JMIR Publications

ABSTRACT

Background:

Generative artificial intelligence (GenAI) holds considerable promise for reducing the academic writing burden on clinical healthcare professionals; however, its deployment without structured oversight introduces citation hallucination—the confident fabrication or misattribution of references—as a critical threat to research integrity. Existing workshop interventions in health professions education (HPE) have not systematically addressed hallucination prevention as a core pedagogical objective.

Objective:

This study describes the development, delivery, and evaluation of a 2-day intensive workshop introducing the Atomic Sentence (AS) method as a structured pedagogical intervention to mitigate AI citation hallucination in clinical academic writing.

Methods:

We designed and evaluated a 2-day academic writing workshop for 18 healthcare professionals across four campuses of a Buddhist medical network in Taiwan. The workshop introduced the AS method—a structured technique in which participants decompose literature into semantically complete, independently readable, and source-traceable knowledge units. GenAI was used exclusively with source-anchored tools, and all AI-generated output was subject to mandatory cross-reference verification. The ADDIE instructional design model organized curriculum development, while the Kirkpatrick evaluation framework guided data collection. Post-workshop outcomes included a 5-domain, 0–10 satisfaction scale and a pre/post research topic transformation analysis using a 4-category classification scheme.

Results:

All 18 participants completed the evaluation (response rate: 100%). Composite satisfaction was 9.63/10 (domain range: 9.22–9.94). Instructional materials and administrative services achieved ceiling-level ratings (M=9.94, SD=0.24). Time allocation showed the highest variance (M=9.22, SD=1.44, range: 5–10), consistent with cognitive load peaks during source verification. Research topic transformation was observed in 12 of 18 participants (66.7%): 8 (44.4%) achieved substantial transformation from vague clinical observations to PICO-structured questions, and 4 (22.2%) refined existing topics by integrating theoretical frameworks or explicit comparison groups. All 18 participants (100%) recommended the workshop.

Conclusions:

The Atomic Sentence method operationalizes citation integrity as a procedural safeguard rather than an ethical exhortation, aligning with ICMJE and WAME mandates that authors bear full responsibility for the accuracy of AI-assisted content. The method is replicable, applicable across research paradigms, and addresses a documented gap in HPE research writing education. Future studies should incorporate objective pre/post knowledge assessments and 3 to 6-month behavioral follow-up to establish Kirkpatrick Level 2–4 evidence.