Bioethical Considerations of Deploying Artificial Intelligence in Clinical Orthopedic Settings: A Narrative Review

Abstract

Artificial intelligence (AI) has emerged in orthopedics with the potential to improve diagnostic accuracy, optimize surgical workflows, and support personalized care. We conducted a narrative review exploring the bioethical considerations of AI use in the orthopedic clinical setting, focusing on 4 core principles—autonomy, beneficence, nonmaleficence, and justice—to provide orthopedists with a practical framework for AI’s implementation. We utilized the Preferred Reporting Items for Systematic Reviews and Meta-Analyses framework to conduct a comprehensive PubMed search; 89 articles were evaluated and 23 met our inclusion criteria. Across these studies, bioethical considerations for the clinical implementation of AI tools consistently emerged, most commonly concerning privacy, bias, transparency, informed consent, and regulation. We offer recommendations for strengthening privacy safeguards, adopting bias mitigation strategies, improving transparency through explainable AI tools, and establishing clear regulatory frameworks with lifecycle evaluation.

Introduction

Artificial intelligence (AI) refers to the study of algorithms that enable machines to perform human-like cognitive functions such as problem-solving, decision making, and learning from experience [6,16]. In recent years, the capabilities of AI have expanded rapidly, driven by advancements in computing power, data availability, and algorithmic innovation. Broadly, AI can be categorized into non-generative and generative types. Non-generative AI is task-specific, focusing on discrete functions for which it has been trained. This includes branches of AI such as machine learning (ML), which applies statistical algorithms to optimize predictions from datasets. ML offers the advantage of continual improvement, as its performance evolves with the addition of new data. In contrast, generative AI is predictive and capable of uncovering patterns or determining outcomes based on input instructions. A prominent example of generative AI is large language models (LLMs), such as ChatGPT, which analyze vast textual datasets using neural networks to generate coherent, human-like responses [30]. Deep learning (DL), a specialized subset of ML, can be used for both non-generative and generative purposes and further enhances data analysis by assigning weights to specific inputs for nuanced decision making [16].

The capacity for AI to analyze big data—large, intricate datasets that challenge conventional processing methods—positions it as a significant tool in healthcare. By uncovering patterns and associations within these datasets, AI has the potential to inform clinical decision making and guide patient care. As data collection becomes increasingly sophisticated, AI demonstrates remarkable promise in healthcare applications.

Within orthopedics, AI is poised to transform healthcare delivery [1]. DL algorithms can identify imaging features such as fracture types [27] and anatomical variation [34], facilitating accurate diagnosis and treatment planning. AI systems have also demonstrated the ability to predict critical clinical outcomes, including those related to spine surgery and arthroplasty, as well as postoperative complications and patient-reported outcome measures [13,19,20]. In the operating room, AI shows promise in workflow optimization, phase and gesture recognition, knot tying, and bone fragment recognition and tracking [6]. Moreover, the language-processing capabilities of AI are fostering clearer and more effective interactions between patients and providers [31]. Collectively, these advancements pave the way for more personalized, patient-centered care.

Despite its potential, AI has significant limitations. While pilot studies often demonstrate improved outcomes, the broader implementation of AI in clinical settings is constrained by underdeveloped regulatory and ethical frameworks. Concerns regarding accuracy, privacy, and bias persist, particularly given AI’s rapid technological development. In early 2025, the United States also introduced sweeping changes within the AI regulatory landscape in favor of accelerated AI innovation and looser mandates on AI governance [15]. Similarly, regulatory bodies in other regions, including Europe, are also shifting toward more innovation-friendly approaches [14]. Understanding how we can balance the drive for innovation with the imperative of patient safety is essential to ensuring that AI is deployed in medicine responsibly and equitably.

This review aims to equip orthopedists with a principles-based bioethical framework for integrating AI into clinical practice. By exploring the ethical dimensions of AI use, this discussion seeks to balance innovation with patient safety and to foster a culture of trust, transparency, and accountability in orthopedic care.

Methods

This review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses framework to ensure a structured process for identifying and analyzing relevant literature [28]. A search of the PubMed database was conducted on December 21, 2024, using predefined keyword combinations related to AI and ethical considerations in orthopedic surgery (Table 1). The search strategy incorporated Boolean operators and synonyms for key terms to capture a broad range of studies.

