Artificial intelligence in anaesthesiology: why don’t we have it in our hands after a decade of innovation? A systematic review and perspective Authors: Rémi Florquin, Philippe Dony Published: 2026 ABSTRACT AbstractBackgroundMachine learning (ML) tools are increasingly integrated into various sectors, including healthcare, where they have demonstrated disruptive potential.While anaesthesia is a specialty historically shaped by technological innovation, the precise clinical impact of ML remains poorly defined. Furthermore, despite a decade of prolific model development, the systematic translation of these computational tools into routine, everyday clinical workflows has stagnated.MethodsA systematic literature review was conducted using the PubMed/Medline and EBSCO databases, supplemented by an analysis of Food and Drug Administration (FDA)-approved medical ML applications. To ensure methodological rigour, the selection process followed the PRISMA 2020 guidelines. Eligible peer-reviewed publications were restricted to those focusing primarily on the development, validation, or implementation of machine learning applications within the field of anaesthesia.ResultsOut of 6,425 screened publications, 1,021 were included in a datasheet. These included 302 reviews or opinion pieces, 5 case reports, and 714 articles focused on tool development or clinical evaluation. Although research topics were highly diverse, the vast majority of studies focused on the prediction of perioperative complications and patient prognosis.ConclusionsMachine learning represents a highly active and debated domain in anaesthesia, characterised by a substantial volume of published research. However, exceptionally few algorithmic models have successfully translated into practice-changing clinical tools. This persistent gap indicates that technical innovation alone is insufficient, as translation is severely hindered by inherent model limitations, software interoperability constraints, and socio-technical challenges within the clinical environment. Addressing the scarcity of hybrid clinician–developer profiles and establishing robust, external field validation are critical pre-requisites for meaningful clinical integration.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12871-026-03995-6.Keywords: Artificial Intelligence, Machine Learning, Innovation, Data-driven change Introduction What is intelligence? Even this simple question can lead to hours of debate. In The Frames of Mind, Howard Gardner proposed a model of multiple intelligences, identifying at least eight distinct types. When we describe someone as "intelligent," what exactly are we referring to? The answer depends on the criteria we use to define intelligence. Is it their ability to analyse complex mathematical problems? Their skill in social interactions and team management? Intelligence is not a singular, fixed trait but rather a diverse and multidimensional concept. Beyond this, adding the term "artificial" further complicates the discussion. In this work, we avoid the broad and ambiguous notion of AI, opting instead for more precise technical terms such as Machine Learning (ML). When the term AI is unavoidable, we refer to it as defined by Russell in Artificial Intelligence: A Modern Approach [1], which describes AI as the study of agents that receive percepts from the environment and perform actions. This definition is both concise and conceptually robust, as it highlights two fundamental aspects of AI: perception, or the integration of information through data, and action, which underscores the importance of the task assigned to the agent. If you ask a neophyte when artificial intelligence was conceived, or at least when its founding paper was published, they will likely guess the 1990s, 1980s, or, at the earliest, the 1970s. You might get a puzzled look if you tell them that the founding paper of modern AI dates back to December 1943 with the paper A Logical Calculus of the Ideas Immanent in Nervous Activity [2]. This paper, authored by a physician and a psychologist, introduces the concept of artificial neural networks through their work in neuroscience, implicitly anticipating the foundational principles of what would later become connectionism. This concept inspired Frank Rosenblatt to create the Mark I perceptron in 1958, which can be seen as the ancestor of current neural network tools. One might ask why ML has only experienced rapid advancements in the past two decades. There are several reasons for this. The digitisation of information and data, is the essential element for the development and validation of ML tools. In the 1980 s, less than 1% of the world's data were in digital form and easily usable by computers, with the rest existing in analogue form. By 2007, this figure had increased to 97% [3]. In terms of quantity, on a global scale, humanity stored 2.6 exabytes (1018) of data in 1986. By 2007, this figure had risen to 295 exabytes. Thus, data are only more readily accessible but also vastly more abundant [3]. Computational power has also increased dramatically, enabling the processing of larger datasets with more computationally intensive algorithms. In 1986, global computing capacity was 3.0 × 108 million instructions per second (MIPS). By 2007, this figure had skyrocketed to 6.4 × 1012 MIPS—an increase of more than 20,000 times [3]. Finally, continuous