Educational Frameworks for Diagnostic Decision-Making in AI-Enhanced Head and Neck Pathology.

Educational Frameworks for Diagnostic Decision-Making in AI-Enhanced Head and Neck Pathology - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Head Neck Pathol . 2026 Jun 4;20(1):60. doi: 10.1007/s12105-026-01909-z Educational Frameworks for Diagnostic Decision-Making in AI-Enhanced Head and Neck Pathology Tiffany Tavares 1, ✉ , Linda Sangalli 2 , Reshma S Menon 3 , Craig S Miller 4 , Guy Hochman 5 Author information Article notes Copyright and License information 1 Department of Comprehensive Dentistry, School of Dentistry, University of Texas at San Antonio, San Antonio, TX USA 2 College of Dental Medicine – Illinois, Midwestern University, Downers Grove, IL USA 3 Tufts University School of Dental Medicine, Boston, MA USA 4 Department of Oral Health Practice, College of Dentistry, University of Kentucky, Lexington, KY USA 5 Barch Ivcher School of Psychology, Reichman University, Herzliya, Israel ✉ Corresponding author. Received 2025 Dec 22; Accepted 2026 Mar 14; Collection date 2026 Dec. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13237305  PMID: 42240861 Abstract Context This paper examines how artificial intelligence (AI) is reshaping diagnostic decision-making in academic and clinical head and neck pathology (HNP) and argues for the redefinition of post-doctoral education in the age of AI. Perspective While AI promises enhanced diagnostic accuracy, reproducibility, and efficiency, its integration also amplifies human cognitive biases and introduces new forms of error. Here we build on the 2019 Accreditation Council for Graduate Medical Education (ACGME) pathology milestones to provide an educational framework that prepares clinicians to engage critically and responsibly with AI-assisted diagnostics. We emphasize the importance of integration of insights from decision science, cognitive psychology, and medical education with recent applications of AI in HNP, including histopathology, radiology, and multi-omics modeling. By mapping how human reasoning processes interact with algorithmic systems, we identify cognitive, ethical, and institutional vulnerabilities that arise in hybrid human–machine decision environments. Our analysis highlights that AI does not eliminate diagnostic uncertainty but redistributes it. Automation bias, overconfidence, and anchoring may distort reasoning when clinicians fail to question algorithmic outputs. To address these challenges, we outline a model of healthcare-optimization training, centered on three core skills: AI literacy, diagnostic skepticism, and ethical transparency. As such, reframing decision-making education in HNP requires moving beyond technical training toward reflective, ethically grounded practice. Reccomendation Embedding AI literacy and critical reasoning within competency-based curricula and institutional governance can ensure that AI enhances rather than replaces human judgment—fostering safer, more equitable, and patient-centered diagnostic care. Keywords: Medical education, Cognitive bias, AI literacy Introduction The rise of artificial intelligence (AI) in diagnostic medicine is rapidly reshaping clinical reasoning. AI tools promise enhanced accuracy, reproducibility, and scalability, particularly in histopathology, radiologic triage, and molecular profiling. Thus, they offer transformative potential for diagnostic and prognostic workflows [ 1 , 2 ]. However, the potential benefits of AI also raise new challenges. AI does not eliminate diagnostic error; it reframes where and how such errors occur. Integration of algorithmic reasoning into clinical workflows also introduces new layers of decision-making that blends human judgment with probabilistic machine outputs [ 3 , 4 ], which depend on technical sophistication as well as the clinician’s interpretive and critical reasoning skills [ 5 ]. As such, experts are recognizing the need for updated competencies in pathology and a new entrustable professional activity (EPA) framework that will help prepare pathologists for AI-supported diagnostics [ 6 ]. In addition, insights provided here into decision science and cognitive psychology support the need for a distinct framework, grounded not only in technical readiness but also in the psychology of human–machine interaction, diagnostic biases, and reflective clinical judgment. In the following sections, we discuss how HNP providers and technological factors interact and sometimes conflict, supporting redefinition of decision-making education in the age of AI. As such, we propose an educational framework to guide AI integration into diagnostic education, emphasizing interpretive literacy, ethical reasoning, and reflective practice. This approach aims to align the promise of AI with the demands of safe, deliberate, and patient-centered care. Applications and Promise of AI in HNP The rapid integration of AI into HNP has generated substantial enthusiasm. AI applications are now being integrated across diverse domains, including histopathology, radiology, molecular profiling, clinical decision support, and electronic health record (EHR) data mining [ 7 ]. By combining computational precision with complex diagnostic data, AI promises to enhance accuracy, reproducibility, and efficiency across the diagnostic continuum. Its use, however, can also introduce notable cognitive challenges for clinicians. In the hierarchical structure of technologies utilizing AI, large learning models may be further classified into generative and non-generative models, based on the task and output data. Some well-known applications of generative AI models include text generators for the purpose of report generation, image generators for clinical, research, and educational purposes, multi-modal models to integrate various data types and multi-agent frameworks utilizing dialogue between different AI models. [ 8 ] High rates of success have been reported with both tabular and text data analysis, with the emergence of generative and non-generative AI models in pathology and medicine; although, their use-cases are still restricted, owing to limited datasets. Multimodal learning is particularly propitious, being able to extract data from diverse datasets, including clinical information, radiographic and histopathology images, and even “omics” data, when available, which can help physicians readily integrate these and expedite final comprehensive report generation. [ 9 ] Non-generative AI models are more well-established and help with predictive tasks, relying on patterns discovered from prior input data and are classified into supervised, unsupervised and reinforcement learning. Supervised models require explicit labeling by expert annotators, whereas unsupervised models learn directly from raw data without such annotation. An example of a supervised model is a convolutional neural network that uses deep learning (DL) to classify histopathology grading of tumors through whole slide image (WSI) analysis. Digital pathology uses this technology, whereby glass slides can be scanned, then studied and analyzed repeatedly via AI models. On the other hand, unsupervised machine learning (ML) in pathology, has played a pivotal role in the discovery of novel findings. Vision transformer architectures trained via self-supervised learning approaches have gained traction as powerful pretrained models that learn transferable image representations, enabling unsupervised or weakly supervised analysis of pathology images without the need for exhaustive expert annotation. These models can be used for biomarker discovery and can be adopted in immunohistochemical analysis or applied to targeted therapeutics by learning patterns in unlabeled data. [ 10 ] Recent data compared certified AI applications to detect metastases and was validated to be extremely capable across various cancer types (breast cancer, melanoma and head and neck cancer), supporting the broader application of AI-enabled digital pathology in routine practice for staging disease [ 11 ]. Similarly, a DL model based on vision transformer (a model that uses self-attention to establish context from image sequences) demonstrated excellent performance in classifying common salivary gland tumor subtypes with WSI; however these models can foster automation bias when analyzing rare and malignant subtypes due to underrepresentation in training data [ 12 ]. Likewise, AI based probabilistic outputs in thyroid