Table 1.

PubMed search terms.

Search strategy

Results

(“Artificial Intelligence” [Mesh] OR “Machine Learning” [Mesh] OR “AI” [Title/Abstract] OR “machine learning” [Title/Abstract] OR “deep learning” [Title/Abstract] OR “Large language model” [Title/Abstract] OR “LLM” [Title/Abstract]) AND (“Orthopedics” [Mesh] OR “Musculoskeletal Diseases” [Mesh] OR “orthopedic” [Title/Abstract] OR “musculoskeletal” [Title/Abstract] OR “Orthopaedic” [Title/Abstract] OR “Orthopaedics” [Title/Abstract]) AND (“Ethics” [Mesh] OR “Bioethics” [Mesh] OR “ethical” [Title/Abstract] OR “ethics” [Title/Abstract] OR “equity” [Title/Abstract] OR “justice” [Title/Abstract] OR “autonomy” [Title/Abstract] OR “accountability” [Title/Abstract] OR “beneficence” [Title/Abstract] OR “nonmaleficence” [Title/Abstract] OR “ethic* principles” [Title/Abstract])

89

Studies were included if they reported on the use of AI or related computational strategies within orthopedic surgery or musculoskeletal care and addressed ethical considerations in the application of these tools. Any study design was eligible for inclusion to ensure a wide range of perspectives, and only articles written in English were considered. We excluded studies on specialties other than orthopedic surgery; robotics or other physical machines without reference to AI; animal models; or research methodology, physician training, or education rather than clinical practice.

Two researchers (L.V.K. and S.D.K.) independently screened the titles and abstracts of retrieved studies for inclusion. Full texts of potentially eligible studies were then reviewed to confirm eligibility. Relevant data were extracted collaboratively by the 2 reviewers, who resolved any disagreements through discussion. To ensure consistency, reviewers were trained in the inclusion and exclusion criteria prior to initiating the study selection process.

The search yielded 89 articles. After screening, applying inclusion and exclusion criteria, and reviewing full texts, 23 papers were included for analysis (Fig. 1 and Table 2).

Fig. 1.

PRISMA flow diagram. PRISMA Preferred Reporting Items for Systematic Reviews and Meta-Analyses.

Table 2.

Summary of findings.

