Wenjuan Qiang, Xiaohong Deng, Tiezhou Hou, Le Qiang
In the no-direct-evidence slice defined by the supplied packet, closed-book models rarely abstained, with drug-level abstention ranging from 5.3% to 33.3%; the evidence-gated protocol required abstention, which all models followed for every no-direct-evidence drug. The same pattern held for recent or low-recognition drugs, where evidence-gated abstention reached 92.0% to 100.0% vs. 8.0% to 49.3% under closed-book answering. Closed-book models also produced high-confidence low-risk responses for DILI-positive drugs, a label-discordant pattern largely removed by evidence gating. Independent expert review of selected responses showed that label discordance did not always imply a clinically unreasonable low-risk category, but identified unsafe reassurance through overconfident wording and under-cautious responses in selected cases. When direct DILI evidence was provided, all models preserved citation-grounded non-abstaining answers. However, they differed in how often they committed to a conclusive rather than an uncertain risk category. Citation-bearing evidence-gated responses cited only the supplied PubMed identifiers and achieved 91.2% to 100.0% concordance with the supplied grade.
BACKGROUND: International undergraduate dental students often learn and communicate in a language other than their first language, affecting terminology acquisition, clinical reasoning, patient explanations, informed consent, clinical communication, and patient safety. AI-enabled language technologies-including generative artificial intelligence (GenAI), large language models (LLMs), and neural machine translation (NMT)-may support these tasks, but their use in this population has not been systematically mapped.
METHODS: Following the Arksey and O'Malley framework, JBI guidance, and PRISMA-ScR, PubMed, Scopus, and Web of Science were searched from inception. Searches were conducted on 1 March 2026 and updated on 28 March 2026, supplemented by reference-list screening and targeted policy and grey-literature searches. Sources were charted by context, population, technology, outcomes, limitations, risks, and implementation, and classified by relevance.
RESULTS: Thirty-six non-policy sources and four policy or governance documents were included. Five involved undergraduate dental students, three addressed international, multilingual, or limited-English-proficiency learners or relevant policies, eight concerned other health-professions contexts, and eight were translation or technical benchmarks without learners. Categories overlapped, and only one source combined undergraduate dental students, an international or multilingual population, and an AI-enabled language technology. Applications included terminology translation, multilingual tutoring, communication rehearsal, and reflective-writing support. No source assessed retained learning, transfer to authentic clinical encounters, or patient-level outcomes.
CONCLUSIONS: AI-enabled language technologies may support supervised language learning and patient-safety-oriented communication training, but direct evidence is sparse, short term, and largely derived from adjacent populations or technical benchmarks. Performance varies by language, task, prompt, model version, and context. Pending direct, comparative, longitudinal, and clinically situated evidence, implementation should include validated local resources, educator oversight, academic-integrity boundaries, privacy safeguards, and staged rehearsal before patient exposure.