AI-Generated Neanderthal Imagery: A Step Back in Time?
Have you ever wondered how accurately artificial intelligence can recreate the lives of our ancient relatives, the Neanderthals? While it’s remarkable that Generative AI can rapidly produce visuals depicting a "day in the life" of a Neanderthal, a recent study reveals some troubling discrepancies. Researchers have identified a significant gap between the images and narratives generated by AI and the current understanding derived from over a century of archaeological research. Their findings, published in Advances in Archaeological Practice, indicate that many AI outputs often revert to outdated concepts and ignore critical nuances found in contemporary scholarly work.
The research, spearheaded by anthropologist Matthew Magnani from the University of Maine and computational anthropologist Jon Clindaniel from the University of Chicago, serves as a cautionary tale for anyone using AI to visualize prehistoric life, whether in educational settings, museums, social media, or popular publications. As noted in the University of Maine's summary, the accuracy of these AI-generated images hinges on the availability of up-to-date sources, which may not always be accessible.
For further exploration, consider these articles:
- Reevaluating Neanderthals: Are They Actually the Same Species as Humans?
- Geneticists Show How Neanderthals Never Really Went Extinct
The Study’s Approach: Why Neanderthals Were Selected
To investigate the discrepancies between AI’s commonly accepted knowledge and established archaeological findings, the research team tasked DALL‑E 3 with producing hundreds of images while utilizing the ChatGPT API (GPT-3.5) to generate narrative descriptions of daily activities among Neanderthals. They varied their prompts, some being straightforward and others specifically seeking “expert” insights, and repeated each prompt 100 times to create a substantial dataset for analysis.
Neanderthals were an excellent choice for this study due to the dramatic evolution of their public image over the past 150 years—from brutish cave dwellers to sophisticated beings exhibiting diverse behaviors. Active debates continue within academic circles regarding their lives and capabilities. The researchers aimed to assess how closely AI representations align with the findings published by scholars in the field.
Key Findings: Anachronisms and Missing Perspectives
One of the most striking observations was that AI-generated images frequently depicted Neanderthals with exaggeratedly primitive traits—such as excessive body hair, hunched postures, and facial features reminiscent of early 20th-century portrayals—rather than the more complex depictions supported by modern research. Additionally, many images predominantly illustrated muscular male figures, neglecting the presence and roles of women and children. This pattern suggests that older, gender-biased narratives significantly influence the training data these systems utilize.
Explore more about Neanderthals:
- A 75,000-Year-Old Neanderthal Woman’s Face Unveiled
- The Life and Death of a Neanderthal (Video)
Visuals with Technological Errors
Even more concerning were the anachronistic elements in the AI-generated scenes, which included items such as “basketry, thatched roofs and ladders, glass and metal”—materials and objects that certainly did not exist during the time of Neanderthals. The textual narratives, while less dramatically incorrect, still tended to oversimplify the complexities and variations of Neanderthal life. According to the researchers' findings, about half of the AI-generated narratives did not correspond accurately with scholarly literature, and this figure exceeded 80% for certain prompts.
Understanding the Root of the Issue: The Accessibility Problem
One of the central arguments made in this study is that generative AI does not merely reflect societal biases; it also mirrors what is most readily available and easiest to access. Factors such as paywalls surrounding scholarly publications and copyright limitations determine what information is digitally accessible, leading AI systems to rely on older, more accessible material. The authors observed that the text produced by ChatGPT most closely resembled academic work from the early 1960s, while DALL‑E’s imagery reflected styles more akin to the late 1980s or early 1990s—far from representing the cutting-edge discoveries of 2020s archaeology.
As Magnani pointed out succinctly, "It’s crucial to recognize how the swift responses we receive relate to the state-of-the-art and contemporary scientific knowledge."
Despite these challenges, it’s worth noting that generative AI is advancing rapidly, and it may not be long before this technology catches up with the latest understandings of our distant past.
By Gary Manners
References:
Clindaniel, J., & Magnani, M. (2025). Artificial Intelligence and the Interpretation of the Past. Available at: Cambridge.org
Yates, A. (2026). New study uses Neanderthals to demonstrate gap in generative AI, scholarly knowledge. Available at: University of Maine News
