Beyond Explainability: The Trust Paradox of Conversational Artificial Intelligence in Endodontics Authors: Mohammed Turky, Paul MH. Dummer Published: 2026 Artificial intelligence (AI) is rapidly reshaping contemporary dental practice, offering unprecedented opportunities to enhance diagnostic precision, treatment planning, workflow efficiency, and personalised patient care.1,2 In endodontics, AI-driven applications have shown considerable promise across a range of clinical domains, including radiographic interpretation, detection of periapical pathosis, root canal morphology analysis, and prediction of treatment outcomes.3 As these technologies increasingly transition from experimental research to clinical implementation, attention has appropriately shifted from algorithmic performance alone toward issues of transparency, interpretability, and clinician trust. In this context, explainable artificial intelligence (XAI) has emerged as an important paradigm, aiming to make AI-driven recommendations more interpretable, auditable, and clinically accountable.4,5 However, the rapid emergence of large language model (LLM)-based conversational AI introduces a distinct challenge that extends beyond model explainability alone: whether highly fluent, synthesised, yet potentially untraceable AI-generated clinical information can be trusted as a foundation for evidence-based decision-making. Ienca et al. present a thought-provoking trust paradox that highlights an unsettling trend in contemporary information environments: as traditional, rigorous institutions increasingly struggle to maintain their credibility, unaccountable digital voices are gaining disproportionate public trust.6 This paradox is particularly pertinent in the context of LLM-based conversational AI, which represents a significant evolution of the trust dilemma that extends beyond the established dynamics of social media platforms. In the realm of traditional social media, users are typically confronted with a range of competing claims made by identifiable sources.6 While these claims may not always be fully credible, the structure of social media allows for some degree of critical comparison and analysis, albeit imperfectly. In contrast, conversational AI powered by LLMs reduces much of the epistemic friction associated with information sourcing.6 These LLM-driven chatbots can generate fluent and synthesised responses by drawing on a vast array of data sources, including scholarly literature, commercial websites, user-generated online forums, and educational materials.7 However, this synthesis often occurs without adequate attribution, obscuring the evidentiary basis behind each assertion made. The implications of this situation extend beyond mere issues of content authenticity; they also raise significant concerns regarding the architectural design of these AI systems. Three critical safeguards that are foundational to evidence-based healthcare practice are frequently absent by default in conversational AI outputs6:1.Provenance: This involves the transparent attribution of sources along with contextual information about their publication, ensuring users can trace the origins of the claims made.2.Uncertainty Calibration: This is the clear communication of the limitations and varying confidence levels associated with the evidence, which is vital for informed decision-making.3.Evidence Trails: This refers to the capability for users to inspect and trace the connections between clinical claims and the supporting data that substantiates them. 1.Provenance: This involves the transparent attribution of sources along with contextual information about their publication, ensuring users can trace the origins of the claims made. Provenance: This involves the transparent attribution of sources along with contextual information about their publication, ensuring users can trace the origins of the claims made. 2.Uncertainty Calibration: This is the clear communication of the limitations and varying confidence levels associated with the evidence, which is vital for informed decision-making. Uncertainty Calibration: This is the clear communication of the limitations and varying confidence levels associated with the evidence, which is vital for informed decision-making. 3.Evidence Trails: This refers to the capability for users to inspect and trace the connections between clinical claims and the supporting data that substantiates them. Evidence Trails: This refers to the capability for users to inspect and trace the connections between clinical claims and the supporting data that substantiates them. The lack of these three safeguards creates an environment where conversational AI systems can project an unwarranted sense of authority, which potentially erodes the transparency essential for fostering professional trust in healthcare contexts. Additionally, LLMs tend to produce "hallucinations"—outputs that, despite being coherent in language and presentation, are factually inaccurate.8 In healthcare settings, these hallucinations can manifest as non-existent references, outdated therapeutic guidelines, or inappropriate clinical recommendations.9 This is particularly concerning because, unlike misinformation seen on social media—which typically can be traced back to identifiable human sources—these hallucinated outputs often lack any clear origin or point of accountability, thereby creating a novel and alarming category of untraceable health misinformation. These issues are especially relevant in the field of endodontics, where clinical decision-making often relies on specialised knowledge that necessitates a nuanced interpretation of diagnostic findings. Additionally, the evolving body of evidence surrounding minimally invasive techniques requires practitioners to engage in context-sensitive treatment planning. The risk posed by conversational AI's fluent yet unverifiable recommendations creates a deceptive façade of expertise, thereby obscuring the foundational evidence-based clinical decisions. To effectively address this emerging trust paradox, it is crucial to advance beyond the concept of mere explainability and strive for a higher form of epistemic accountability within conversational AI systems. This necessitates the following10:1.Mandatory source attribution that ensures users are aware of the origins of the information presented.2.Explicit signalling of uncertainty, so users are informed about the limitations and reliability of the claims being made.3.The implementation of independent safety audits of health-oriented conversational AI systems before they are widely adopted in clinical education and patient decision-support contexts. 1.Mandatory source attribution that ensures users are aware of the origins of the information presented. Mandatory source attribution that ensures users are aware of the origins of the information presented. 2.Explicit signalling of uncertainty, so users are informed about the limitations and reliability of the claims being made. Explicit signalling of uncertainty, so users are informed about the limitations and reliability of the claims being made. 3.The implementation of independent safety audits of health-oriented conversational AI systems before they are widely adopted in clinical education and patient decision-support contexts. The implementation of independent safety audits of health-oriented conversational AI systems before they are widely adopted in clinical education and patient decision-support contexts. These measures are essential to restore trust in healthcare decision-making processes and to mitigate the risks associated with the integration of conversational AI into clinical practice.