nodule evaluation can support risk stratification and prognostication, but may also induce overconfidence. [ 13 , 14 ] Deep learning models have shown promising results in recognizing and quantifying intricate patterns within histopathologic images, overcoming the limitations of classical imaging analysis [ 9 ] For example, an AI-derived intraepithelial lymphocyte (IEL) score outperformed both binary and WHO grading systems in predicting malignant transformation reducing inter- and intra-observer variability; however, these models can create an anchoring bias, leading to overreliance on AI-generated scores. [ 15 , 16 ] Some highly sought-out applications of AI in the field of medical education include augmented and virtual reality tools to help train the next generation of healthcare teams. This momentum is further reflected in recent proposals to introduce EPA in medical education, particularly in diagnostic pathology, embracing what has worked well in image-based medical specialties such as radiology. [ 17 ] Overall, these applications demonstrate that AI can serve as a powerful augmentative tool in HNP, extending interpretive, analytical, prognostic and integrative capabilities beyond traditional manual analysis while it simultaneously may bias clinical reasoning and decision-making, requiring careful implementation and oversight. The Human Side of Diagnostic Reasoning Clinical decision-making in HNP is inherently complex, [ 18 , 19 ] with high inter-observer variability and frequent misclassification in oral and laryngeal lesions. [ 20 – 22 ] Thus, formal understanding of the cognitive mechanisms that underlie professional judgment is particularly important in HNP. It should be a central component of post-doctoral education, where pathologists are expected to transition from competent practitioners to independent experts who supervise, mentor, and critically evaluate diagnostic decisions. The integration of AI into pathology requires clinicians who can interpret, question, and contextualize machine outputs. To achieve this goal, it is essential to understand how humans make diagnostic decisions: what cognitive processes guide them, where they excel, and where they are susceptible to systematic error. Traditionally, decision-making in medicine is taught under the dual-system account. According to this view, reasoning relies on two independent modes of thought: System 1, which is fast, intuitive, automatic, and associative, and System 2, which is slow, deliberate, and rule based. System 1 generates immediate impressions and heuristic judgments, often based on salient but limited cues, while System 2 monitors, verifies, or overrides these intuitive outputs through effortful analysis [ 23 , 24 ]. This framework has been highly influential in medical and dental education, as it offers an accessible language for distinguishing between intuitive “pattern recognition” and reflective “analytic reasoning” in clinical contexts. In pathology, for example, intuitive processes may accelerate the recognition of familiar patterns, while deliberate processes enable careful evaluation of competing diagnostic possibilities. [ 22 , 25 ] Although the dual-system framework provides a useful general conceptualization, recent developments in decision science challenge its overly simplistic dichotomous nature. Clinical judgments rarely conform to a strict separation of “fast and biased” versus “slow and accurate.” Instead, they emerge from a dynamic interaction between processing modes, task features, and contextual constraints [ 26 , 27 ]. For example, decision quality depends on the compatibility between the thinking style and the task requirements [ 26 , 28 ]. Intuition can sometimes lead to more accurate and consistent outcomes than deliberation, while deliberate reasoning may introduce biases when it is misaligned with context. These findings underscore that both novice and experienced pathologists are susceptible to systematic error, regardless of whether they approach reasoning intuitively or analytically. The complexity of diagnostic reasoning in HNP is further illustrated by the systematic biases that have been documented across clinical contexts [ 29 ]. These biases are not restricted to novices; indeed, research suggests that seasoned experts are often less aware of their own cognitive vulnerabilities and therefore more likely to perpetuate them. [ 30 , 32 ] Collectively, these tendencies underscore the need for a more nuanced understanding of judgment processes, one that transcends the simplistic assumption that expertise or deliberate reasoning guarantees accuracy [ 27 ]. As diagnostic decisions reflect a dynamic interplay of cognitive systems, task demands, and contextual factors, clinicians are vulnerable to predictable judgmental errors and variability. Importantly, the very same mechanisms that shape human judgment also determine how pathologists may engage with AI: whether they adopt its recommendations uncritically, resist them completely, or integrate them in a balanced and reflective manner [ 33 – 35 ]. Thus, appreciating the cognitive foundations of diagnostic reasoning is a prerequisite for critically assessing AI tools, deciding when and how to employ them, and avoiding technology-amplified biases. This understanding also underscores the educational imperative: post-doctoral training must prepare pathologists not only to master diagnostic content but also to recognize, monitor, and manage the cognitive processes that shape their decisions. Limitations and Risks of AI in HNP Despite its promise, the clinical implementation of AI in HNP remains constrained by significant technical, infrastructural, and governance challenges. Many models perform well in research settings, yet falter in real-world deployment where data heterogeneity, workflow complexity, and regulatory oversight play decisive roles. [ 36 ] The reliability of AI depends heavily on the quality and representativeness of the training datasets. Numerous models are trained on small, single-center, or research-curated datasets that underrepresent rare cancers, minority populations, and atypical presentations, resulting in uneven performance—high accuracy for “textbook” cases but poor generalization to complex or marginalized populations. Data collected in tertiary centers and digital repositories may not translate effectively across scanners, institutions, or resource-limited settings [ 37 – 39 ]. Such gaps risk widening existing healthcare disparities: models developed to improve accessibility may inadvertently exacerbate inequity if training data are not inclusive. [ 37 ] Further limitations stem from data quality, annotation bias, and lack of external validation, which compromise reproducibility and clinical robustness [ 40 ]. Other technical vulnerabilities include adversarial data manipulation (intentional modification of input data to negatively affect the performance of a model or alter its behavior), model drift (the decline of model performance or accuracy over time as new data differs from training data), and software decentralization, which can complicate accountability and interoperability when independent hospitals, vendors, and research algorithms are involved [ 3 ]. Without consistent oversight, AI may silently alter performance and blur responsibility when diagnostic errors occur. Implementation also requires robust digital infrastructure, cybersecurity safeguards, and cross-disciplinary coordination. Adversarial data manipulation and malware vulnerabilities can corrupt diagnostic pipelines [ 3 ]. Moreover, AI development in pathology is highly decentralized: commercial vendors, research groups, and hospital teams often produce proprietary algorithms with minimal interoperability. This fragmentation complicates accountability, version tracking, and regulatory supervision. Without standardized validation and post-deployment monitoring, silent software updates can alter model behavior without the clinician's awareness, creating patient safety risks. Finally, when diagnostic errors occur, responsibility is diffuse. It is often unclear whether fault lies with the clinician, the institution, or the algorithm’s developer. Effective integration, therefore, depends on clear governance frameworks that define liability, require independent auditing, and ensure algorithm retraining and performance reporting [ 3 , 41 ]. Collaborative oversight involving data scientists, engineers, and clinicians is essential