Expert Evaluation of AI-Generated Cultural Heritage Imagery: A Dataset of 84 Images from Seven Generative AI Systems Across Twelve Heritage Elements

Description

This dataset contains expert evaluation scores for 84 AI-generated images of twelve cultural heritage elements (spanning Ancient Egyptian, Greek, Roman, Islamic, Romanesque, Gothic, Renaissance, Plateresque, and Modernist traditions), each produced by seven widely used generative AI systems (Canva, Perplexity, Adobe Firefly, ChatGPT, Gemini, Grok, and Copilot) via zero-shot prompting. An interdisciplinary panel of experts in archaeology, art history, architecture, and heritage education rated each image on a 12-item, 5-dimension Likert-scale instrument (historical veracity, formal coherence, stylistic adequacy, contextual fit, and didactic reliability), following the methodological protocol proposed by Galindo-Durán, Prego de Oliver-López & Bernal-Bravo (2026, *Journal of Cultural Heritage*). Data are anonymized and provided in long format with an accompanying codebook.

Steps to reproduce

Steps to Reproduce Twelve heritage elements spanning different periods and cultural traditions were selected as generation targets (Doric capital, Pointed arch, Egyptian obelisk, Gothic rose window, Islamic horseshoe archway, Romanesque cathedral tower, Renaissance façade, Plateresque façade, Modernist façade, Lighthouse of Alexandria, Piece of sigillata, Iberian ceramic). For each heritage element, a single English-language prompt was submitted, under zero-shot conditions (no prior examples, no iterative correction), to each of seven generative AI image systems: Canva AI Image Generator (3.2), Perplexity (Ask v3, Image Engine 2025), Adobe Firefly (Model 3), ChatGPT (GPT-5.1 + DALL·E 4.1), Gemini (2.0 Advanced, Nanobanana Vision), Grok (Grok-2 Vision), and Copilot Designer (2025), all accessed in November 2025. Only the first image generated by each system was retained, with no post-generation edits, under each platform's default settings (1:1 aspect ratio, default style, no negative prompts), producing a corpus of 12 × 7 = 84 images. An interdisciplinary panel of experts in archaeology, art history, architecture, and heritage education independently evaluated the images via an online questionnaire (Google Forms). Each expert rated every image assigned to them on the 12-item, 5-point Likert scale (1 = very low; 5 = very high) described in the Items_Codebook sheet, with a "Not applicable" option available per item. Not every expert rated every heritage element (voluntary interdisciplinary panel), resulting in an incomplete but robust cross-classification design with 6 to 11 respondents per element (11 experts in total). Raw responses were exported, expert identities were replaced with anonymous codes (E01-E11), "Not applicable" entries were recoded as missing (NA), and the data were restructured into the long format provided in the Raw_Data sheet, following the methodological protocol of Galindo-Durán, Prego de Oliver-López & Bernal-Bravo (2026, Journal of Cultural Heritage, https://doi.org/10.1016/j.culher.2026.08.010).

Institutions

• University of Almería

  Andalusia, Almería
• Universidad Rey Juan Carlos

  Madrid, Madrid

Categories

Related Links

Licence