Imagine relying on a cutting-edge tool to depict ancient life, only to find it’s stuck in the past—literally. A groundbreaking study reveals that AI’s portrayal of Neanderthals is not just outdated but downright inaccurate, relying on century-old stereotypes rather than modern scientific understanding. But here’s where it gets controversial: could this be a symptom of a larger issue in how AI learns and perpetuates information? Let’s dive in.
Over the past four decades, technology has transformed phones and computers into the world’s most accessible library. With generative AI, answers to questions about ancient humans or even heart rate changes are delivered in seconds. Yet, speed doesn’t always equate to accuracy. This gap is the focal point of a new study led by Matthew Magnani, an assistant professor of anthropology at the University of Maine, and Jon Clindaniel, a professor of computational anthropology at the University of Chicago. Published in Advances in Archaeological Practice, their research poses a deceptively simple question: When AI reconstructs daily life from the distant past, does it reflect modern science or cling to outdated notions?
The researchers chose Neanderthals as their test case—a species that has been misunderstood for over a century. Early depictions painted them as hunched, primitive, and barely human. However, recent studies reveal a far more sophisticated reality: Neanderthals possessed cultural skills, complex social structures, and physical diversity. This evolution in understanding made them the perfect subject to test AI’s ability to keep up with scientific progress.
But here’s the kicker: Even when explicitly asked to be accurate, AI often defaults to older, more accessible data rather than current research. Magnani emphasizes, ‘It’s crucial to examine the biases embedded in these technologies. We need to understand how the quick answers we get align with contemporary scientific knowledge.’
To test this, Magnani and Clindaniel launched their project in 2023, using two popular AI tools: DALL-E 3 for images and ChatGPT (GPT-3.5 model) for text. For images, they crafted four prompts—two without accuracy requests and two based on expert knowledge. Each prompt was run 100 times, generating 400 images. Some runs allowed DALL-E 3 to refine the prompt, while others required strict adherence to the original text. For text, they generated 200 one-paragraph descriptions, half from basic prompts and half from prompts asking the AI to respond as a Neanderthal expert.
The results were eye-opening. Many AI-generated images depicted Neanderthals as heavily hunched, covered in thick body hair, and ape-like—traits rooted in early 20th-century misconceptions. Women and children were rarely included, with most scenes focusing on muscular adult males. The text fared no better: about half of the descriptions contradicted modern scholarly understanding, with one prompt yielding over 80% inaccurate responses. Both images and text often mixed timelines, featuring advanced tools like glass, metal, and thatched roofs—technologies far beyond Neanderthal capabilities.
By comparing AI outputs with decades of archaeological literature, the researchers found that ChatGPT’s text aligned most closely with scholarship from the early 1960s, while DALL-E 3’s images matched the late 1980s to early 1990s. This lag highlights a critical issue: AI often learns from older, more accessible data, even when accuracy is requested.
And this is the part most people miss: Much scientific research remains locked behind paywalls due to early 20th-century copyright rules. Open access publishing only gained momentum in the early 2000s, making older material easier for AI to access. Clindaniel notes, ‘Making anthropological datasets and scholarly articles AI-accessible is key to improving accuracy.’ The researchers faced this challenge firsthand, relying on abstracts instead of full-text papers from the 1920s onward due to limited availability.
Why does this matter beyond archaeology? Generative AI is reshaping how we create and trust images, writing, and sound. While it empowers individuals to explore history and science, it also risks spreading outdated stereotypes on a massive scale. In fields like archaeology and anthropology, where public understanding relies heavily on visuals and narratives, inaccuracies can cement misconceptions. Neanderthals are just one example—the same risks apply to countless cultures and historical periods.
Magnani sees this as both a warning and an opportunity: ‘Our study provides a framework for researchers to assess the gap between scholarship and AI-generated content. Teaching students to approach AI critically will foster a more informed society.’
The practical implications are clear: AI tools should be used cautiously, especially in education and science communication. While teachers, students, and journalists can benefit from AI’s speed, they must question its sources. The study also underscores the need for open access research, as making modern studies more accessible could help AI reflect current knowledge rather than repeating the past.
Finally, this research offers a method for testing AI accuracy across disciplines. As AI becomes ubiquitous, such tools are essential to ensure technology enhances learning rather than distorting it. But here’s the question: As AI continues to evolve, who is responsible for ensuring it stays updated with the latest science? Is it developers, researchers, or the public? Share your thoughts in the comments—let’s spark a conversation!
