Teaching the Atomic Sentence Method for Source-Verified, AI-Assisted Literature Synthesis to Clinical Health Care Professionals: Single-Cohort Feasibility and Acceptability Study
Open Access
- 27 August 2026
- journal article
- Published by JMIR Publications Inc. in JMIR Formative Research
- Vol. 10, e98343
Abstract
Background: Generative AI can reduce the academic-writing burden on clinical health care professionals, but unsupervised use introduces citation hallucination (the confident fabrication or misattribution of references), which threatens research integrity. When a machine invents a source, it is termed “hallucination,” and when a person does it, it is termed “fabrication,” yet both are equally unacceptable. Existing health professions education writing workshops have rarely translated this concern into a concrete, reproducible source-verification procedure. Objective: This study aimed to describe the development and delivery of a 2-day workshop teaching the Atomic Sentence method, a source-anchored knowledge-modeling technique for AI-assisted literature synthesis, and evaluate its feasibility, acceptability, and short-term effect on research-idea development among clinical health care professionals. Methods: We conducted a single-cohort educational program evaluation of a 2-day workshop (≈16 contact hours) for 18 health care professionals across 4 campuses of a Buddhist medical network in Taiwan. Curriculum development used the ADDIE (Analysis, Design, Development, Implementation, and Evaluation) model; outcomes were framed with the Kirkpatrick model (levels 1‐2). The workshop implemented a published 7-step AI-assisted research workflow (topic exploration, literature search, knowledge-base management, reading, synthesis, writing, and peer-review simulation). Citation integrity was protected by confining citation generation to a source-grounded tool (NotebookLM) combined with atomic-sentence extraction and mandatory cross-reference verification; other, nongrounded tools supported discovery, reading, and writing. Feasibility was assessed by the workshop completion rate and the questionnaire response rate. Outcomes were an acceptability questionnaire covering 5 domains (usefulness of instructional materials, curriculum planning, time allocation, personal-goal attainment, and administrative support; each rated 0‐10) and a pre/post research-topic-transformation analysis (4 categories; 2 independent coders, Cohen κ); responses to an open-ended item on improvement suggestions were analyzed by qualitative content analysis. Reporting follows the GAMER (Guidance for AI Use in Medical Education Reporting) guidance for AI use in medical education. Results: All 18 (100%) participants completed the workshop and evaluation. Acceptability was high across all 5 domains (each median 10; overall median 10, IQR 9.3-10). The most variable domain was time allocation (range 5-10). Short-term improvement in the research topic (substantial transformation or refinement) occurred in 11 (61%; κ=0.77) participants; 2 (11%) had no structured question before or after. All 18 (100%) would recommend the workshop. Conclusions: A short workshop teaching the Atomic Sentence method was feasible and well accepted and was associated with short-term research-idea development in a multidisciplinary clinical cohort. Whether the method reduces citation hallucination or improves citation accuracy was not tested and requires controlled studies with objective outcomes, blinded raters, and longitudinal follow-up.Keywords
This publication has 26 references indexed in Scilit:
- Hallucinated citations produced by generative artificial intelligence may constitute research misconduct when citations function as data in scholarly papersAccountability in Research, 2026
- Reporting guideline for the use of Generative Artificial intelligence tools in MEdical Research: the GAMER StatementBMJ Evidence-Based Medicine, 2025
- The ethics of ChatGPT in medicine and healthcare: a systematic review on Large Language Models (LLMs)npj Digital Medicine, 2024
- AI and science: what 1,600 researchers thinkNature, 2023
- Fabrication and errors in the bibliographic citations generated by ChatGPTScientific Reports, 2023
- Learning to Fake It: Limited Responses and Fabricated References Provided by ChatGPT for Medical QuestionsMayo Clinic Proceedings: Digital Health, 2023
- On the Dangers of Stochastic Parrots2021
- Going Beyond “Not Enough Time”: Barriers to Preparing Manuscripts for Academic Medical JournalsTeaching and Learning in Medicine, 2019
- Instructional Design: The ADDIE Approach2009
- The qualitative content analysis processJournal of Advanced Nursing, 2008