advancements in computer science and increasingly sophisticated algorithmic architectures enable the effective utilisation of this vast amount of data and computational power. Artificial intelligence, concurrently with advances in robotics, computer science, advanced automation, and the Internet of Things, constitutes a driving force behind the ongoing Fourth Industrial Revolution. Like its predecessors, this revolution is set to transform society, the nature of work, and human interaction [4]. Artificial intelligence is a prominent topic impacting all sectors of society and spanning generations. Frequently serving as a focal point in marketing discourse, it has rapidly become integrated into daily human life, often operating imperceptibly. Medical practice is already being affected in varying proportions depending on specialisation. While anaesthesia is traditionally recognised as an innovative, technologically oriented discipline, its integration of clinical AI tools has been unexpectedly slow. A significant amount of research on AI in anaesthesia has been published over the years, with a recent paper that explores decision-making aids for increasing the quality of care [5]. The transition of these technologies into daily clinical practice remains limited. This disparity raises critical questions regarding whether current AI capabilities in medicine are partly driven by commercial enthusiasm, or if systemic barriers are preventing the deployment of genuinely useful tools. The primary objective of this analysis is to evaluate the evolution of these limitations over the past decade to identify the precise obstacles blocking the clinical translation of AI in anaesthesia. In addition, this review aims to formulate actionable solutions to enhance the effective integration of AI in clinical workflows. To achieve this goal, a systematic analysis of machine learning applications in anaesthesia is needed to assess the current state of the field. Methods FDA review In March 2025, an author (RF) conducted a systematic review of the artificial intelligence and machine learning-enabled medical devices of the U.S. Food and Drug Administration (FDA). Each instance with an anaesthesia label has been reviewed, and its FDA clearance has been examined to verify its relevance to anaesthesia-related applications. The information collected in a table included the decision date, submission number, name, and company. The application was then categorised on the basis of its intended use. Literature review In March 2025, an author (RF) conducted a systematic literature review via the PubMed/MedLine search engine. The PRISMA 2020 flow diagram for systematic reviews was used to guide the literature review, and the PRISMA 2020 checklist was applied to ensure the quality of the search and methodology [6]. The keywords used were 'anaesthesia' or 'anaesthesiology' combined with 'machine learning' or 'artificial intelligence'. The period of study was from the first of January 2000 to the first of March 2025. In June 2025, a second round of research was conducted on MedLine via the MESH method with Anaesthesia and Artificial Intelligence research as the major topic. In addition, a review of the EBSCO database was conducted with the keywords anaesthesia AND (machine learning or artificial intelligence or deep learning or neural network). Title screening was followed by abstract screening. For articles not written in French, English, or Dutch, an online translator was utilised (Reverso.net®). Each result that met the selection criteria and did not fall under the exclusion criteria was recorded in a table. In cases of uncertainty regarding the relevance of including an article, the second author (PD) decided whether the article was off-topic or not. The collected data included the year of publication, title, DOI and type (case report, review or opinion, and development or validation). The duplicates were automatically detected by a formula in the table that detects similar DOIs. For articles focused on tool development or clinical evaluation, we collected additional information, including the country of the first author’s institution. The topics addressed were categorised into seven areas: prediction of complications, depth of anaesthesia, techniques, organisational aspects, ultrasound, monitoring, and chatbots. Complications included undesirable events related to anaesthesia (e.g. difficult intubation) or surgery but where the anaesthetist is involved in prevention or treatment (e.g. surgical bleeding). Selection criteria The publications had to be indexed in PubMed/MedLine or EBSCO.. The primary focus had to be the use of machine learning tools related to anaesthesia or perioperative medicine, whether in the preoperative, intraoperative, or postoperative setting. The types of publications included literature reviews (regardless of type), opinion articles or letters to the editor, case reports, and studies on clinical development or evaluation. Exclusion criteria Articles without available abstracts were rejected. For development or evaluation articles, if the type of algorithm was not specified in the abstract, it was sought in the full text. If no description of the mathematical tool was provided, the article was excluded. Language was not considered a reason for exclusion due to the accuracy provided by today's machine translators. Retracted