Stable diffusion models reveal a persisting human and AI gap in visual creativity

Referee: [Abstract and Methods] Abstract and Methods: The reported creativity gradient and the claim of a persisting human-AI gap rest on subjective ratings, yet the manuscript provides no inter-rater reliability statistics (e.g., Cronbach’s alpha, ICC, or Fleiss’ kappa), no explicit operational definition or rating criteria for creativity, and no controls for potential confounds such as technical execution quality or stylistic familiarity. Without these, it is unclear whether the observed ordering reflects creativity differences or systematic rater biases against AI-generated images.

Authors: We agree that these elements are necessary for robust interpretation of subjective ratings. In the revised manuscript we will add Cronbach’s alpha and ICC values computed across the 255 human raters in the Methods section. We will also insert an explicit operational definition of visual creativity, grounded in established criteria of originality, novelty, and contextual appropriateness. To address potential confounds, we will expand the Methods to detail the anonymous, randomized presentation of images (which reduces stylistic familiarity effects) and add a dedicated limitations paragraph discussing the possibility of execution-quality biases. We note that the observed gradient remained stable across multiple rating conditions, which is consistent with prior creativity research, but we will not claim this fully rules out bias without additional controls. revision: yes

Referee: [Results] Results: The abstract states that human and AI raters showed “vastly different creativity judgment patterns,” but the manuscript does not report the specific statistical comparisons, effect sizes, or agreement metrics between the two rater groups. This information is necessary to evaluate whether the differing patterns support the interpretation of uniquely human perceptual capacities.

Authors: We will revise the Results section to report the requested statistics. This includes mean differences between human and GPT-4o ratings with accompanying t-tests or mixed-effects models, effect sizes (Cohen’s d), and agreement metrics such as Pearson correlations and, where ratings permit, Cohen’s kappa or intraclass correlations between the two rater groups. We will also describe qualitative differences in judgment patterns (e.g., which image attributes each group weighted more heavily) to support the claim of distinct evaluative criteria. revision: yes

Referee: [Discussion] Discussion: The conclusion that GenAI models face unique challenges in visual domains because creativity depends on “perceptual nuance and contextual sensitivity” assumes the ratings isolate these capacities. However, absent bias controls or validation against objective creativity markers, alternative explanations (e.g., raters penalizing AI images for lacking human-like execution cues) remain viable and would weaken the contrast to language-centered tasks.

Authors: We accept that subjective ratings cannot fully isolate perceptual nuance from other cues and that alternative explanations remain plausible. In the revised Discussion we will explicitly enumerate these alternatives, including execution-cue penalties, and temper our claims accordingly while still highlighting the human–AI rater disagreement as evidence of differing judgment bases. We will reference perceptual-processing literature to support the contrast with language tasks. Because the current dataset does not contain separate objective creativity markers or execution-quality ratings, we cannot add post-hoc validation analyses; we will instead frame this as a limitation and propose it as a target for future studies. revision: partial