Drivers of Oncologist Preference of AI-Generated Literature Review in a Randomized Mixed-Methods Study

Technical Summary: Drivers of Oncologist Preference of AI-Generated Literature Review in a Randomized Mixed-Methods Study

Problem Statement
While Large Language Models (LLMs) are increasingly adopted by physicians for clinical reasoning and administrative tasks, the specific design features that render AI-generated literature reviews useful and trustworthy remain unclear. Previous evaluations suggest that model accuracy alone does not guarantee clinical utility; in some cases, physician-AI collaboration performs worse than AI alone. In oncology, where evidence evolves rapidly and decisions are high-stakes, there is a critical gap in understanding how clinicians evaluate, verify, and incorporate AI-generated reports. Specifically, it is unknown whether modifying report structures to explicitly highlight evidence hierarchy (e.g., ranking studies by strength) improves perceived utility, or if such design changes inadvertently degrade trust and workflow fit.

Methodology
The authors conducted a remote, within-participant, randomized mixed-methods study involving 34 oncology physicians (fellows, residents, and attendings) from multiple institutions, primarily Stanford and UCLA.

• Study Design: Participants reviewed five oncology clinical vignettes. For each vignette, they evaluated four blinded AI-generated literature review reports:
  1. Standard OpenEvidence: A widely used clinical AI tool.
  2. Modified OpenEvidence: A report generated by re-prompting a model (GPT-5.2 Thinking) to reorganize the existing OpenEvidence output (without supplementing it with outside information) to prioritize references by evidence strength (guidelines and large trials) and explicitly grade evidence.
  3. ChatGPT: A general-purpose LLM (GPT-5.2 Thinking).
  4. DoximityGPT: A general-purpose LLM (DoxGPT).
• Data Collection:
  • Quantitative: Participants rated each report on a 7-point Likert scale for overall utility and six subdomains (clarity, confidence in action, citation quality, order/flow, uncertainty communication, and length). They also selected the "best" and "worst" reports and indicated preferred report length.
  • Qualitative: 20 semi-structured interviews were conducted (some with multiple participants), totaling nearly 6 hours of audio. These were analyzed using a prespecified, LLM-assisted qualitative pipeline. Transcripts were processed into structured JSON files to extract themes, likes/dislikes, and design requirements, which were then synthesized into a cross-interview knowledge base.
• Analysis: Quantitative data were analyzed using linear mixed-effects models with random intercepts for participants and vignettes to account for clustering. Qualitative data were synthesized to identify recurring design requirements and trust drivers.

Key Results

• Quantitative Findings: Contrary to the hypothesis that an evidence-graded, reorganized report would improve utility, the Modified OpenEvidence report was rated significantly lower in overall utility than the Standard OpenEvidence report (mean difference: −0.96; 95% CI, −1.26 to −0.66; P<.001). The modified report also scored lower across all six subdomains, with the largest deficits in communication of uncertainty, order/flow, and clarity.
  • Standard OpenEvidence performed similarly to ChatGPT and DoximityGPT in overall utility but significantly outperformed both in citation quality.
  • ChatGPT had significantly lower citation quality ratings than OpenEvidence.
  • DoximityGPT received improved ratings for length but lower ratings for citation quality.
• Qualitative Findings: Analysis of 20 interviews identified six overarching themes and seven specific design requirements. Oncologists prioritized:
  • Rapid Orientation: Concise, top-loaded summaries suitable for time-constrained workflows.
  • Verifiability: Direct, clickable citations to specific guideline sections or trial data; recognizable guideline sources (e.g., NCCN, ASCO) were crucial.
  • Evidence Transparency: Explicit communication of uncertainty and separation of evidence presentation from AI recommendations.
  • Formatting: Scannable layouts with selective tables and bolded guidelines; avoidance of dense paragraphs or "AI-sounding" editorial tones.
  • Trust Erosion: Trust deteriorated rapidly with citation mismatches, buried provenance, evidence misclassification, and overconfident recommendations.

Key Contributions

1. Design over Accuracy: The study provides empirical evidence that report design and presentation significantly influence perceived clinical utility, independent of the underlying model's factual accuracy. A report with identical source data but a different structure (evidence-graded) was perceived as less useful, suggesting that "being right" is insufficient for adoption.
2. Specific Design Requirements: The paper delineates concrete design requirements for clinical AI, including the need for "traceable text" (direct links to source passages), explicit uncertainty quantification, and the separation of literature review from AI-generated recommendations.
3. Methodological Innovation: The authors demonstrate a prespecified, LLM-assisted qualitative analysis workflow. By converting interview transcripts into structured JSON and synthesizing them into a machine-readable knowledge base, they offer a reproducible method for aggregating large volumes of qualitative physician feedback while maintaining auditability.

Significance and Claims
The authors claim that for clinical AI to be effective, it must be designed as a tool for rapid orientation, evidence retrieval, and verification rather than an autonomous decision-maker. The study argues that the "form implies function" in clinical AI; the layout and hierarchy of an output shape whether a clinician treats it as a trusted assistant or a risky recommendation engine.

The paper posits that trust is a design feature. Features such as visible guideline anchoring, bolded evidence sources, and direct hyperlinks are not merely cosmetic but are core components of clinical usefulness. The failure of the modified report (which attempted to foreground evidence strength) suggests that current attempts to visualize evidence hierarchy may inadvertently compromise usability if they disrupt the flow or clarity of the report.

Ultimately, the authors conclude that clinical AI evaluations must move beyond benchmarking model accuracy to empirically evaluating report design and physician-AI interaction. They emphasize that effective collaboration depends on how AI is integrated into human workflows, noting that "presentation cues can influence perceived expertise" analogous to the "white coat effect" in medicine. The study calls for future work to isolate specific interface elements to determine their causal impact on clinical verification behavior and decision quality.