Can AI Predict Publication? Multimodal Large Language Models and the Structural Determinants of Surgical Scholarship

BackgroundWhether artificial intelligence can identify publishable scientific work is untested. We evaluated whether a multimodal large language model (MLLM) could predict, from poster content alone, which abstracts at the American Association for the Surgery of Trauma (AAST) Annual Meetings reached publication, and characterized the investigator, institutional, and domain level determinants situating model performance.MethodsWe retrospectively analyzed 260 abstracts from the 2021-2022 AAST Annual Meetings. Bibliographic searches confirmed publication, and investigator sex, race, and ethnicity were inferred from public data. Variables were compared by t-tests and Pearson χ² or Fisher exact tests, with multivariable logistic regression. GPT-4.1 scored poster images on a six-domain rubric averaged over 30 iterations.ResultsGPT-4.1 predicted publication with 58.5% accuracy overall (P = .009), but performance was domain-dependent, 74.2% in Violence, Societal, and Behavioral and 62.2% in Hemorrhage, Resuscitation, and Vascular Control, falling to chance in Critical Care and Outcomes and Systems, Technology, and Process Optimization. 142 (54.6%) reached publication at a mean of 13.4 months. Multicenter origin was the only independent predictor (P = .02). Hispanic investigators were underrepresented among first (P = .03) and senior (P = .01) authors, absent from the published Hemorrhage, Resuscitation, and Vascular Control subset.DiscussionThe model predicted publication where scientific content carried the signal and fell to chance where advancement depended on institutional factors absent from the poster-multicenter scaffolding, mentorship, and senior author fluency. Publication is determined jointly by what is legible on the page and structural advantage the model cannot read.