, and determine whether a function should be updated or restricted. Bias monitoring and post-deployment surveillance belong to the same process. The aim is not simply to keep a clinician nominally “in the loop”, but to ensure that AI-supported decisions remain reviewable, contestable, and under clinical control. The proposed architecture is shown in Figure 4.