papers and those not accepted by peer-reviewed journals (preprint) were excluded from the analysis. Ex vivo studies, articles related to veterinary anaesthesia, studies focused on the influence of genetics in anaesthesia, animal studies, and those concerning purely local anaesthesia were excluded. The type of algorithm, research question, number of patients, quality of the work, type of funding and results were not used as exclusion criteria. The objective was to provide an overview of available tools and visions in the literature. Publications without DOIs or PMIs were excluded. Results FDA review Of the 1,016 machine learning tools approved by the FDA, 17 were listed under the anaesthesia category. Their approval dates range from 1997 to 2024, with the majority granted between 2023 and 2024. (see Table 1). Table 1List of FDA-approved ML-based productsDate of Final DecisionSubmission NumberDeviceCompanySubject09/13/2024K240929Sleep Apnea Notification Feature (SANF)Apple Incsleep apnea07/15/2024K240285Huxley SANSA Home Sleep Apnea Test (1000–00)Huxley Medicalsleep apnea07/02/2024K240555Tyto Insights for Crackles DetectionTyto Care Ltdcrackles detection02/09/2024K231355AuroraEnsoDatasleep quality02/06/2024DEN230041Sleep Apnea FeatureSamsung Electronics Co., Ltdsleep apnea12/13/2023K232237Tyto Insights for Wheeze DetectionTyto Care Ltdwheeze detection10/06/2023K232787ScanNav Anatomy Peripheral Nerve BlockIntelligent Ultrasound Limitedlabel image regional anaesthesia03/31/2023DEN200011Masimo SafetyNetMasimo Corporationopiod induced respiratory depression02/23/2023K222579Belun Sleep System BLS-100Belun Technology Company Limitedsleep apnea02/17/2023DEN210022PMD-200Medasense Biometrics Ltdnociception level10/05/2022K221892VISIONAIRPacificMD Biotech Llcmeasurement of the nasal respiratory airway01/02/2020K183089Air NextNuvoAir ABspirometry testing02/04/2019K181916d-Nav SystemHygieia, Incinsulin management05/05/2017K163665MATRx plusZephyr Sleep Technologiessleep apnea11/05/2015K150102SnoreSoundsAPPIAN MEDICAL, INCsleep apnea04/24/2015K142918EndoTool SubQMONARCH MEDICAL TECHNOLOGIESinsulin management03/11/1997K955841SLEEP SYSTEM FOR CLINICS & RESEARCHCOMPUMEDICS PTD LTDsleep apneaThe information listed includes the date of the decision, the device, the company that developed the product and the subjectSource: https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices List of FDA-approved ML-based products The information listed includes the date of the decision, the device, the company that developed the product and the subject Source: https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices Upon analysing individual approvals, 8 devices (47.1%) were designed for the detection of sleep apnoea. Two (11.8%) were related to insulin therapy management, and two (11.8%) to the detection of pathological sounds during pulmonary auscultation. One device (5.9%) was developed to measure the nasal cavity, one (5.9%) to analyse spirometry results, one (5.9%) to detect opioid-induced hypoxia, one (5.9%) to evaluate a nociception score, and one (5.9%) for image labelling in locoregional anaesthesia. Literature review The searches yielded a total of 6,425 abstracts. After removing duplicates and applying the exclusion criteria, 1,021 entries were included in the final analysis (see Fig. 1). Two abstracts could not be retrieved, and 9 were retracted after their publication. PRISMA flow diagram. Legend: Source: Page MJ, et al. BMJ 2021;372:n71. 10.1136/bmj.n71 Of these, 714 (69.93%) were articles focused on the development or evaluation of a tool based on, or supported by, machine learning. A further 302 (29.58%) were reviews or opinion papers, and 5 articles (0.49%) were case reports (see Fig. 2). Type of publications. Legend: proportion of type of publications of machine learning in anaesthesiology. Development and evaluation paper (D&E), reviews and opinion and case reports A notable trend is the sharp increase in publications over time. More than half (56.12%) of all included papers were published within the last 2.5 years, despite the search covering the past 25 years (see Fig. 3). Time distribution of publications. Legend: timeframe of publication about machine learning in anaesthesiology. Note that 2025 was in progress and that the review took place in March The countries of origin for the development and evaluation studies demonstrated significant geographical variation. Two countries accounted for more than half (50.14%) of these publications: the United States (25.77%) and the People's Republic of China (24.36%). The top five countries (also including South Korea, Canada, and Taiwan) together accounted for 67.78% of the total (see Fig. 4). Major publication countries. Legend: representation of the top 10 countries producing publications on machine learning in anaesthesia. Note that China refers to the People Republic of China and Taiwan to the Republic of China, and Korea to South Korea In terms of subject matter, the majority (57.70%) of development or evaluation papers focused on patient prognosis or the prediction of specific complications. A further 11.34% addressed the depth of anaesthesia via EEG. Studies examining techniques or methods represented 8.54%, whereas 7.00% focused on organisational aspects or patient flow or cost optimisation. Papers related to ultrasound accounted for 5.74%, those on monitoring