to maintain transparency, ensure equitable model performance, and sustain trust in AI-assisted diagnostics. These limitations underscore that successful AI implementation in HNP is not solely a technical endeavor, but also an educational one. Clinicians must understand not only how AI systems operate but also how to interpret, question, and integrate their probabilistic outputs responsibly. Strengthening AI literacy and diagnostic skepticism at the post-doctoral level is therefore essential for safe translation of these technologies. A Promise and a Challenge: When Human and Machine Errors Intersect Beyond technical issues, AI interacts with and often amplifies human cognitive biases. For example, AI’s perceived objectivity and precision can create a “reliability halo,” encouraging clinicians, especially less experienced clinicians, to defer to its outputs instead of engaging in independent scrutiny [ 42 , 43 ]. Similarly, automation bias refers to the tendency to over-trust algorithmic outputs at the expense of critical evaluation. [ 44 , 45 ] In pathology, clinicians may accept AI-generated classifications without weighing the contradictory evidence from histopathological or clinical sources [ 42 , 43 ]. Most AI outputs are probabilistic and opaque, generating predictions without clear explanations. Clinicians may misinterpret these predictions as facts, attributing certainty to statistical estimations, and continual use without scrutiny can reinforce misplaced confidence in those predictions [ 3 ]. These issues are particularly dangerous in HNP, where rare subtypes, borderline lesions, and complex anatomic relationships demand nuanced interpretation. Thus, the tendency to blindly accept AI recommendations risks perpetuating misdiagnosis. Overconfidence is another common bias [ 46 , 47 ], among medical professionals, including HNP providers [ 41 ]. Less experienced clinicians tend to accept AI predictions as precise and trustworthy, even if the underlying models are unstable or poorly calibrated [ 41 ]. This can erode epistemic humility, the professional recognition that all diagnoses carry a degree of uncertainty. In a domain as complex as HNP, such misplaced confidence can curtail dialogue, discourage second opinions, and suppress healthy skepticism. Confirmation bias further contributes to the preferential seeking or interpretation of findings that support a favored diagnosis, while overlooking disconfirming evidence [ 45 , 47 ]. AI outputs may reinforce this by framing uncertainty as certainty, inadvertently narrowing diagnostic reasoning. Indeed, initial AI suggestions can bias final grading or classification in H&N histopathology [ 42 ], highlighting the importance of maintaining critical appraisal skills in conjunction with AI-assisted interpretation. In addition, anchoring occurs when clinicians place disproportionate weight on initial impressions, such as the first visible abnormality on a slide, and fail to adjust when new evidence emerges. [ 47 – 50 ] Initial AI-generated diagnoses can serve as an anchor that influences evaluators’ tendency to adhere to that frame and overlook discrepant histological or radiographic features. These anchors can exert a stronger influence than peer input because of their perceived objectivity. [ 42 ] Finally, long-term reliance on AI may result in deskilling (diminishing human expertise), particularly at the postgraduate level. Medical education relies on repetition of judgment under uncertainty; it is through grappling with ambiguous slides, discrepant clinical histories, and imperfect evidence that future specialists acquire expertise. A recent multicenter, retrospective observational study investigated the risk of deskilling among endoscopists by assessing the rate at which they detected adenomas before and after using AI-assisted colonoscopy (AAC) technology. They reported an absolute 6% reduction in adenoma detection rate after three-month exposure to AAC, indicating potential deskilling of endoscopists [ 51 ]. Possible explanations for deskilling and overreliance in computer-aided detection (CAD) workflows have been suggested by evaluating eye-tracking when using these technologies in analyses of colorectal and breast imaging, where reduced time was spent reviewing the images and increased false positive rates were observed when CAD was used. [ 52 , 53 ] As an imaging-based specialty, pathologists are at risk of similar phenomena. If trainees defer to AI for definitive answers, their ability to synthesize pathological, radiographic, and clinical data may weaken over time [ 48 , 54 ] For example, while studies on oral epithelial dysplasia (OED) grading and multiomics modeling in HNP illustrate how automation can streamline routine assessments, they may also potentially limit experiential learning [ 15 ]. The same architecture that leads to cognitive biases also determines whether pathologists will trust, question, or dismiss AI-generated outputs. Designing Educational Frameworks for AI-Augmented Decision-Making The rapid adoption of AI technologies, including machine learning (ML), deep learning (DL), computer vision (the processing and interpretation of visual data through machine learning techniques and AI), natural language processing (technology that enables computers to understand, interpret and manipulate human language), and large language models (e.g., GPT, Claude), is transforming healthcare education. These systems now support a wide range of activities, from admissions and tutoring to virtual simulation, adaptive learning, and clinical assessment [ 55 ]. Their integration into training programs offers unprecedented opportunities to personalize learning, expand exposure to rare cases, and enhance diagnostic reasoning. Yet the educational promise of AI also introduces a profound pedagogical challenge (Fig. 1 ). Traditional healthcare curricula were designed around human expertise, emphasizing foundational knowledge, critical appraisal, ethical judgment, and professionalism. In the AI era, these competencies must evolve. Clinicians should learn not only how to use AI tools but how to think with them—critically, reflectively, and responsibly. This requires reframing medical education toward what we term healthcare-optimization training : a model that integrates AI literacy and diagnostic skepticism with the enduring principles of human accountability, ethical integrity, and patient-centered care. Fig. 1. Open in a new tab AI: artificial intelligence. The diagram summarizes the promises (upper blue panel) and challenges (lower pink panel) of applying AI in head and neck pathology. To mitigate these challenges, users must acquire specific skills (purple left-hand panel) and consider warnings (red right-hand panel) when implementing AI tools Revised Core Competencies in the AI Era Integrating AI into diagnostic education requires redefining core professional competencies rather than merely adding new technical skills. The American Medical Association (AMA) has issued policy recommendations that encourage education on the capabilities and limitations of AI [ 56 ]. However, current accreditation bodies, including the Accreditation Council for Graduate Medical Education (ACGME), Accreditation Council for Continuing Medical Education (ACCME), and CanMEDS have not yet formally incorporated AI-specific competencies. Expanding on Forney and McBright’s work applied to radiology training, recently, Vos et al. proposed a new EPA entitled “using AI in diagnostic pathology practice.” [ 6 , 57 ] They also suggest that certain traditional competencies such as grossing, autopsies and intraoperative histologic examination, as well as in multidisciplinary diagnostic and management settings will remain primarily the task of the pathologist. However, they speculate that microscopic evaluation and reporting are likely to be taken over by AI, including the detection, characterization, quantification and prediction of disease through automated image analysis. Furthermore, with the broader implementation of AI in pathology, they expect competencies to include more self-directed learning and quality control activities. These would shift the role of a pathologist from primarily a diagnostician to that of a data analyst, utilizing foundational biomedical knowledge and traditional pathology training to select, evaluate, interpret and analyze data inputs and outputs for quality and accuracy. Additionally, greater emphasis is expected in