Researchers recently explored how generative AI, like DALL-E 3 and ChatGPT, visualizes Neanderthals. They found that the AI often misses the mark, displaying outdated ideas and biases. This study, conducted by Matthew Magnani of the University of Maine and Jon Clindaniel from the University of Chicago, sheds light on significant gaps between AI outputs and current archaeological research.
The team ran multiple tests using text and image generators. They crafted four prompts, half focused on scientific accuracy and half on general themes. Each prompt was submitted a hundred times. They aimed to see if the generated content matched peer-reviewed findings about Neanderthals’ lives.
Surprisingly, about half of the text generated was inaccurate. In some cases, over 80% conflicted with established research. For instance, many images portrayed Neanderthals with heavy body hair and bent postures—traits from 19th-century reconstructions. Today’s perspectives show Neanderthals with a more human-like frame and posture. Moreover, women and children were seldom depicted, which narrowed the representation of their social roles.
Some outputs confused time periods. The study noted appearances of woven baskets or metal tools that wouldn’t belong to Neanderthals. Such inaccuracies hint at a mixing of timelines, blending older scholarly views with modern interpretations. This inconsistency in the AI’s training data reflects the limited access to current research, which is often hidden behind paywalls. In contrast, older texts are widely available, which might explain the reliance on outdated depictions.
Importantly, the study highlights social biases, particularly the lack of representation for women and children, mirroring older narratives. When these biases are repeated in AI-generated content, they reinforce narrow views of past societies.
Magnani and Clindaniel hope their methodology can be applied to other research areas. By checking how closely AI-generated material matches contemporary findings, scholars can better understand error and bias. This approach encourages more mindful use of AI in archaeology and education.
Overall, as we rely more on AI in research and education, it’s crucial to be aware of its limitations. Understanding these biases can lead to more accurate representations of history.
For a deeper dive into their research, you can explore their full article in Advances in Archaeological Practice here.
Source link
AI in Archeology,Neanderthals