techniques accounted for 5.46%, and studies evaluating chatbots or large language models (LLMs) accounted for 4.20% (Fig. 5). Theme of publications. Legend: proportion of theme of machine learning publications in anaesthesia Discussion Our review illustrates that artificial intelligence in anaesthesia has generated substantial discussion in the literature. Numerous applications have been described and evaluated, with approximately two-thirds of the publications focusing on the development or evaluation of tools and one-third comprising reviews or opinion writings, indicating that the debate remains active. Interestingly, remarkably few case reports or FDA-approved products exist. The majority of AI-related papers on anaesthesia focus on prognosis or complication prediction. This emphasis is understandable since the availability of data is vital for developing and evaluating machine learning tools. Complications are routinely documented in patient records, and it is often a matter of cross-referencing these data with relevant clinical parameters to attempt prediction. This explains why such research dominates the field, whether addressing intraoperative or postoperative complications. It is a promising area: overall mortality, delirium, postinduction hypotension, and difficult intubation have all been studied across various contexts, with mixed results in terms of performance. The assessment of anaesthetic depth is also well represented in the literature, largely because of the availability of EEG data, which provide a rich dataset for algorithm development. This application is the one integrated into anaesthesia practices, with many commercial products utilising proprietary algorithms to monitor the depth of anaesthesia. In terms of techniques, studies have investigated applications such as closed-loop systems that adjust syringe pump infusion rates, prediction models for spinal anaesthetic dosing, and ventilators optimised through predictive algorithms to determine ideal ventilation parameters. Although a wide range of topics is covered, the effectiveness of these tools varies, and few, if any, have been adopted into routine clinical practice. Integration difficulties, a lack of validation by health authorities, or disappointing results may cause this relative delay between what is technically possible and what is routinely used. Some publications address organisational aspects, such as predicting operating times to optimise patient flow in the operating theatre or forecasting cost. While not directly related to bedside anaesthetic practices, these tools may offer valuable support to clinical leadership and administrative planning, but few such tools have been successfully integrated into clinical practice to date. With respect to ultrasound, despite the abundance of potential imaging data, the number of published ML applications remains low compared with those in radiology. This is unexpected, given the clear potential in terms of both education and patient safety. One of the few FDA-approved tools is, in fact, an image labelling system for local and regional anaesthesia, suggesting promise in this area. The relative lag in ML development may be attributed to a lack of industry investment, or perhaps to the fact that, unlike radiologists, anaesthetists rarely archive or label ultrasound images systematically. There are emerging ideas for applying ML in the field of monitoring, such as the analysis of arterial pressure waveforms or the assessment of nociception. The practical efficacy of these tools has yet to be determined but has the potential to increase the accuracy of our intraoperative awareness. Notably, LLMs such as ChatGPT® and Claude®, which are highly visible and widely accessible to the general public, have also sparked interest in the anaesthesia community. For example, there have been proposals to explore the use of LLMs in high-stakes assessments, such as board examinations to assess their applicability for information retrieval such as the correlation between the instruction given and the recommendation for the adaptation for stopping anticoagulants before neuraxial anaesthesia. These discussions reflect the increasing integration of LLMs into both professional and private spheres and, perhaps, also a degree of apprehension among some practitioners. This concern may stem from a perceived erosion of the clinician’s traditional role as a gatekeeper of medical knowledge, a shift that arguably began with the widespread availability of information via the internet. A notable observation is the extremely low number of published case reports in the field. This indicates a lack of bottom-up feedback, suggesting that these tools are still infrequently used outside their original development teams. The absence of input from end-users reflects limited integration into routine clinical practice. For developers to create robust and reliable tools, it is crucial for them to receive feedback when their models fail or yield unexpected results. The current scarcity of such reports, whether due to an actual lack of usage or the difficulties associated with publishing in today’s academic environment, represents a significant barrier to the ongoing improvement and broader adoption of ML in anaesthesia. No product, statistical or pharmacological, is ever perfect or fully understood at launch, and case reports are