the understanding of foundational informatics in the development and implementation of AI models in diagnostic and operational workflows. Consequently, quality improvement-, quality assurance-, and self-assessment-related competencies will gain further importance for the identification and mitigation of biases and prevention of deskilling [ 6 ]. Rather than create separate AI competencies training pathologists in this new paradigm, we suggest that AI-assisted workflows are inevitable, contemporary expressions of the principles of pathology practice. Responsible integration of AI into traditional pathology educational competencies should be considered, for example, including AI validation for clinical and specimen data extraction, bias awareness during diagnostic reasoning, algorithmic performance monitoring during quality assurance tasks, and transparency and equity considerations with respect to AI use when considering professionalism standards. Thus, clinicians should learn not only how to use AI tools but how to think with them—critically, reflectively, and responsibly in every aspect of their practice. AI does not eliminate error but can influence where and how errors occur. This requires reframing medical education to address these new vulnerabilities. Healthcare-Optimization Training: A framework for AI-augmented decision-making education One approach to address these vulnerabilities is to implement healthcare-optimization training —a model that integrates AI literacy and diagnostic skepticism with the enduring principles of human accountability, ethical integrity, and patient-centered care. This training would incorporate a decision-making framework that facilitates the identification of important tasks of the practice of pathology where delegation to AI tools and workflows are indicated or appropriate, indicated with caution, or never appropriate. This structured approach would also identify risks for cognitive biases and their mitigation when incorporating AI into HNP training (Table 1 ). For example, main decision-making nodes relate to the determination of: 1) clinical adequacy, 2) specimen adequacy 3) microscopic adequacy 4) diagnostic accuracy 5) reporting adequacy 6) quality assurance and regulatory compliance and 7) ethical considerations, professionalism and lifelong learning. Table 1. Decision-Making Matrix in HNP with AI Considerations Component AI Delegation category Bias/risk Mitigation Clinical adequacy Assess completeness of clinical data Appropriate with Caution Context neglect, automation bias (accepting AI summaries even if incomplete) Cross-check AI summary with clinician notes; require human validation prompts Flag missing clinical data Appropriate Data visibility bias (hidden or unstructured data ignored by AI) Force-complete clinical intake fields; highlight missing elements Determine when clinical uncertainty limits interpretation Never Appropriate to Delegate AI overconfidence bias Require human judgment documentation; teach uncertainty framing Specimen adequacy Assess tissue adequacy Appropriate with Caution Anchoring (trusting AI adequacy labels) Human override rules; correlate with gross exam Recognize artifacts Appropriate with Caution Pattern-matching bias (AI confuses artifacts with neoplasm) Human training in artifact libraries; cross-sectional review Assess specimen labeling and chain of custody Appropriate Workflow omission bias Barcode verification; audit logs Microscopic adequacy Field selection & quality check Appropriate Automation bias Random spot-checks Detection of atypia, dysplasia, invasive features Appropriate with Caution Confirmation bias, AI hallucination Blind second reads; mandate human review of all flagged foci Rare variant recognition Never Appropriate Data scarcity bias Required human expert sign-off Diagnostic adequacy Establishing the diagnostic category (reactive, benign, dysplastic, malignant, etc.) Appropriate with Caution Anchoring, overtrust, base-rate neglect Mandatory human synthesis; require AI uncertainty quantification Selecting ancillary tests Never Appropriate Algorithmic echoing (ordering tests AI was trained on) Human clinical reasoning; standard operating procedures for test indication Multimodal data integration (morphology + clinical + AI) Appropriate with Caution Information overload bias Structured interpretation templates Reporting adequacy Generate structured reporting (synoptic elements) Appropriate Omission bias Auto-validation, completeness checks Narrative interpretation & clinical impact statement Never Appropriate Descriptive drift, AI verbosity bias Human-authored contextualization Risk stratification & prognostic parameter synthesis Appropriate with Caution Over-reliance on AI prognostic models Transparency requirements; documentation of model limits Quality assurance and regulatory compliance Quality assurance monitoring Appropriate Automation omission bias Dashboard review by quality assurance lead Peer review / secondary sign-out requirements Never Appropriate Delegation drift Hard-coded second sign-out for certain diagnoses Compliance audits, documentation Appropriate Compliance blind spots in AI Human verification & audit trails Ethical considerations, professionalism and lifelong learning Equity monitoring (performance across demographics) Appropriate Hidden bias, dataset bias Bias audits; require disaggregated performance data Human–AI interaction safety (automation bias, overreliance) Never Fully Delegable Overtrust, diminished vigilance Teach “trust-but-verify,” simulation training Consent, transparency, and explainability Appropriate with Caution Opaque model reasoning Require explainability statements in report addenda Reflect on diagnostic challenges and cognitive errors Never Anchoring, confirmation, self-assessment bias Structured reflection, peer review, case logs Maintain currency with WHO revisions, CAP/ASCP guidelines, and evidence updates Appropriate Knowledge decay, outdated protocol reliance CME, scheduled guideline review, QA updates Ensure high-quality interpersonal communication with staff, clinicians, and teams Never Communication bias, hierarchy bias Structured reporting templates, interdisciplinary rounds Evaluate clinician satisfaction with pathology communication Appropriate Feedback bias (positive/negative extremes) Systematic surveys, closed-loop follow-up Assess need for subspecialty consults and quality of referral Never Referral bias, overreliance on familiarity Case-based checklists, consult guidelines Identify AI-driven bias and adjust interpretation accordingly Appropriate with Caution Automation bias, dataset bias, overtrust in AI Periodic AI performance audits, bias review, human verification Communicate AI involvement transparently in clinical interactions and tumor boards Appropriate Opacity bias, overreliance perception Document AI usage, explanatory statements, ensure interpretability Open in a new tab This table outlines major decision-making nodes and whether delegation to AI tools would be appropriate, potential biases and risks involved with this approach and strategies for their mitigation AI Literacy with Epistemic Humility A foundational understanding of how algorithms are designed, trained, validated, and limited is fundamental to AI literacy [ 6 , 57 – 59 ]. Clinicians must understand the strengths and limitations of model architectures, data representativeness, and bias sources to critically interpret outputs rather than accepting them as objective truths. AI literacy thus bridges technical awareness with epistemic humility, enabling clinicians to appreciate both the capabilities and fallibility of algorithmic reasoning, addressing both automation bias and overconfidence. Diagnostic Skepticism Diagnostic skepticism, the disciplined capacity to question AI-generated outputs before incorporating them into clinical reasoning, also is key to this training. Because algorithms can reproduce biases or generate errors and “hallucinations” (plausible but factually incorrect or fabricated AI output), learners must cultivate reflective reasoning to evaluate when an AI recommendation is consistent with evidence (i.e., clinical context, histopathological patterns, and molecular data) and when it demands scrutiny [ 48 , 60 , 61 ]. This skepticism protects against automation bias, confirmation bias and anchoring, and promotes independent judgment—skills essential for responsible collaboration between humans and machines. Transparency