Author (year)	Type of study	Key findings
Myers et al (2020) [25]	Review	Gives examples of AI in orthopedics across AI subtypes (ML, DL, NLP, the internet of things); describes ethical (privacy, bias, risk of errors, exploitation, regulation) and clinical (overreliance on AI, reduction of clinical skills) considerations
Rizk et al (2024) [33]	Review	Explores ML use in orthopedic oncology (diagnostic imaging, histologic evaluation, survival prediction, molecular analyses) and ethical concerns (privacy, consent, bias)
Carroll et al (2024) [7]	Review	Guides orthopedic surgeons in prompt engineering for LLMs to generate reliable clinical outputs; describes ethical considerations (privacy, data protection, bias, fairness, accountability, professional autonomy, patient confidentiality, informed consent)
Giorgino et al (2023) [11]	Narrative review	Explores diverse applications of ChatGPT in orthopedics (diagnosis, treatment planning, patient education, research); describes ethics concerns (data privacy, security, accountability, validation, potential biases)
Vaishya et al (2024) [37]	Scoping review	Reviews the use of ML and DL in regenerative orthopedics; includes ethical considerations (potential biases, fair access to AI); notes the issues of transparency and trust, encourages clear guidelines and regulation
Papalia et al (2024) [29]	Systematic review	Evaluates AI use in detection, management, and prognosis of bone metastases in orthopedic oncology; includes analysis of ethical considerations (transparency and risk of bias, need to ensure AI systems are trained with diverse datasets)
Chang et al (2020) [8]	Narrative review	Reviews ML applications in spine surgery (predicting postoperative outcomes, computer vision, computer-assisted navigation, and other intraoperative support tools); discusses the black box of AI, bias in dataset training, insensitivity to impact, and reward hacking (machines cheating the system to achieve an outcome)
Karaibrahimoglu et al (2024) [17]	Scoping review	Describes AI use in telehealth for work-related musculoskeletal disorders; notes ethical themes (equitable access, quality care, professional liability, accountability, patient privacy, informed consent, autonomy, data protection, confidentiality)
Zsidai et al (2023) [39]	Review	Describes ways AI systems can enhance orthopedic research by providing an overview of existing AI applications in medicine; discusses ethical considerations (black box decision making, transparency, risk of bias)
Banatwala et al (2024) [3]	Narrative review	Explores applications of AI in orthopedics in low- and middle-income countries (fracture detection, spine imaging, bone tumors, joint surgery, identifying disabilities, surgical training) and ethical concerns (patient privacy, consent, data safety, data monitoring, bias, accountability)
Gossec et al (2020) [12]	Guideline	The EULAR addresses the use of big data through AI and ML in musculoskeletal diseases and ethical challenges (privacy, confidentiality, security, data ownership, data minimalization, flow of data)
Birkhoff et al (2021) [6]	Systematic review	Reviews AI use in the OR (workflow prediction, surgical phase recognition, knot tying, registration, and bone tracking); most uses show improved performance but lack regulatory systems for development and ethical use (especially on a large scale)
Huffman et al (2024) [16]	Review	Reviews AI applications in orthopedics (personalized prediction of patient outcomes, diagnostic and imaging analysis, predicting resource use); describes ethical concerns (patient safety, need for transparency on when and how models work, evaluation of performance, possible spectrum bias)
Morya et al (2024) [23]	Review	Describes ChatGPT use in orthopedics (diagnostics, processing large volumes of information, preoperative and rehabilitation planning, communication, patient education), plus its limitations (biases, inaccuracies, lack of transparency, inability to do higher-order judgment or logical reasoning); recommends ethical, legal interventions
Pressman et al (2024) [30]	Systematic review	Identifies ethical concerns of AI use in surgery and how the 4 ethical principles are represented in them; defines how common concerns with AI (accuracy, bias, patient confidentiality, responsibility) fit within the bioethics framework
Siddiqui et al (2024) [35]	Retrospective study	Examines biases in AI algorithms for segmentation of hip and knee bony anatomy on plain X-rays, proposing an automated hip/knee segmentation system, and found that a balance and group-specific modeling approach had the greatest value for fairness metrics; proposes strategies to reduce bias in imaging algorithms (training models on diverse datasets, doing longitudinal outcome studies)
Lans et al (2023) [21]	Systematic review	Examines whether prognostic ML models for orthopedic surgery outcomes account for SDOH, with most models reporting 1+ SDOH (usually age or gender/sex) in model development and a minority including race/ethnicity; concludes that ML models have bias and should be used clinically with caution
Oettl et al (2024) [26]	Review	Considers the use of AI in research, with guidelines on uses (classification, regression, clustering, text/image/video generation) and ethical concerns with each (eg, how to evaluate AI outputs); concludes that evaluating these systems is a continuous process
Musat et al (2024) [24]	Review	Examines AI use in predicting and preventing sports injuries (analyze large datasets, identify patterns that signify potential injuries, create personalized training, and recovery plans); lists ethical challenges of AI (data privacy, interpretability, data handling guidelines)
Taylor et al (2024) [36]	Cross-sectional study	Assesses accuracy and reliability of ChatGPT-generated responses to FAQs about TKA; concludes that it can provide accurate answers but concerns remain (inaccuracy, tendency to “hallucinate”)
Khosravi et al (2024) [18]	Retrospective study	Uses generative DL technology to examine radiographic differences based on race in patients undergoing TKA; demonstrates the importance of identifying racial differences and discusses how understanding race-based differences can help uncover disparities in datasets and models that could lead to biases in AI
Balch et al (2024) [2]	Scoping review	Describes the use of ML to predict PROMs for incapacitated patients and the ethical concerns of this use of ML (autonomy, trust, conflicting outputs in different models); argues that ML should complement but not replace provider/surrogate decision making
Chester and Mandler (2025) [9]	Comparative study	Explores the reliability of ChatGPT by assessing its agreement with orthopedics consensus statements; discusses ethical limitations of clinical use (accountability for decision making, performance)

AI Artificial intelligence, DL deep learning, EULAR European League Against Rheumatism, LLMs large language models, ML machine learning, NLP natural language processing, OR operating room, PROMs patient-reported outcome measures, SDOH social determinants of health, TKA total knee arthroplasty.