essential for building end-user confidence as they provide valuable insights into these products. The significant gap between the number of ML tools developed or described in the literature and the number of FDA-approved products, or the relatively small number incorporated into clinical practice, raises several important questions. One may question whether the results presented in these studies are truly generalisable, or whether the tools underperform once implemented clinically owing to overfitting during development (a prediction too close to the training data, leading to a drop in performance on new data) or the use of suboptimal models that were not identified because validation was limited to internal datasets. Additionally, the challenge of validating such tools, which are often highly sensitive to population disparities or local factors, may discourage developers from progressing beyond the initial publication. Data: a critical issue and major challenge For the authors of this paper, one of the most significant challenges in the development of ML tools is the issue of data, its quality, quantity, and lack of standardisation. Data are often referred to as the new gold. It forms the foundation upon which all machine learning models are developed. The effectiveness of an algorithm in performing a specific task is entirely dependent on the quality of the data it is trained on. For instance, if a hospital provides a dataset to develop a model for predicting postoperative mortality, the model’s performance will be strongly influenced by the quality of that data. The type of model selected will also rely on the type of data (nominal, ordinal, numerical, etc.) together with the volume of data available. More complex models typically require larger datasets, although complexity does not always guarantee better results. Once a model has been trained, it needs to be evaluated on an independent dataset and, most importantly, tested in real-world clinical settings to determine its practical applicability. This final stage is the most challenging to implement but is essential for ensuring the tool’s relevance. Let us consider that the model developed above shows excellent results and attracts interest from a hospital in another country. However, clinical practice, equipment, data encoding methods, and patient characteristics, such as the prevalence and types of chronic diseases, contrast significantly. In such a case, it is highly likely that the model, despite performing skilfully in its original environment, will prove less effective in this new setting. This is a classic example of poor generalisability across populations. An additional issue that often only becomes evident only after clinical implementation (and typically after performance results have already been published), is poor generalisability within the original population. This may have occurred if the training data were erroneously encoded or lacked relevance to the clinical question. This underscores the importance of having domain expertise to critically evaluate which data should be included in model development. In summary, the development and implementation of ML tools face significant challenges due to the lack of standardisation in databases, variability in data quantity and quality, and difficulties in generalising results both across and within populations. This can be addressed through the creation of multicentre, and ideally multinational, databases to enable external validation and generalisation of models. This effort must be accompanied by strict attention to data protection and cybersecurity. The persistent gap between technological innovation and clinical translation A paper from 2014 by O’Sullivan [7] had already outlined both the potential benefits and the implementation challenges associated with clinical decision support systems, the most complex of which are ML tools. At that time, the publication highlighted that the medical sector lagged behind other industries, such as finance, in adopting these technologies. The primary barriers identified included the need for expert knowledge to ensure consistent application, difficulties with interoperability between software systems, lack of standardisation in data collection, challenges in customisation, the burdensome maintenance of bespoke or in-house programmes, limited time available to clinical innovators, and concerns surrounding the black box nature of many ML tools. Since then, significant technological advancements have transformed daily life through widespread digitisation, and AI tools have become ubiquitous in the public domain. Concurrently, a decade of productive academic research has yielded hundreds of medical models featuring more efficient, sophisticated, and interpretable architectures. However, despite this substantial progress in data and computer science, the meaningful integration of artificial intelligence into routine clinical anaesthetic practice remains limited. This enduring stagnation suggests that the primary obstacles may no longer be purely mathematical or technological, but rather institutional, regulatory, or operational, requiring a systematic evaluation to determine why medicine has failed to capitalise on these computational developments. Chokepoints Figure 6 illustrates what the authors identify as the main bottlenecks to the integration of ML tools in