It is impressive how Generative AI can conjure a “day in the life” image of a Neanderthal in seconds. But a new study suggests those scenes often come with a dash of built-in time travel. Researchers comparing AI-created Neanderthal images and narratives to a century of scholarly writing found the outputs repeatedly echoed outdated ideas, missed key nuances in modern research, and reproduced familiar biases, concludes their study published in Advances in Archaeological Practice.
The work, led by anthropologist Matthew Magnani (University of Maine) and computational anthropologist Jon Clindaniel (University of Chicago), is being framed as a warning for anyone using AI to “illustrate” prehistory - whether in classrooms, museums, social media, or popular articles. As the University of Maine summary puts it, accuracy often depends on whether the systems can access up-to-date sources in the first place.
- Reevaluating Neanderthals: Are They Actually the Same Species as Humans?
- Geneticists Show How Neanderthals Never Really Went Extinct
What the Researchers Tested (and Why Neanderthals)
To probe the gap between AI “common knowledge” and archaeology, the team prompted DALL‑E 3 to generate hundreds of images and used the ChatGPT API (GPT‑3.5) to produce narrative descriptions of daily Neanderthal life. Some prompts were plain; others explicitly asked for “expert” knowledge. They ran each prompt 100 times to build large samples they could analyze systematically.
Neanderthals were an ideal test case precisely because their public image has swung dramatically over the last 150+ years - from brutish cavemen to complex humans with varied behaviors - and because debates about their lives remain active in the literature. The study then compared AI outputs to a large corpus of published Neanderthal scholarship using multimodal computational methods (including CLIP embeddings), asking: how close is the AI’s “past” to what researchers actually write?
The Biggest Gaps: Outdated Bodies, Anachronistic Tech, Missing People
One key finding is AI imagery often defaulted to Neanderthals with exaggeratedly archaic features - heavy body hair, stooped postures, and faces that resemble early 20th‑century portrayals more than the range suggested by modern research. The images also frequently centered heavily muscled males while sidelining women and children, a pattern the authors interpret as a sign that older, gendered narratives still dominate the training “memory” these systems draw from.
Image of 'Neanderthal daily life' produced by prompt in the study. (Advances in Archaeological Practice, 2025)
More jarring were the technological mashups. The University of Maine write-up notes AI scenes inserting things like “basketry, thatched roofs and ladders, glass and metal” into Neanderthal contexts - objects and materials that don’t belong in that timeframe.
Text outputs came off as less flamboyantly wrong, but still tended to flatten variation and sophistication. In the Maine summary, the researchers report that roughly half of the AI narration failed to align with scholarly knowledge, rising above 80% for one prompt.
Why It Happens: The “Access Problem” Behind AI History
A key argument in the paper is that currently generative AI doesn’t just reflect bias in society, it reflects what’s easiest to ingest. Scholarly publishing paywalls and copyright constraints shape what’s digitally accessible, and that can skew the “average” knowledge AI systems absorb toward older, more available material. In their analysis, the authors found ChatGPT’s Neanderthal text most closely resembled scholarship from around the early 1960s, while DALL‑E’s imagery aligned more with the late 1980s/early 1990s - hardly the cutting edge of 2020s archaeology.
Magnani put it bluntly in the university release: “It’s consequential to understand how the quick answers we receive relate to state-of-the-art and contemporary scientific knowledge.”
With that said, Geberative AI is improving at a rapid rate, it surely won't be long before this modern tool catches up with the past.
By Gary Manners
References
Clindaniel, J., & Magnani, M. 2025. Artificial Intelligence and the Interpretation of the Past. Available at: https://www.cambridge.org/core/journals/advances-in-archaeological-practice/article/artificial-intelligence-and-the-interpretation-of-the-past/8FE3F2CB6BBFAD49F75FFC3031158A5A
Yates, A. 2026. New study uses Neanderthals to demonstrate gap in generative AI, scholarly knowledge. Available at: https://umaine.edu/news/2026/02/new-study-uses-neanderthals-to-demonstrate-gap-in-generative-ai-scholarly-knowledge/