Results

Of the 23 studies analyzed, publication dates ranged from 2020 to 2025, with most published in 2024. Most were narrative, systematic, or scoping reviews. The applications of AI, including ML and DL, in orthopedic surgery were diverse, ranging from clinical decision making, such as diagnosis and treatment planning, to preoperative preparation, patient education, outcome prediction, and regenerative orthopedics. Across studies, several ethical concerns consistently emerged, including privacy, bias, confidentiality, transparency, and regulation (Fig. 2).

Fig. 2.

Framing of commonly cited ethical concerns within the core bioethical principles.

Privacy and Data Protection

The development of AI-based algorithms requires the collection, processing, and validation of large datasets, often containing sensitive patient information. This presents significant concerns for privacy and data protection, as multiple studies highlighted [3,7,11,17,24,25,30,33,37]. One instance involved a 2015 agreement between DeepMind, a Google-owned British software company, and the UK’s National Health Service (NHS), in which patient data from 1.6 million individuals was improperly shared, violating the Data Protection Act [4,25]. As AI use becomes widespread, authors like Rizk et al have urged for careful handling of data, tight cybersecurity, and robust safeguards to protect patient information [33]. Karaibrahimoglu et al recommended implementing security measures like maintaining access controls, employing data encryption, creating safe storage procedures, preserving data integrity, and adhering to relevant privacy laws [17]. Gossec et al introduced the concept of Privacy by Design (PbD), advocating for privacy strategies at every stage of data handling, from collection to analysis [12].

Risk of Bias

Bias was a pervasive concern, identified in 16 of the 23 articles reviewed [3,7,8,11,16,18,21,23,25,29,30,33,35 -37,39]. The principle of “garbage in, garbage out,” as illustrated by Giorgino et al, underscores how biased or unrepresentative training data can propagate errors in AI systems [11]. Myers et al noted that much of the training data used in the United States is based on the medical records of white men, leading to reduced accuracy for women, ethnic minorities, and other underrepresented groups [25].

Chang et al introduced the term “distributional shift” to describe the training of datasets that may be outdated or biased toward socioeconomic status or race [8]. Both Carroll et al and Pressman et al elaborated that biased algorithms perpetuate inequities and discriminatory practices, as well as amplify existing inaccuracies and suppress new opinions, thus slowing rather than advancing developments [7,30]. This concern is particularly critical in orthopedics, where social determinants of health (SDOH) are rarely included in training datasets, as highlighted by Lans et al [21]. Efforts to mitigate bias included work by Siddiqui et al, who implemented trade-offs between fairness and performance in AI imaging tools [35], and Khosravi et al, who used generative models to expose and address racial disparities in imaging datasets, particularly among African American patients with osteoarthritis [18]. Spectrum bias, or differences in patient populations between primary and tertiary care settings, was also flagged as a potential issue by Huffman et al [16].

Informed Consent and Confidentiality

Both Karaibrahimoglu et al and Banatwala et al emphasized the importance of maintaining patient autonomy through informed consent and confidentiality [3,17]. Carroll et al described the importance of informed consent in maintaining patient autonomy and ensuring patients can make well-informed decisions regarding their treatment and use of their health data [7]. Challenges arise when determining when consent is necessary and how best to obtain it. Rizk et al noted that informed consent may be required for ML applications that directly influence treatment decisions but not for tools with minimal patient impact, such as triage imaging [33]. Pressman et al argued for comprehensive disclosure of risks and benefits when incorporating LLMs into patient care but noted that patients might not comprehend the implications of AI use, complicating the informed consent process [30]. Cases involving incapacitated patients, as described by Balch et al, further highlighted ethical ambiguities surrounding consent when deploying AI tools such as Patient Preference Predictors [2].

Transparency

Transparency challenges were noted in 8 studies, particularly the “black box” phenomenon, wherein the decision-making processes of AI systems are inaccessible to human understanding [8,16,23 -25,30,37,39]. In healthcare, this lack of explainability can undermine patient and provider trust.

