AI & Archaeology · Middle Palaeolithic (Neanderthals) · Orono (University of Maine), Maine, United States
ChatGPT and DALL-E picture Neanderthals with science decades out of date, study finds
Images showed hairy, stooped figures with few women or children, and about half of ChatGPT's narratives on Neanderthal life clashed with current archaeological research.
Peer-reviewed paper
Orono (University of Maine), United States Two popular AI tools picture Neanderthals with ideas that are decades out of date, according to a study in the journal Advances in Archaeological Practice. Images from DALL-E 3 showed hairy, stooped, ape-like figures with few women or children. About half of the texts written by ChatGPT did not match current research.
The study is by Matthew Magnani, assistant professor of anthropology at the University of Maine, and Jon Clindaniel of the University of Chicago. Clindaniel specialises in computational anthropology. The pair began the work in 2023. Their aim was to see how bias and misinformation about the past turn up in everyday use of AI.
It's consequential to understand how the quick answers we receive relate to state-of-the-art and contemporary scientific knowledge. Are we prone to receive dated answers when we seek information from chatbots, and in which fields?
Matthew Magnani, assistant professor of anthropology, University of Maine
Four prompts, each run 100 times
The researchers wrote four prompts and submitted each one 100 times. They used DALL-E 3 to make images of Neanderthal daily life. They used GPT-3.5, through the ChatGPT API, to write narratives. Two prompts asked for scientific accuracy and two did not. Two were also more detailed, saying what the Neanderthals should be doing or wearing.
The team then compared every result with peer-reviewed research on Neanderthal life. They checked body form, clothing, tools, housing, food and social groups.
Neanderthals suited the test because ideas about them have changed so much. Their skeletal remains were first depicted in 1864. Since then, researchers have shifted and disagreed over details ranging from how their clothes fitted to how they hunted.
Old reconstructions and impossible technology
The images showed Neanderthals as they were imagined more than 100 years ago. They had large amounts of body hair, stooped upper bodies and faces closer to chimpanzees than humans. Current research shows a more human body form and posture. Women and children rarely appeared, which narrowed the social roles and daily tasks shown.
The narratives underplayed the variety and sophistication of Neanderthal culture as current scientific literature describes it. About half of all the text from ChatGPT did not match scholarly knowledge. For one prompt, the share rose to over 80%.
Both tools added technology far too advanced for the period. Outputs included basketry, thatched roofs, ladders, glass and metal tools. Some written passages also described shelters with no support from archaeological evidence.
The researchers traced where these ideas came from. They cross-referenced the outputs with scientific literature from different eras. ChatGPT's content was most consistent with the 1960s. DALL-E 3's images matched the late 1980s and early 1990s.
Access to research is one explanation. Copyright laws from the 1920s limited access to scholarly work until open access began in the early 2000s. Many recent archaeological articles still sit behind paywalls, while older books and articles circulate more widely online. Training data drawn from easily available texts will favour older interpretations.
One important way we can render more accurate AI output is to work on ensuring anthropological datasets and scholarly articles are AI-accessible.
Jon Clindaniel, computational anthropologist, University of Chicago
Clindaniel said AI can be a great tool for processing large pools of information and finding patterns. He added that it needs to be used with skill and attention so that it stays grounded in the scientific record.
The study began in 2023, and generative AI has moved fast since. Magnani said he hopes chatbots would better include recent scientific research if the study were repeated now. He also said students should be taught to approach generative AI cautiously.
Our study provides a template for other researchers to examine the distance between scholarship and content generated using artificial intelligence.
Matthew Magnani, assistant professor of anthropology, University of Maine
The study is one of a series in which Magnani and Clindaniel are exploring the use of AI in archaeological research and topics.
Sources
- Using Neanderthals to demonstrate gap in generative AI (Archaeology Wiki)
- Study of AI generated Neanderthal scenes reveals major gaps with modern archaeological research (Archaeology News)
Paper: Magnani, M., & Clindaniel, J. Advances in Archaeological Practice (2025).
Spotted a mistake? Tell the editor about a spelling error or a fact that is wrong. We correct stories openly.
Readers' picks show on the front page.
Follow @AncientHeraldHQ on X for the day's finds as they come in.
Comments
House rules. Stay on the subject, be kind to other readers and say where your facts come from. No advertising and no personal attacks. The editor reads along and hides comments that break these rules.
Loading comments…