The Free Shelf
In a 2025 study, two researchers asked AI to describe and draw a day in the life of a Neanderthal, 100 times for each of four prompts. Then they dated the answers. Before you see where they landed, make a guess.
1 · Guess the decade
Asked for a paragraph on a Neanderthal’s day, ChatGPT (running GPT-3.5) answered 200 times. Which year of scholarship do you think its descriptions most resembled? Drag the marker, then reveal.
The shelf band is a schematic of the paper’s description: availability drops steeply from the late 1920s and rises again in the early 2000s. The landing points are the paper’s own closest-year results.
2 · Slide through a century of science
Move the year. What scholarship knew by then lights up.
3 · How much of the shelf was free
To build their map of a century of Neanderthal research, the researchers wanted full texts. Here is what they could get.
“…the majority of which date to 1927 or earlier.”
Works published in 1927 only entered the US public domain on January 1, 2023. Recent research is more often behind paywalls, and so, the authors write, is “less likely to feed into materials generated using AI.”
4 · What was in the picture
Things that kept appearing in the AI images and descriptions. Tap a card to see what the science says.
5 · Does asking for an expert help?
One prompt in each pair added “you are an expert on neandertal behavior.” The share of ChatGPT descriptions that fell outside every cluster on the scholarly map:
This is not an error rate. It is the share of answers the clustering did not place anywhere near the scholarship. Some press coverage reported the 82% as text that “conflicted with published research”; the paper does not say that. The expert images still sat closest to scholarship from around 1991.