- The paper shows that atomic fact-checking increases clinician trust by decomposing LLM recommendations into individually verifiable claims linked to guidelines.
- A rigorous RCT of 356 clinicians across 21 oncology cases revealed a substantial trust boost, with effect sizes far exceeding traditional methods.
- The findings suggest that verifiability through atomic fact-checking is a key design principle for enhancing clinical AI adoption in oncology.
Atomic Fact-Checking as a Mechanism to Enhance Clinician Trust in LLM-Driven Oncology Decision Support
Background and Motivation
The deployment of LLMs for oncology decision support is constrained not by technical accuracy but by limited clinician trust, particularly in high-stakes clinical workflows. Traditional transparency strategies focus on natural language explanations and guideline source citations, operating under the premise that revealing AI reasoning pathways and evidence sources enhances trust calibration. However, behavioral data from clinical settings indicate inconsistent benefits from these transparency methods, with some explainability techniques even reducing appropriate acceptance rates among domain experts. This gap mandates the exploration of alternative mechanisms grounded in cognitive and workflow validity rather than technical introspection.
Atomic Fact-Checking (AFC) operationalizes a paradigm shift: instead of holistic reasoning appraisal, AFC atomizes LLM recommendations into discrete, independently verifiable factual claims, each linked directly to authoritative guideline passages. This shifts clinician interaction from reflection on AI logic to itemized correspondence checking, fundamentally altering the cognitive demands of verification and trust formation.
Methods
A randomized controlled trial (RCT) was implemented involving 356 board-certified clinicians (radiologists, radiation oncologists, medical oncologists) who provided 7,476 trust ratings across 21 curated oncology cases encompassing seven cancer entities. The trial compared five presentation conditions for LLM-generated recommendations: (1) bare recommendations; (2) recommendations with explanations; (3) recommendations with guideline citations; (4) recommendations with both explanations and citations; (5) recommendations augmented with AFC (explanation, citation, atomic fact decomposition).
Randomization was stratified by specialty to ensure distributional parity. Trust, the primary endpoint, was measured using a validated 5-point Likert scale immediately after assessment of each AI recommendation. The analytic framework utilized linear mixed-effects modeling to account for repeated measures and case/participant heterogeneity, with effect sizes captured as Cohen's d.
Key Results
Quantitative Findings
AFC produced a large, statistically robust increase in clinician trust relative to all control arms:
- Mean trust score with AFC: 3.80 (SD: 0.82), compared to 2.59–3.09 across all traditional transparency methods.
- Cohen's d (AFC vs. pooled controls): 0.94 (95% CI: 0.88–1.00; P < .001), far exceeding the "large" convention in behavioral science.
- Proportion expressing trust (score ≥4): 66.5% with AFC (935/1,407) vs. 26.9% in control arms, an absolute increase of 39.5 percentage points (NNT = 2.53).
- Dose-response gradient observed: Explanations and citations alone produced incremental but small-to-moderate effect sizes (d = 0.25–0.50) but did not meaningfully approach AFC's effect magnitude.
These effects proved consistent across specialties (d = 0.80–1.03), cancer types (d = 0.80–1.10), and levels of clinical and AI/LLM experience. Residents/fellows showed the most pronounced absolute trust increases, but even highly experienced clinicians and frequent LLM users exhibited significant AFC-driven trust gains. Response distributions in AFC arms were strongly right-skewed with marked ceiling effects, confirming not only statistical but practical significance.
Control and Subgroup Analyses
Pairwise contrasts among control arms demonstrated that citations meaningfully outperformed explanations, with the two in combination providing additive effects, but still representing only modest improvements over opaque LLM outputs. The superiority of AFC (d = 0.78 vs. the next-best method) indicates that decompositional verification, not just citation or rationale, is the primary driver.
Subgroup analyses showed that clinicians with limited prior AI exposure were less responsive, suggesting an interaction between baseline AI awareness and the capacity to benefit from AFC. However, these analyses should be interpreted cautiously due to small sample sizes in some subgroups and lack of direct behavioral endpoints.
Theoretical and Practical Implications
Cognitive and Human Factors
The effectiveness of AFC is attributable to its explicit reduction of cognitive load. Classical explainability burdens clinicians with the task of validating AI logic holistically—a high-complexity, low-yield activity in the context of intricate guideline-driven medicine. AFC replaces this with rapid fact-matching, directly aligning with existing guideline-centric verification strategies entrenched in oncology. This reframing of transparency as verifiability rather than explainability leverages domain expertise and reduces the risk of both over- and under-trust.
Impact on Clinical AI Adoption
These findings have significant implications for the design of clinical AI platforms:
- Design Principle Shift: Interface-level prioritization of verifiability (AFC) should supersede isolated investments in explainability for AI intended for primary recommendation support.
- Grounding in Evidence: Direct linkage to validated clinical guidelines enables responsible AI augmentation, enhancing acceptability without sacrificing accountability.
- Augmented Intelligence Model: AFC operationalizes human-AI collaboration, maximizing the strengths of both machine synthesis and expert confirmation.
Limitations and Future Research
While the effect on trust is pronounced, several limitations temper direct clinical translation:
- The study relies on self-reported trust in a vignette-based survey rather than observed real-world behavior. The intention-behavior gap is non-trivial in operational clinical settings with competing demands and risk aversion.
- All AI recommendations were pre-validated, precluding assessment of trust calibration under error or disagreement conditions. Overtrust is a potential risk; future evaluation must address whether AFC artificially inflates trust when recommendations deviate from ground truth or best practice.
- Generalizability outside North American/European academic contexts and to additional specialties is uncertain.
Conclusion
Atomic fact-checking produces a substantially larger increase in clinician trust in LLM-generated oncology recommendations than traditional transparency methods, including explanations and source citations. By decomposing outputs into verifiable claims with direct guideline linkage, AFC provides a cognitively tractable pathway for calibrated trust, with consistent effects across specialty, experience, and clinical scenarios. The findings position verifiability—not explainability—as the dominant design axis for clinical AI intended for integration into high-stakes workflows.
Further research is required to validate AFC-based trust mechanisms in live clinical environments, under error conditions, and across additional domains. However, the current quantitative evidence robustly motivates the prioritization of AFC in AI-augmented decision support system design for oncology and likely beyond.
