AI assistance can make scientific diagrams more explicit while weakening their visual cohesion, according to model-based evaluations reported in a new preprint. Among 369 revisions that passed recorded fidelity checks, clarity ratings rose by an average of 0.428 points and aesthetic coherence ratings fell by 0.111 points on five-point scales. These results describe one evaluation workflow, rather than demonstrated improvements in reader understanding.
Xiaoqi Zhao, Youwei Pang and Kailai Zhou of Nanyang Technological University, Singapore, examined 371 figures from 200 papers. Their arXiv preprint was posted on November 5, 2026; peer review has not been verified. The study distinguishes motivation figures, which introduce a problem or contrast, from framework figures, which explain a method’s overall organization.
The main sample contained 160 papers and 320 figures, balanced across four periods and four computer-vision topics: object detection, image segmentation, image restoration and 3D vision. A separate sample of 40 papers contributed 51 figures deliberately selected for visual conventions. That sample was excluded from temporal averages. Ratings of original figures showed no uniform decline across periods, but the comparison cannot establish whether AI adoption changed scientific illustration: the original authors’ use of AI was generally unknown.
A single multimodal model, gpt-6-astra, rated clarity, organization and readability, graphical consistency, aesthetic coherence, purposeful visual inventiveness and design specificity. Each figure received three fresh judgments per version, summarized by the median for each dimension. External captions and source identifiers were withheld during scoring, although text inside figures remained visible. Evaluation took place on October 31–November 1, 2026.
Revisions used gpt-image-2.5, source captions, manuscript explanations and feedback from the evaluator. Protected scientific content, including numerical values and data-bearing image regions, was restored from the originals where necessary. Of 371 attempts, two failed fidelity checks because scientific connections were incorrect or missing; neither received revised quality scores. Passing these model-based checks did not amount to independent expert verification.
The reported increase in a post hoc average of all six dimensions was larger for motivation figures than framework figures in the main sample. Numerical subgroup effects were not available in the supplied material. Case analyses suggest that added explanations can clarify comparisons while leaving typography, hierarchy and visual economy unresolved. Because revisions combined added information, layout changes and rendering, the study cannot isolate which component produced a rating change.
Several limitations constrain interpretation. Sampling was not probabilistic, period lengths and source populations differed, and eligibility rules were relaxed for 23 of the latest period’s 40 papers. The same model supported collection, revision feedback, fidelity review and scoring in separate sessions. Human comprehension, human preferences and cross-model agreement were not assessed; repeated judgments do not provide independent validation.
The rubric and equal-weight overall average are unvalidated, and the scores are ordinal. The methods specify 95% intervals from 5,000 paper-level bootstrap resamples that preserve dependence between figures from the same paper, but numerical intervals for the headline effects are unavailable in the supplied text. Such intervals would not resolve selection or evaluator bias. The authors advocate researcher-led design, with AI helping explore alternatives and authors retaining responsibility for explanatory purpose and scientific relationships.
Cover illustration generated with AI; conceptual, not research evidence.