TY - GEN
T1 - AI Pictures and the Thousand-Word Problem
T2 - A Feature-Level Reanalysis of User Perceptions of Visual News Summaries
AU - Stabauer, Martin
AU - Mittermayr, Julia Christine
PY - 2026/6/26
Y1 - 2026/6/26
N2 - This paper investigates which visual features contribute to the effectiveness of AI-generated news images and how these machine-produced representations align with human evaluations. Building on a previous survey in which participants rated AI-generated images illustrating news articles, we reanalyze the dataset using interpretable visual feature coding, regression models, and a “Reverse Engineering” approach in which ChatGPT evaluates the same images along the same dimensions. The results show strong and consistent effects across four outcome variables. Textual cues and visual realism substantially improve perceived clarity and interpretability, while visual complexity and symbolic clutter reliably reduce them. Human figures enhance content and quality evaluations but matter less once realism and text are available. Older participants evaluate images more critically, though demographic effects remain modest compared to feature-level influences. Reverse Engineering reveals that ChatGPT systematically rates images more positively than human respondents, suggesting that current models may overestimate the communicative adequacy of their own outputs. Taken together, the findings identify a stable set of design principles for effective AI-generated news visuals and point toward hybrid human–AI workflows that incorporate realism, textual scaffolding, tone alignment, and visual simplicity.
AB - This paper investigates which visual features contribute to the effectiveness of AI-generated news images and how these machine-produced representations align with human evaluations. Building on a previous survey in which participants rated AI-generated images illustrating news articles, we reanalyze the dataset using interpretable visual feature coding, regression models, and a “Reverse Engineering” approach in which ChatGPT evaluates the same images along the same dimensions. The results show strong and consistent effects across four outcome variables. Textual cues and visual realism substantially improve perceived clarity and interpretability, while visual complexity and symbolic clutter reliably reduce them. Human figures enhance content and quality evaluations but matter less once realism and text are available. Older participants evaluate images more critically, though demographic effects remain modest compared to feature-level influences. Reverse Engineering reveals that ChatGPT systematically rates images more positively than human respondents, suggesting that current models may overestimate the communicative adequacy of their own outputs. Taken together, the findings identify a stable set of design principles for effective AI-generated news visuals and point toward hybrid human–AI workflows that incorporate realism, textual scaffolding, tone alignment, and visual simplicity.
UR - https://www.scopus.com/pages/publications/105043941079
U2 - 10.1007/978-3-032-30047-8_15
DO - 10.1007/978-3-032-30047-8_15
M3 - Conference proceedings
SN - 978-3-032-30046-1
T3 - Lecture Notes in Computer Science
SP - 233
EP - 249
BT - HCI in Business, Government and Organizations - 13th International Conference, HCIBGO 2026, Held as Part of the 28th HCI International Conference, HCII 2026, Proceedings
A2 - Nah, Fiona Fui-Hoon
A2 - Siau, Keng Leng
PB - Springer
CY - Cham
ER -