Granice wiarygodności AI w cyfrowej narracji Grodziska w Sopocie

Abstract

Objective: This article examines the reliability of generative artificial intelligence in visualizing the early medieval past, using the Sopot stronghold as a case study. It asks whether increasingly specialized prompts improve historical accuracy and how AI-generated historical images are perceived by museum audiences. Method: A qualitative comparative experiment used a four-level prompt matrix across three topics–stronghold architecture, amber craftsmanship, and dress and weaponry–and three tools: ChatGPT (DALL·E 3), Gemini 3 (Nano Banana 2), and Adobe Firefly (Image 4). The design comprised 36 planned combinations. Several images were generated for each combination without changing the prompt, and the result considered the best within that series was selected for analysis. The assessment sheet contains 37 records because two specialist Firefly variants for dress and weaponry were retained. No additional reference materials were supplied to the models. Both authors first assessed the selected images independently using the same criteria, then compared their observations and agreed on the final assessments after jointly re-examining the images. The experiment was complemented by a pilot audience study (N = 32). Results: Increasing prompt specialization did not produce a consistent improvement in historical accuracy. Performance depended on both the tool and the task. Architectural scenes were generally easier to render than technological processes and typological details of dress and weaponry. Visually convincing images could still contain anachronisms and incorrect elements of material culture. In the audience study, photographs of reconstructions were considered more credible than AI-generated visualizations. Conclusions: A specialist prompt should not be treated as a guarantee of archaeological accuracy. In museum practice, generative AI can support prototyping, education, and alternative interpretations, but images that fill gaps in the archaeological record should be clearly labelled, documented, and subject to expert review. In such cases, they are better treated as visual hypotheses than automatically as reconstructions.