Additionally, transparency must be reframed as a professional and ethical obligation. Learners and clinicians should be able to articulate how AI tools have influenced their reasoning, justify their reliance on specific outputs, and clearly communicate uncertainty to colleagues and patients. Integrating these competencies shifts AI from a passive shortcut into an active partner in clinical learning, fostering accountability, interpretive depth, and patient-centered decision-making. [ 62 ] Pedagogical Strategies and Implementation Traditional lecture-based models have evolved into interactive, competency-driven learning models that emphasize problem-solving and reflection. [ 63 ] Such approaches, like flipped classrooms, small-group discussions, and case-based learning, form the foundation for AI-enhanced education in HNP. AI can strengthen these pedagogies by enabling simulation, adaptive feedback, and exposure to diverse and rare diagnostic scenarios. AI systems can provide real-time, individualized feedback, highlighting missed features, refining questioning, and prompting reevaluation—thereby promoting self-directed learning while complementing, not replacing, patient encounters [ 64 ]. Thus, we propose development of key cognitive skills that move the learner from basic self-awareness to human-Al reconciliation and simulation-based practice. This sequencing can foster metacognitively aware, bias-resistant, and diagnostically accurate pathologists equipped to work safely and effectively with AI. Cognitive Bias Recognition Recognition of these cognitive biases is the first step in establishing awareness of systematic thinking errors that may influence diagnostic decisions and are foundational to metacognitive skill development and safe AI implementation . Non-AI Assessments The constant and growing possibilities for incorporation of AI-driven processes in education and clinical practice can lead to deskilling (erosion of previous skills), “never-skilling” (lack of specific essential skill development), mis-skilling” (amplification of errors and bias) [ 65 , 66 ]. Unaided diagnostic reasoning is an essential skill which allows identification of cognitive biases and knowledge gaps and help build and calibrate confidence [ 30 ]. Programs should incorporate workflows without the use of AI to ensure learners maintain independent clinical reasoning, diagnostic accuracy and metacognitive awareness, preserving the human expertise at the heart of clinical decision-making. Confidence Calibration When implementing AI tools, helping trainees distinguish their perceived diagnostic performance with their actual performance is a necessary calibration step, as both overconfidence and underconfidence can affect patient safety. Students in initial stages of training consistently exhibit overconfidence bias, often influenced by anchoring. Sample exercises to achieve confidence calibration consists of trainees reviewing slides independently, recording their confidence level, then comparing them with AI-generated outputs and the actual diagnoses. AI-generated outputs and confirmed final diagnoses act as benchmarks for calibration. [ 42 , 48 ] Decision Audits and Reflection Decision audits further encourage metacognitive awareness. By comparing initial diagnostic judgments with those revised after consulting AI suggestions, learners can identify tendencies toward automation bias, algorithm aversion (loss of confidence, distrust, or negative bias towards AI outputs), or opacity of models [ 61 , 67 , 68 ]. Guided reflection has been shown to improve diagnostic decision-making, especially in complex cases and is most effective when immediate feedback is included and contrasting cases are also presented [ 68 – 70 ]. Trainees should first review slides and make their initial diagnosis using traditional pathology training methods (blinded to AI), review AI generated classification and diagnostic outputs, then determine the final diagnosis. This exercise should be followed by a formal reflective activity indicating whether the AI output changed their initial diagnosis and why, also elucidating additional datapoints that would increase confidence in the assessment. Concomitantly, trainees should identify potential influential cognitive biases. This structured reflection cultivates balanced collaboration between human reasoning and algorithmic input. AI Discrepancy Assessment Importantly, discrepancies among AI systems, arising from dataset imbalance, algorithmic variation, or differences in model architecture, should be reframed as teaching opportunities. Exposure to conflicting outputs helps learners appreciate the non-deterministic nature of AI and reinforces the need for evidence-based evaluations [ 71 – 73 ]. Through guided analysis, students can compare AI outputs against clinical guidelines, justify agreement or disagreement, and articulate their reasoning, strengthening accountability and critical reflection [ 73 ]. For example, trainees can review slides with outputs from three different AI models showing conflicting dysplasia grading, followed by an assessment of which model is trustworthy, justifying their decision by applying foundational medical knowledge that support or refute their conclusions, citing relevant evidence. The same process can be modeled for assessment of discrepancies between human and AI outputs. [ 42 ] Simulation-Based Practice Simulation exercises, once limited to manikins or physical models, now can incorporate AI-driven virtual environments that replicate complex diagnostic situations [ 74 ]. In pathology, AI-powered slide repositories and case-based modules allow learners to engage with both common and rare lesions previously encountered only in textbooks, promoting iterative diagnostic reasoning and deliberate practice [ 69 , 74 ]. While AI can democratize access to high-quality educational resources, disparities in digital literacy, technology access, and faculty readiness risk widening the educational divide. [ 75 – 77 ] Development of this pedagogical process requires access to high-quality WSI, publicly accessible databases integrating histological, genomic, and clinical data, and standardized annotation protocols. They should be integrated into a comprehensive educational framework that includes traditional microscopy, case-based learning, and expert consultation. Figure 2 maps traditional ACGME milestones for pathology [ 78 ] to major decision-making nodes and AI workflow training considerations, while Fig. 3 ties them to the proposed pedagogical strategies. Fig. 2. Open in a new tab Decision-making nodes, ACGME pathology milestones, and AI workflow training considerations. Legend: Patient Care (PC); PC1- Reporting; PC2: Grossing; PC3: Clinical Consultation, including On-Call Interactions; PC4: Interpretation and Diagnosis; PC5: Intra-Operative Consultation, including Frozen Section; PC6: Autopsy. Medical Knowledge (MK); MK1: Diagnostic Knowledge; MK2: Clinical Reasoning. Systems-Based Practice SBP1: Patient Safety and Quality Improvement; SBP2: Systems Navigation for Patient-Centered Care; SBP3: Physician Role in Health Care System; SBP4: Informatics; SBP5: Accreditation, Compliance and Quality. Practice-Based Learning and Improvement (PBLI); PBLI1: Evidence-Based Practice and Scholarship; PBLI2: Reflective Practice and Commitment to Personal Growth. Professionalism (PROF); PROF1: Professional Behavior and Ethical Principles; PROF2: Accountability and Conscientiousness; PROF3: Self-Awareness and Help-Seeking. Interpersonal and Communication Skills (ICS); ICS1: Patient and Family-Centered Communication; ICS2: Interprofessional and Team Communication; ICS3: Communication. within Health Care Systems This figure maps the proposed major decision-making nodes to traditional ACGME pathology milestones [ 75 ] while highlighting AI workflow training considerations and the necessary core skills Fig. 3. Open in a new tab This figure maps the proposed major decision-making nodes to skillsets for risk mitigation (AI literacy, diagnostic skepticism and transparency) and pedagogical strategies in HNP Ethical, Legal, and Institutional Implications of AI in HNP Integration of AI into HNP raises profound ethical, legal, and institutional challenges. The distinctive complexity of HNP makes it a domain particularly sensitive to questions of responsibility, trust, and governance [ 22 , 26 ] While AI systems promise to enhance consistency and efficiency in diagnosis, implementation