TLDR: A study on ChatGPT4o’s image generation reveals significant biases in its depiction of museum curators. AI-generated images overwhelmingly feature young, male, and Caucasian individuals, sharply contrasting real-world demographics where women and older professionals are prevalent. While avoiding some traditional stereotypes, the AI introduces new ones, portraying curators as fashionable ‘yuppie professionals’ with specific attributes like beards and clipboards, raising concerns about the perpetuation of inaccurate professional portrayals.
A recent study titled “Draw me a curator” Examining the visual stereotyping of a cultural services profession by generative AI, delves into how popular generative Artificial Intelligence (AI) models, specifically ChatGPT4o, visually depict museum curators. The research, conducted by Dirk HR Spennemann, reveals significant biases in these AI-generated images, which often contradict real-world demographics of the profession. You can find the full research paper here: Research Paper.
Curators play a vital, often behind-the-scenes, role in preserving cultural heritage and creating exhibitions. While popular culture sometimes stereotypes them as ‘nerdy’ or ‘old-fashioned,’ this study aimed to see if AI-generated visuals perpetuate these or other stereotypes.
How the Study Was Conducted
The researcher prompted ChatGPT4o to generate 230 images of typical curators across seven different museum types: art, fashion, natural history, social history, maritime, science, and technology. The prompts were intentionally broad to avoid influencing the AI’s output. Each image was then analyzed for characteristics such as gender, age, attire, and the presence of stereotypical items like beards, glasses, clipboards, or digital tablets. The background museum settings were also examined.
Key Findings: A Skewed Reality
The study uncovered several striking discrepancies between AI-generated depictions and actual curator demographics:
Gender Imbalance: A staggering 96.5% of the AI-generated curators were men. Only a handful of women were depicted (3.5%), all of whom were young and professionally dressed. This sharply contrasts with real-world statistics, where women constitute a significant majority of museum professionals in many English-speaking countries (e.g., 49% in the UK, 58.6% in the USA, and up to 72% in Australia).
Youthful Bias: The AI predominantly portrayed curators as young, with 79% falling into this age group. This is a stark difference from reality, where a large portion of curators are middle-aged or older (e.g., over 50% middle-aged in Australia, and a growing number of older curators in the USA).
Ethnic Homogeneity: All AI-generated curators were depicted as Caucasian. This is significantly higher than actual representation in countries like the USA (82.6% Caucasian) and the UK (64% white museum staff), suggesting a bias in the AI’s training data, likely derived from predominantly English-speaking sources.
Stereotypical Attributes: While the AI didn’t always show curators as ‘nerdy’ or ‘stuck in time,’ it did consistently include certain stereotypical attributes. Beards were common, especially in art, natural history, and maritime museum curators. Clipboards were frequently seen in art and fashion museum settings, while tablets were popular in technology museums. Name tags were common in science, natural history, and maritime museums but absent in art and fashion contexts. Glasses appeared inconsistently.
Fashionable Professionals: Instead of the traditional ‘dusty’ image, AI-generated curators often resembled ‘yuppie professionals’ or figures from fashion advertisements, appearing elegantly dressed with stylish hair and well-trimmed beards. Their attire varied by museum type, with art and social history curators often in informal blazers and pants, while natural history curators were more formally dressed.
Traditional Museum Settings: Most museum backgrounds generated by the AI were grand, 19th-century ‘hallowed halls,’ with the exception of science and technology museums, which featured more modern designs.
Also Read:
- Unpacking Bias in AI: Which Part of Vision-Language Models is More Stereotypical?
- How Algorithms Shape Academic Visibility: A Gender Fairness Study
The Dissonance and Implications
Interestingly, when ChatGPT4o was directly asked to describe a typical curator, its textual response acknowledged diversity in gender and ethnicity, indicating a ‘textual understanding’ that contradicts its visual output. This suggests that biases are introduced at multiple stages of the AI image generation process, from the initial prompt formulation by ChatGPT to the image rendering by DALL-E.
The findings highlight a significant risk: if these AI-generated images are uncritically accepted, they could perpetuate inaccurate portrayals of museum professionals. Given the copyright-free nature of AI-generated visuals, their widespread use in media, websites, and presentations could reinforce these biases, shaping a false public perception of who curators are and what they look like.