Zsidai et al and Pressman et al highlighted the importance of explainable AI (XAI) tools, which provide insights into the decision-making processes of AI models [8,25,30,39]. This is significant in healthcare, where decision making depends upon providers’ ability to trace their logic through events and their outcomes, a process that is lost with a black box model [30,39]. Furthermore, mistrust due to the lack of transparency in AI-based decision making may hinder the adoption and application of AI technology [23,37]. To address this lack of trust, Musat et al recommended using methods such as XAI tools that make AI decision pathways more transparent [24]. These include tools such as Local Interpretable Model-Agnostic Explanations (LIME), developed by Ribeiro et al [32], and SHapley Additive exPlanations (SHAP), proposed by Lundberg and Lee [22]. Both methods generate explanations of AI outputs and can be used in a variety of contexts, as both LIME and SHAP can be applied to any ML model [22,32].

Regulation

Authors stressed the importance of regulation of AI systems in ensuring safety and trust while noting the uncertainties of when, how, and by whom AI should be regulated [3,6,11,16,26,30,37,39]. Myers et al recommended lifecycle-based regulatory frameworks that evaluate AI tools at every stage of development and implementation [25]. Lifecycle regulation was also supported by Oettl et al, who called for recurrent and interdisciplinary evaluation [26]. Pressman et al raised concerns about responsibility in the event of AI-driven errors. While some argue that physicians should bear ultimate responsibility, a lack of clear guidelines complicates accountability, underscoring the need for robust regulatory frameworks [30].

Other Ethical Considerations

Overreliance on AI could erode clinical judgment, as noted by Myers et al, Banatwala et al, and Carroll et al [3,7,25]. Further, there are limits to AI’s ability to employ nuanced judgment for complex tasks in orthopedic surgery, requiring physicians to preserve professional autonomy, remain vigilant and up to date on education and training, and consider AI as a complement to provider decision making rather than a replacement [2,7,17]. The phenomenon of AI hallucination, whereby models generate plausible but incorrect outputs, could result in misinformation [9,11,30,36]. If undetected, decisions based on AI hallucination, and thus inaccurate information, could put patients at serious risk [9,11]. Chang et al noted that machines may learn to cheat the system to achieve a particular outcome (reward hacking), again spreading misinformation [8]. They also note that AI tools are insensitive to impact, highlighting how these technologies may underestimate the consequences of predicting certain outcomes [8]. Finally, access to AI tools remains a significant equity issue, as concerns about fair distribution across diverse patient populations and inequitable access to resources, such as broadband internet, now carry even broader implications for accessibility [37].

Discussion

Since at least 1942, humans have been considering the ethical implications of AI/robotics. In that year, science fiction writer Isaac Asimov, in one of his short stories, put forth the “Three Laws of Robotics,” the first of which dictated that a robot should not harm a human through action or inaction [38]. Almost a century later, the principle of “do no harm” remains central to AI innovations. As our review highlights, ethical challenges are widely acknowledged, but discussions about them in the literature often lack depth, consistency, and actionable solutions.

We found in our review that privacy and data protection continue to be critical concerns. Breaches of sensitive patient information, such as the DeepMind-NHS case, illustrate how inadequate safeguards can undermine patient trust and violate autonomy [4,25]. Strategies such as PbD, as proposed by Gossec et al, provide a proactive approach to embedding privacy considerations into all stages of AI development [12]. Furthermore, adherence to established privacy regulations such as the Health Insurance Portability and Accountability Act (HIPAA), General Data Protection Regulation (GDPR), and Genetic Information Nondiscrimination Act is essential to protecting patient information and fostering trust [7,24,30].

Bias in AI models poses another significant ethical challenge. As highlighted in multiple studies, unrepresentative training datasets can perpetuate inequities and reduce the accuracy of AI systems for underrepresented groups, including women and ethnic minorities [8,25]. Efforts to mitigate bias, such as incorporating diverse datasets and addressing SDOH, align with the principle of justice and promote equitable outcomes [18,21]. In addition, techniques like those explored by Siddiqui et al, which balance fairness and performance in AI models, offer valuable insights into addressing this challenge [35].