must be guided by frameworks that balance innovation with prudence, ensuring that AI functions as a decision-support tool rather than a replacement for human expertise. Ethically, the adoption of AI in pathology intersects with the core principles of beneficence, non-maleficence, autonomy, and justice [ 79 ]. AI tools may improve patient outcomes by reducing errors and expediting diagnoses, thereby serving beneficence. Yet risks of algorithmic bias, opaque decision-making, and over-reliance on technology can introduce potential harm [ 41 ]. Respecting patient autonomy also requires transparency. Clinicians must be able to explain AI-supported decisions in comprehensible terms to patients and colleagues. This aligns with calls for epistemic humility in medicine, emphasizing acknowledgment of uncertainty and limits of knowledge. [ 80 ] Without humility and understanding, there is a risk of overstating AI’s accuracy and underestimating its vulnerabilities. From a legal perspective, AI raises open questions of liability. If an algorithm contributes to a diagnostic error, responsibility ultimately rests with the pathologist, as professional standards of care continue to place clinicians at the center of accountability [ 81 ]. This legal position underscores why AI should not be positioned as an autonomous decision-maker but rather as a supportive adjunct. In practice, this requires institutions to establish clear governance mechanisms for the validation, monitoring, and retraining of algorithms over time [ 41 ]. Pathology-specific evidence of diagnostic error further illustrates the stakes. Discrepancies in HNP remain a persistent problem, demanding institutional safeguards even in human-only practice. [ 22 , 26 ] AI, if poorly governed, risks amplifying rather than mitigating these errors. Institutional integration of AI must be accompanied by educational reform. Training should focus on equipping clinicians with the technical literacy and critical reasoning skills necessary to evaluate AI recommendations within the clinical context. This reflects broader arguments for embedding humility into professional systems, ensuring that both individuals and institutions retain the ability to challenge automated outputs. [ 42 ] At an institutional level, governance should incorporate continuous algorithmic monitoring, clear delineation of responsibility, and processes for appeal and oversight. These structures parallel existing quality assurance mechanisms in pathology but require adaptation to account for novel risks such as dataset drift (changes to input data over time relative to model training data, adversarial vulnerabilities, and shifting regulatory standards. [ 82 ] Without these safeguards, the promise of AI may erode trust in technology and the institutions deploying it. Conclusion AI is transforming HNP, expanding diagnostic capacity while challenging long-held assumptions about expertise, reasoning, and accountability. Recent work [ 6 ] outlined infrastructural and regulatory competencies required for the responsible adoption of AI in pathology training. Our approach diverges by emphasizing the cognitive and decision-making dimensions of AI integration. Rather than focusing solely on formal EPAs or systemic implementation, we analyze how clinicians interpret, trust, or over-rely on algorithmic outputs, and how educational frameworks can mitigate such risks. This focus is especially critical in HNP, where diagnostic uncertainty can be high, and the interpretive demands are nuanced. Our proposed framework complements existing competency-based initiatives by addressing the human factors that shape diagnostic decisions in AI-enhanced environments, aiming not only to integrate technology but to enhance clinical reasoning and ethical reflection in parallel. The intersection of human and machine reasoning introduces new vulnerabilities that demand deliberate educational reform. Reframing decision-making education for the AI era means teaching clinicians not only how to use intelligent systems, but how to think with them. AI literacy, diagnostic skepticism, and ethical transparency must therefore be treated as core professional skills, alongside reflective practice and continuous awareness of bias. Trainees must, therefore, be formally instructed to analyze AI-assisted outputs within pathology practice. Future educational development with relevant stakeholders can expand on the intersection of these AI-assisted workflows within the six core ACGME pathology competency domains and milestones to create AI-related themes for each subcompetency. By embedding these principles into post-doctoral education, competency-based assessments, and institutional governance, HNP can evolve toward an epistemically humble, technologically fluent, and ethically resilient model of clinical reasoning. In this model, AI does not replace the pathologist’s judgment; it refines it—turning human–machine collaboration into a catalyst for safer, fairer, and more patient-centered care. Author Contributions All authors whose names appear on the submission made substantial contributions to the conception or design of the work; drafted the work or revised it critically for important intellectual content; approved the version to be published; and agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. Funding This study was not supported by any funding. Data Availability No datasets were generated or analysed during the current study. Declarations Conflict of interest The authors declare no competing interests. Ethical Approval This article does not contain any studies with human participants or animals performed by any of the authors. Informed Consent For this type of study, informed consent is not required. For this type of study, consent for publication is not required. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Pham TD, Teh M-T, Chatzopoulou D, Holmes S, Coulthard P (2024) Artificial intelligence in head and neck cancer: innovations, applications, and future directions. Curr Oncol 31(9):5255–5290 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Mirfendereski P, Li GY, Pearson AT, Kerr AR (2025) Artificial intelligence and the diagnosis of oral cavity cancer and oral potentially malignant disorders from clinical photographs: a narrative review. Front Oral Health 6:1569567 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Rao KN, Fernandez-Alvarez V, Guntinas-Lichius O, Sreeram MP, de Bree R, Kowalski LP, Forastiere A et al (2025) The limitations of artificial intelligence in head and neck oncology. Adv Ther 42(6):2559–2568 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Poller DN, Johnson SJ, Bongiovanni M (2020) Measures to reduce diagnostic error and improve clinical decision making in thyroid FNA aspiration cytology: a proposed framework. Cancer Cytopathol 128(12):917–927 [ DOI ] [ PubMed ] [ Google Scholar ] 5. Jussupow E, Spohrer K, Heinzl A, Gawlitza J (2021) Augmenting medical diagnosis decisions? An investigation into physicians’ decision-making process with artificial intelligence. Inf Syst Res 32(3):713–735 [ Google Scholar ] 6. Vos S, Hebeda K, Milota M, Sand M, Drogt J, Grünberg K, Jongsma K (2025) Making pathologists ready for the new artificial intelligence era: changes in required competencies. Mod Pathol 38(2):100657 [ DOI ] [ PubMed ] [ Google Scholar ] 7. Alter IL, Chan K, Lechien J, Rameau A (2024) An introduction to machine learning and generative artificial intelligence for otolaryngologists—head and neck surgeons: a narrative review. Eur Arch Otorhinolaryngol 281(5):2723–2731 [ DOI ] [ PubMed ] [ Google Scholar ] 8. Rashidi HH, Pantanowitz J, Hanna MG, Tafti AP, Sanghani P, Buchinsky A, Fennell B et al (2025) Introduction to artificial intelligence and machine learning in pathology and medicine: generative and nongenerative artificial intelligence basics. Mod Pathol 38(4):100688 [ DOI ] [ PubMed ] [ Google Scholar ] 9. Rashidi HH, Pantanowitz J, Chamanzar A, Fennell B, Wang Y, Gullapalli RR, Tafti A et al (2025) Generative artificial intelligence in pathology and medicine: a deeper dive. Mod Pathol 38(4):100687 [ DOI ] [ PubMed ] [ Google Scholar ] 10. Pantanowitz L, Pearce T, Abukhiran I, Hanna M, Wheeler S, Soong TR, Tafti AP et al (2024) Non-generative artificial intelligence (AI) in medicine: advancements and applications in supervised and unsupervised machine learning. Mod Pathol 38:100680 [ DOI ] [ PubMed ] [ Google Scholar ] 11. van Dooijeweert C, ter