anaesthesia. In the authors' view, anaesthetists do not lack innovative ideas, rather, the main obstacles lie in limited time, insufficient technical expertise in mathematics or computer science, and a lack of collaboration with ML experts. As a result, relatively few ideas are converted into factual research projects. Chokepoints in innovation. Legend: chokepoints in the integration of ML tool in the practice of aneasthesiology Once a concept has been formulated, it must be developed into a structured research project, a step often hindered by the data-related issues described earlier. Additional logistical and infrastructure challenges arise. Data must be securely stored and protected, raising concerns about cybersecurity and incurring costs that institutions may be unwilling or unable to bear. The next phase is publication, the cornerstone of scientific progress. However, this presents difficulties, including the well-documented issue of publication bias, with a propensity to favour positive results, whether genuine or overstated. Moreover, the time and effort required to publish research can be substantial, especially when it is balanced with demanding clinical responsibilities. Peer reviewers may also lack familiarity with ML methodologies, further complicating the publication process. A possible solution can be found in the growing adoption of explainable AI (XAI). This concept promotes the use of algorithms with transparent structures (so-called white-box models, as opposed to black-box models) where the internal workings are opaque to the user. Alternatively, XAI may involve supplementary algorithms designed to explain how the primary model reaches its conclusions. Such approaches can enhance user trust, facilitate the understanding of errors, and enable meaningful feedback to developers [1, 8]. For these reasons, explainable AI should be considered essential in the application of AI in medicine. Finally, the greatest bottleneck is the integration of these tools into everyday clinical practice. Crucially, the deployment of AI-powered systems frequently encounters passive or active resistance among physicians; this professional reluctance is often rooted in automation bias, where clinicians either over-rely on or completely reject algorithmic suggestions, alongside a perceived threat to clinical autonomy and a fear of skill degradation [9]. Furthermore, navigating highly divergent regional regulatory frameworks, such as varying data handling policies and stringent privacy regulations (e.g., GDPR or HIPAA), imposes a substantial compliance burden that severely restricts international collaboration and multisite tool deployment [9, 10]. While AI-powered tools hold the potential to democratise expertise and mitigate regional imbalances in underserved areas, they also risk widening the digital divide, as resource-limited regions frequently lack the infrastructural readiness and high-quality representative data required to implement such sophisticated technologies [8, 10]. As demonstrated in this work, while there is an abundance of proposed applications in the literature, very few have successfully transitioned into routine clinical use. This translation gap is multifactorial and may be attributed to several distinct barriers: limited commercial interest from the medical industry, a lack of confidence among end-users, insufficient external clinical validation, and a fundamental mismatch between the technical capabilities of the developed tools and the practical realities of daily clinical workflows. The peak of the innovation hype cycle? "AI" has already gone through two periods of "winter" for various reasons that are outside the scope of this study, after following the classical Gartner Hype Cycle. There are two potential perspectives to consider: a pessimistic one, viewing us as being at the peak of hype just before disillusionment, or a more optimistic view that sees hype as progressing linearly alongside productivity. This perspective may depend on the actual results of widely integrating AI into our practice. We do not need just another screen (we already have too many). An ergonomic workspace should ideally include one integrated screen. Additionally, we should move beyond relying on local algorithms designed for a very specific population with a rigid, black-box architecture. What we truly need are tools that meet our field requirements, are intelligently integrated into our computing systems, and are easy to maintain and adjust for different populations. The goal is to streamline our daily practices and genuinely enhance patient outcomes. As recommended in a paper by Maheshwari et al. [11], critical implementation must be supported by education, interdisciplinary collaboration, data sharing, multicentre validation, and translational research. Technological innovation in itself has no significance unless it is integrated into a cohesive global system that enables its potential to be realised. To achieve this goal, we must now include them in a strategic vision of our profession, through material and human investment specifically geared to integrate this into our daily practices and applying the various recommendations for the successful integration of those tools [9, 11]. The ideal AI tool It is a comprehensive system capable of performing multiple tasks, including prediction and optimisation. It