Informed consent and transparency are closely linked and essential for maintaining patient autonomy. However, our review identifies significant ambiguities regarding when and how consent should be obtained, particularly for applications with minimal direct patient impact [33]. Pressman et al’s emphasis on XAI tools offers a practical solution for bridging the gap between algorithmic complexity and patient or clinician understanding [24,30]. Improving transparency through XAI can not only enhance trust but also ensure accountability in clinical decision making.

The issue of regulation underscores the need for a lifecycle-based approach, as proposed by Myers et al, to evaluate AI tools at every stage of their development and use [25]. Clear guidelines on accountability, particularly in cases of AI-driven errors, remain a critical gap [30]. Addressing these regulatory challenges is essential to creating a robust framework that ensures safety and trust.

Emerging considerations, such as AI hallucination and reward hacking, highlight the challenge of misinformation. For instance, Chang et al describe how reward hacking can result in algorithms prioritizing metrics at the expense of accuracy or relevance [8]. These risks underscore the importance of maintaining clinician oversight and professional autonomy, ensuring that AI complements rather than replaces human judgment [2,7,17].

The 4 moral principles introduced by Beauchamp and Childress—autonomy, beneficence, nonmaleficence, and justice—remain relevant and provide a structured framework for addressing the ethical implications of AI in orthopedics [5]. Although some authors, such as Floridi and Cowls, advocate for tech-specific frameworks, their analysis reveals that most principles converge on these core bioethical values [10]. This suggests that adapting these principles to contemporary needs can effectively guide the ethical deployment of AI while maintaining alignment with established healthcare values.

Practical Recommendations

To advance ethical AI integration in orthopedics, the following steps are recommended:

1. Strengthen privacy protections: Incorporate PbD strategies and ensure compliance with privacy regulations such as HIPAA and GDPR during AI development.

2. Mitigate bias: Use diverse and representative datasets while integrating SDOH to promote equitable outcomes.

3. Enhance transparency: Employ XAI tools to improve the interpretability of AI systems for clinicians and patients.

4. Clarify informed consent: Develop clear guidelines for obtaining consent tailored to the context of specific AI applications.

5. Adopt lifecycle regulation: Implement continuous, interdisciplinary evaluations of AI systems to ensure safety and adaptability over time.

6. Address emerging risks: Establish safeguards to mitigate AI hallucination and reward hacking, including using high-quality data sets and maintaining close clinician oversight.

As AI continues to evolve amid shifting regulatory landscapes, embracing these recommendations and grounding them in a bioethical framework will support meaningful and consistent ethical discourse.

Limitations

This review has several limitations. First, we relied solely on PubMed for literature retrieval, and relevant studies may have been excluded from disciplines such as computer science, engineering, or social sciences that are critical to understanding AI use. Thus, we recommend that future reviews incorporate databases such as IEEE Xplore, Scopus, or Web of Science. Second, a narrative review is suitable for summarizing diverse literature but lacks the methodological rigor of a systematic review, such as applying inclusion and exclusion criteria, quality appraisals, and comprehensive search strategies that reduce potential bias in study selection. Third, the heterogeneity of included studies makes direct comparisons challenging. Finally, AI technologies and regulatory frameworks are rapidly evolving, and so some of the findings of our review may be less applicable over time. Continuous reassessment of the ethical landscape will be crucial.

Conclusion

The integration of AI into orthopedics promises transformative advancement while posing complex ethical challenges, including issues related to privacy, bias, informed consent, transparency, and regulation. This review underscores the importance of addressing these challenges proactively, guided by the bioethical principles of autonomy, beneficence, nonmaleficence, and justice.

Key recommendations for future efforts include strengthening privacy safeguards, adopting bias mitigation strategies, improving transparency through XAI tools, and establishing clear regulatory frameworks with lifecycle evaluation. In addition, prioritizing equity and access will be essential to ensure that AI technologies benefit all patient populations, particularly those from underserved or marginalized communities.

As the field of orthopedics continues to embrace AI, it is imperative to maintain rigorous ethical standards and foster interdisciplinary collaboration among clinicians, AI developers, ethicists, and policymakers. By doing so, we can harness the potential of AI to enhance patient care while addressing its risks responsibly and equitably.