Hoeve ND, Nguyen T, Breimer G, Blokx W, Stathonikos N, van Diest PJ, Flach R (2025) Broad implementation of AI for lymph node assessment: Insights from a head-to-head comparison of two applications. J Clin Oncol 43:e13679–e13679 [ Google Scholar ] 12. Alsanie I, Shephard A, Azarmehr N, Vargas P, Pring M, Rajpoot NM, Khurram SA (2025) Exploring the feasibility of AI-based analysis of histopathological variability in salivary gland tumours. Sci Rep 15(1):29171 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Wu Q, Wang X, Liang G, Luo X, Zhou M, Deng H, Zhang Y, Huang X, Yang Q (2023) Advances in image‐based artificial intelligence in otorhinolaryngology–head and neck surgery: a systematic review. Otolaryngol Head Neck Surg 169(5):1132–1142 [ DOI ] [ PubMed ] [ Google Scholar ] 14. Sant VR, Radhachandran A, Ivezic V, Lee DT, Livhits MJ, Wu JX, Masamed R, Arnold CW, Yeh MW, Speier W (2024) From bench-to-bedside: how artificial intelligence is changing thyroid nodule diagnostics, a systematic review. J Clin Endocrinol Metab 109(7):1684–1693 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Shephard AJ, Mahmood H, Raza SEA, Khurram SA, Rajpoot NM (2025) A novel AI-based score for assessing the prognostic value of intra-epithelial lymphocytes in oral epithelial dysplasia. Br J Cancer 132(2):168–179 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Alajaji SA, Khoury ZH, Jessri M, Sciubba JJ, Sultan AS (2024) An update on the use of artificial intelligence in digital pathology for oral epithelial dysplasia research. Head Neck Pathol 18(1):38 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Hanna MG, Pantanowitz L, Jackson B, Palmer O, Visweswaran S, Pantanowitz J, Deebajah M, Rashidi HH (2025) Future of artificial intelligence (AI)-machine learning (ML) trends in pathology and medicine. Mod Pathol 38:100705 [ DOI ] [ PubMed ] [ Google Scholar ] 18. Song B, Yadav I, Tsai JC, Madabhushi A, Kann BH (2025) Artificial intelligence for head and neck squamous cell carcinoma: from diagnosis to treatment. Am Soc Clin Oncol Educ Book 45(3):e472464 [ DOI ] [ PubMed ] [ Google Scholar ] 19. Broggi G, Maniaci A, Lentini M, Palicelli A, Zanelli M, Zizzo M, Koufopoulos N, Salzano S, Mazzucchelli M, Caltabiano R (2024) Artificial intelligence in head and neck cancer diagnosis: a comprehensive review with emphasis on radiomics, histopathological, and molecular applications. Cancers 16(21):3623 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Fathallah L, Tulunay OE, Feng J, Husain M, Jacobs JR, Al-Abbadi MA (2006) Histopathologic and cytopathologic diagnostic discrepancies in head and neck region: pitfalls, causes, and preventive strategies. Otolaryngol Head Neck Surg 134(2):302–308 [ DOI ] [ PubMed ] [ Google Scholar ] 21. Upadhyaya JD, Fitzpatrick SG, Cohen DM, Bilodeau EA, Bhattacharyya I, Lewis JS Jr, Lai J et al (2020) Inter-observer variability in the diagnosis of proliferative verrucous leukoplakia: clinical implications for oral and maxillofacial surgeon understanding: a collaborative pilot study. Head Neck Pathol 14(1):156–165 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Woolgar JA, Ferlito A, Devaney KO, Rinaldo A, Barnes L (2011) How trustworthy is a diagnosis in head and neck surgical pathology? A consideration of diagnostic discrepancies (errors). Eur Arch Otorhinolaryngol 268(5):643–651 [ DOI ] [ PubMed ] [ Google Scholar ] 23. Stanovich KE, West RF (2000) Advancing the rationality debate. Behav Brain Sci 23(5):701–717 [ DOI ] [ PubMed ] [ Google Scholar ] 24. Kahneman D, Frederick S (2002) Representativeness revisited: attribute substitution in intuitive judgment. Heuristics Biases Psychol Intuitive Judgm 49(49–81):74 [ Google Scholar ] 25. Sibbald M, Sherbino J, Preyra I, Coffin‐Simpson T, Norman G, Monteiro S (2017) Eyeballing: the use of visual appearance to diagnose ‘sick.’ Med Educ 51(11):1138–1145 [ DOI ] [ PubMed ] [ Google Scholar ] 26. Ayal S, Rusou Z, Zakay D, Hochman G (2015) Determinants of judgment and decision making quality: the interplay between information processing style and situational factors. Front Psychol 6:1088 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Hochman G (2024) Beyond the surface: a new perspective on dual-system theories in decision-making. Behav Sci 14(11):1028 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Norman GR, Monteiro SD, Sherbino J, Ilgen JS, Schmidt HG, Mamede S (2017) The causes of errors in clinical reasoning: cognitive biases, knowledge deficits, and dual process thinking. Acad Med 92(1):23–30 [ DOI ] [ PubMed ] [ Google Scholar ] 29. Norman G (2005) Research in clinical reasoning: past history and current trends. Med Educ 39(4):418–427 [ DOI ] [ PubMed ] [ Google Scholar ] 30. Eva KW (2005) What every teacher needs to know about clinical reasoning. Med Educ 39(1):98–106 [ DOI ] [ PubMed ] [ Google Scholar ] 31. Croskerry P (2013) From mindless to mindful practice—cognitive bias and clinical decision making. N Engl J Med 368(26):2445–2448 [ DOI ] [ PubMed ] [ Google Scholar ] 32. Kostopoulou O, Sirota M, Round T, Samaranayaka S, Delaney BC (2017) The role of physicians’ first impressions in the diagnosis of possible cancers without alarm symptoms. Med Decis Making 37(1):9–16 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Frenkenberg A, Hochman G (2025) It’s scary to use it, it’s scary to refuse it: the psychological dimensions of AI adoption—anxiety, motives, and dependency. Systems 13(2):82 [ Google Scholar ] 34. Grote T, Berens P (2020) On the ethics of algorithmic decision-making in healthcare. J Med Ethics 46(3):205–211 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Patil SV, Myers CG, Lu-Myers Y (2025) Calibrating AI reliance—A physician’s superhuman dilemma. JAMA Health Forum 6(3):e250106–e250106 [ DOI ] [ PubMed ] [ Google Scholar ] 36. Reis-Filho JS, Kather JN (2023) Overcoming the challenges to implementation of artificial intelligence in pathology. JNCI J Natl Cancer Inst 115(6):608–612 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Norori N, Hu Q, Aellen FM, Faraci FD, Tzovara A (2021) Addressing bias in big data and AI for health care: a call for open science. Patterns 2(10):100347 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Cheng CH, Shi S-S (2025) Artificial intelligence in cancer: applications, challenges, and future perspectives. Mol Cancer 24(1):274 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Vaidya A, Chen RJ, Williamson DF, Song AH, Jaume G, Yang Y, Hartvigsen T et al (2024) Demographic bias in misdiagnosis by computational pathology models. Nat Med 30(4):1174–1190 [ DOI ] [ PubMed ] [ Google Scholar ] 40. Stenzinger A, Alber M, Allgaeuer M, Jurmeister P, Bockmayr M, Budczies J, Lennerz J et al (2022) Artificial intelligence and pathology: from principles to practice and future applications in histomorphology and molecular profiling. Seminars in cancer biology, vol 84. Academic Press, Cambridge, pp 129–143 [ DOI ] [ PubMed ] [ Google Scholar ] 41. Finlayson SG, Subbaswamy A, Singh K, Bowers J, Kupke A, Zittrain J, Kohane IS, Saria S (2021) The clinician and dataset shift in artificial intelligence. N Engl J Med 385(3):283–286 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 42. Bellahsen-Harrar Y, Lubrano M, Lépine C, Beaufrère A, Bocciarelli C, Brunet A, Decroix E et al (2025) Exploring the risks of over-reliance on AI in diagnostic pathology. What lessons can be learned to support the training of young pathologists? PLoS ONE 20(8):e0323270 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Evans H, Snead D (2024) Why do errors arise in artificial intelligence diagnostic tools in histopathology and how can we minimize them? Histopathology 84(2):279–287 [ DOI ] [ PubMed ] [ Google Scholar ] 44. Croskerry P (2003) The importance of cognitive errors in diagnosis and strategies to minimize them. Acad Med 78(8):775–780 [ DOI ] [ PubMed ] [ Google Scholar ] 45. Graber ML, Franklin N, Gordon R (2005) Diagnostic error in internal medicine. Arch Intern Med 165(13):1493–1499 [ DOI ] [ PubMed ] [ Google Scholar ] 46. Berner ES, Graber ML (2008) Overconfidence as a cause of diagnostic error in medicine. Am J Med 121(5):S2–S23 [ DOI ] [ PubMed ] [ Google Scholar ] 47. Baron J, Ritov I (1994) Reference points and omission bias. Organ Behav Hum Decis Process 59(3):475–498 [ DOI ] [ PubMed ] [ Google Scholar ] 48. Brunyé TT, Mitroff SR, Elmore JG (2025) Artificial intelligence and computer-aided diagnosis in diagnostic decisions: 5 questions