should be intuitive for anaesthetists to use, with results that are transparent and explainable (functioning as a “white box”). Its outputs should provide added value compared with conventional scoring systems (“useful”). The ideal AI should allow easy retraining on new populations, demonstrate robustness against noise or irrelevant data, and operate efficiently with seamless, ergonomic integration into operating room computer systems. Finally, it must be free from ethnic or demographic bias. An AI of this capacity is on the border between a dream in terms of usefulness, and a nightmare. Some optimistic papers [10, 12] predict that these tools will not replace clinical practice but will help the decision-making process and increase its efficiency, while emphasising the importance of using them intelligently. However, with integrated, overpowered and omniscient AIs, we might wonder what an anaesthetist could still be useful for. Limitations Our study has several limitations. Firstly, it is based solely on published research, which presents a publication bias. This is relevant given that projects yielding negative or unexpected results, which the authors of this work experienced directly, are less likely to be published. As a consequence, researchers may unknowingly duplicate previous efforts, thereby wasting time and resources on developments that have already been explored but not reported. Alternatively, it may provide an unduly optimistic overall impression of the current capabilities of ML by publishing only positive results. Second, we acknowledge the potential presence of adherence bias. Both authors, while not uncritical enthusiasts, are nonetheless convinced of the value of machine learning in their daily clinical practice. Although the literature reflects an ongoing debate on the role of ML in anaesthesia, which fosters creative and constructive discourse, our perspective may be subtly influenced by our professional engagement with the field. A further limitation lies in the possibility of novelty bias. As with many emerging technologies that generate curiosity among anaesthetists, including ourselves, there is a tendency to become overly enthusiastic about innovations. While machine learning is by no means a new concept, its recent clinical applications are, and this calls for a measured and critical approach when assessing its potential. Finally, we have made every effort to avoid one-sided referencing. We remain critically aware of the limitations and potential pitfalls of machine learning, recognising that unmet expectations could undermine trust in future ML tools. Our aim is to contribute to a balanced and rigorous discussion that supports the responsible integration of these technologies into anaesthetic practices. Conclusion This systematic review underscores a significant paradox in modern anaesthesia: despite a decade of prolific academic research and tool development, the clinical translation of machine learning (ML) remains critically limited. While the potential of ML to enhance workflow efficiency and patient safety is widely acknowledged, innovation alone has proven insufficient to disrupt routine practice. An abundance of theoretical models characterises the current landscape, but a severe deficit in full clinical validation and integration, leaving anaesthesia notably behind some other medical specialities. To bridge this gap between development and clinical reality, a fundamental shift in strategy is required. Future efforts must move beyond merely generating models for publication and focus on establishing robust implementation frameworks. This transition requires supporting the emergence of hybrid anaesthetist-developer profiles, fostering a clinical culture of beta-testing, and promoting widespread algorithmic literacy among practitioners. Furthermore, better exploitation of the vast, underutilised data generated daily in operating theatres, while strictly addressing inherent big data biases, is essential. Ultimately, without a deliberate, multidisciplinary effort to overcome these systemic and educational barriers, machine learning will remain a theoretical promise rather than a practice-changing force in anaesthesia. Supplementary Information Supplementary Material 1. Acknowledgements I would like to thank my father for his invaluable technical help and my wife and family for their support. Registration Not Applicable (literature review). Clinical trial number Not Applicable (literature review). Abbreviations Artificial Intelligence Food and Drug Administration Machine Learning Millions of instructions per second Large Language Models Authors’ contributions RF determined the study concept and the research question. The methodology was discussed collaboratively between RF and PD. The collection of abstracts was carried out by RF, followed by the first selection. PD conducted a second selection reading. RF produced the statistics as well as the figures. RF wrote the article, which was reviewed and completed by PD. Funding No external funding. Data availability Tables of data available through the supplementary material. Declarations Ethics approval and consent to participate Not Applicable (literature review). Consent for publication Not Applicable (literature review). Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement Tables of data available through the supplementary material.