for medical informatics and human-computer interface research. J Am Med Inform Assoc 33:ocaf123 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Tversky A, Kahneman D (1974) Judgment under uncertainty: heuristics and biases: biases in judgments reveal some heuristics of thinking under uncertainty. Science 185(4157):1124–1131 [ DOI ] [ PubMed ] [ Google Scholar ] 50. Meyer AND, Payne VL, Meeks DW, Rao R, Singh H (2013) Physicians’ diagnostic accuracy, confidence, and resource requests: a vignette study. JAMA Intern Med 173(21):1952–1958 [ DOI ] [ PubMed ] [ Google Scholar ] 51. Budzyń K, Romańczyk M, Kitala D, Kołodziej P, Bugajski M, Adami HO, Blom J et al (2025) Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterol Hepatol 10(10):896–903 [ DOI ] [ PubMed ] [ Google Scholar ] 52. Du-Crow E, Astley SM, Hulleman J (2020) Is there a safety-net effect with computer-aided detection? J Med Imaging 7(2):022405–022405 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 53. Troya J, Fitting D, Brand M, Sudarevic B, Kather JN, Meining A, Hann A (2022) The influence of computer-aided polyp detection systems on reaction time for polyp detection and eye gaze. Endoscopy 54(10):1009–1014 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Salloch S, Eriksen A (2024) What are humans doing in the loop? Co-reasoning and practical judgment when using machine learning-driven decision aids. Am J Bioeth 24(9):67–78 [ DOI ] [ PubMed ] [ Google Scholar ] 55. Gordon M, Daniel M, Ajiboye A, Uraiby H, Xu NY, Bartlett R, Hanson J et al (2024) A scoping review of artificial intelligence in medical education: BEME Guide No. 84. Med Teach 46(4):446–470 [ DOI ] [ PubMed ] [ Google Scholar ] 56. HOD (2019) ACTION: Council on Medical Education Report 4 adopted as amended at, filed. rotr. Augmented intelligence in medical education (Resolution 317-A-18). Available at: https://www.ama-assn.org/system/files/cme-report-4-a19-annotated.pdf?utm_source=chatgpt.com . Accessed 29 Aug 2025 57. Forney MC, McBride AF (2020) Artificial intelligence in radiology residency training. In Seminars in musculoskeletal radiology, vol 24, no 01. Thieme Medical Publishers, pp 74–80 [ DOI ] [ PubMed ] 58. Liaw W, Kueper JK, Lin S, Bazemore A, Kakadiaris I (2022) Competencies for the use of artificial intelligence in primary care. Ann Fam Med 20(6):559–563 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 59. Schwendicke F, Chaurasia A, Wiegand T, Uribe SE, Fontana M, Akota I, Tryfonos O, Krois J (2023) Artificial intelligence for oral and dental healthcare: core education curriculum. J Dent 128:104363 [ DOI ] [ PubMed ] [ Google Scholar ] 60. Arora A, Alderman JE, Palmer J, Ganapathi S, Laws E, Mccradden MD, Oakden-Rayner L et al (2023) The value of standards for health datasets in artificial intelligence-based applications. Nat Med 29(11):2929–2938 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 61. Hasanzadeh F, Josephson CB, Waters G, Adedinsewo D, Azizi Z, White JA (2025) Bias recognition and mitigation strategies in artificial intelligence healthcare applications. NPJ Digit Med 8(1):154 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 62. McKay F, Williams BJ, Prestwich G, Bansal D, Hallowell N, Treanor D (2022) The ethical challenges of artificial intelligence-driven digital pathology. J Pathol Clin Res 8(3):209–216 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 63. Knollmann-Ritschel BEC, Regula DP, Borowitz MJ, Conran R, Prystowsky MB (2017) Pathology competencies for medical education and educational cases. Acad Pathol 4:2374289517715040 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 64. Elendu C, Amaechi DC, Okatta AU, Amaechi EC, Elendu TC, Ezeh CP, Elendu ID (2024) The impact of simulation-based training in medical education: A review. Medicine 103(27):e38813 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 65. Schwartzstein RM (2024) Clinical reasoning and artificial intelligence: can AI really think? Trans Am Clin Climatol Assoc 134:133 [ PMC free article ] [ PubMed ] [ Google Scholar ] 66. Abdulnour R-E, Gin B, Boscardin CK (2025) Educational strategies for clinical supervision of artificial intelligence use. N Engl J Med 393(8):786–797 [ DOI ] [ PubMed ] [ Google Scholar ] 67. Dratsch T, Chen X, Rezazade Mehrizi M, Kloeckner R, Mähringer-Kunz A, Püsken M, Baeßler B, Sauer S, Maintz D, Pinto dos Santos D (2023) Automation bias in mammography: the impact of artificial intelligence BI-RADS suggestions on reader performance. Radiology 307(4):e222176 [ DOI ] [ PubMed ] [ Google Scholar ] 68. Drew BJ, Harris P, Zègre-Hemsey JK, Mammone T, Schindler D, Salas-Boni R, Bai Y, Tinoco A, Ding Q, Xiao H (2014) Insights into the problem of alarm fatigue with physiologic monitor devices: a comprehensive observational study of consecutive intensive care unit patients. PLoS ONE 9(10):e110274 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 69. Prakash S, Sladek RM, Schuwirth L (2019) Interventions to improve diagnostic decision making: a systematic review and meta-analysis on reflective strategies. Med Teach 41(5):517–524 [ DOI ] [ PubMed ] [ Google Scholar ] 70. Mamede S, Schmidt HG (2023) Deliberate reflection and clinical reasoning: founding ideas and empirical findings. Med Educ 57(1):76–85 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 71. Weiner EB, Dankwa-Mullan I, Nelson WA, Hassanpour S (2025) Ethical challenges and evolving strategies in the integration of artificial intelligence into clinical practice. PLOS Digit Health 4(4):e0000810 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 72. GM Council (2025) Artificial intelligence and innovative technologies. [cited January 12, 2025]. . Available from: https://www.gmc-uk.org/professional-standards/learning-materials/artificial-intelligence-and-innovative-technologies . Accessed 11 Aug 2025. 73. Lee CC, Low MYH (2024) Using genAI in education: the case for critical thinking. Front Artif Intell 7:1452131 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 74. Jacobs SM, Lundy NN, Issenberg SB, Chandran L (2023) Reimagining core entrustable professional activities for undergraduate medical education in the era of artificial intelligence. JMIR Med Educ 9:e50903 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 75. García-Torres D, Vicente Ripoll MA, Fernández Peris C, Mira Solves JJ (2024) Enhancing clinical reasoning with virtual patients: a hybrid systematic review combining human reviewers and ChatGPT. Healthcare Basel 12(22):2241 ( MDPI ) [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 76. Saroha S (2025) Artificial intelligence in medical education: promise, pitfalls, and practical pathways. Adv Med Educ Pract. 10.2147/amep.s523255 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 77. Van De Werfhorst HG, Kessenich E, Geven S (2022) The digital divide in online education: inequality in digital readiness of students and schools. Comput Educ Open 3:100100.797 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 78. Accreditation Council for Graduate Medical Education (2019) Pathology milestones. Published February 2019. Accessed 15 Nov 2025 79. Beauchamp TL, Childress JF (2019) Principles of biomedical ethics [ DOI ] [ PubMed ] 80. Hochman G (2025) Epistemic humility and epistemic confidence: competing ethical forces in clinical medicine. SocArXiv . 10.31235/osf.io/m6bzr_v1 81. Gerke S, Minssen T, Cohen G (2020) Ethical and legal challenges of artificial intelligence-driven healthcare. Artificial intelligence in healthcare. Academic Press, pp 295–336 [ Google Scholar ] 82. Topol E (2019) Deep medicine: how artificial intelligence can make healthcare human again. Hachette, UK [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement No datasets were generated or analysed during the current study. Articles from Head and Neck Pathology are provided here courtesy of Humana Press ACTIONS View on publisher site PDF (2.3 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Assessment of an App Based on Artificial Intelligence for Purulent AOM Diagnosis in a Pediatric Department#982734
• Validation of the Utility of an Intelligent Visual Acuity Diagnostic System for Children#981608
• AI-Empowered Nuclear Medicine Education, Part 3: Practical AI Applications for Learners.#659540
• From Task Automation to AI-Mediated Judgment in Higher Education#610602
• Development and Demonstration of Intelligent Assessment Based on Multi-modal Information Fusion for Tumor Risk and Diagnosis and Treatment#1060300