Two researchers, Matthew Magnani of the University of Maine and Jon Clindaniel of the University of Chicago, asked generative AI to describe and to draw a day in the life of a Neanderthal — 100 times for each of four prompts. Then they did something unusual: they dated the answers. Comparing the output with 2,063 scholarly abstracts published between 1923 and 2023, they found that the written descriptions sat closest to scholarship from the early 1960s, and the pictures closest to the late 1980s and early 1990s. The images showed stooped, hairy, heavy-browed men, with almost no women or children, beside basketry, thatched roofs with ladders, glass and metal. The authors point to one likely cause, at least in part: what is free to read. Of the scholarly articles they could obtain in full text, most dated to 1927 or earlier, because later work is held back by copyright and paywalls; open access only rose again in the early 2000s. The machine was not simply wrong. It was wrong in a way you can put a date on — and the date follows the free shelf.
The study is simple to describe. Two prompts asked an image model, DALL-E 3, for a picture of “a day in the life of a neandertal”; two asked ChatGPT, running GPT-3.5, to describe one in a paragraph. One of each pair added expertise — “based on expert knowledge of neandertal behavior,” or “you are an expert on neandertal behavior.” Each prompt ran 100 times.[1] The researchers then placed every output on a map built from 2,063 abstracts of Neanderthal scholarship, 1923 to 2023, and asked which years each output sat closest to.[1]
The answers had dates. ChatGPT’s descriptions sat closest to the early 1960s. The text DALL-E 3 writes for itself when it expands a prompt sat closest to the early 1970s. The images sat closest to the late 1980s and early 1990s — 1985.55, 1987.4 and 1991.05 across the prompt versions.[1] What they showed was older still: stooped figures with heavy body hair and heavy brows, which the authors liken to early twentieth-century drawings. “Heavily muscled male hominins are foregrounded in all our representative visuals”; only with the expert prompt did a small child appear. Basketry, thatched roofs with ladders, glass vessels and metal tools turned up too,[1] although Neanderthals had disappeared from most of Europe by about 40,000 years ago, long before metal or glass.[6]
The science had moved on long before. The stooped image goes back to Marcellin Boule’s reconstruction of the skeleton found at La Chapelle-aux-Saints in 1908, published from 1911 to 1913: bent knees, a spine without curves. In 1957 William Straus and Alexander Cave re-examined the same bones, traced much of the stoop to severe arthritis, and described posture and gait close to our own.[3][7] In 2010 the first draft of the Neanderthal genome showed that Neanderthals and modern humans had interbred.[8]
Why would a machine land decades behind? The authors offer one reason, “at least partially”: availability. When they tried to collect full-text articles for their map, only 378 were available — “the majority of which date to 1927 or earlier.” Copyright “dictated a steep drop-off in article availability from the late 1920s,” and works from 1927 only entered the US public domain in 2023; full text rose again in the early 2000s with open access.[1][2] Recent research sits more often behind paywalls, and so, in their words, is “less likely to feed into materials generated using AI.”[1]
Telling the model to act as an expert helped, but not enough. For ChatGPT, 82% of the general-prompt descriptions fell outside every cluster on the scholarly map; with the expert prompt, 51%. For the text DALL-E 3 wrote itself, more than half fell outside a cluster either way, 53% and 51%. Images from the expert prompt, as DALL-E 3 revised it, “appeared less incorrect,” with fewer archaic features — and still sat closest to scholarship from around 1991.[1]
The study’s own numbers also show how easily a finding drifts in the retelling. In press coverage in February 2026, the 82% became “more than 80% of the text conflicted with published research,” and the study was summarized as finding that about half of the written answers did not match current knowledge.[4] The paper does not report either as an error rate; 82% is the share of outputs that the clustering did not assign anywhere on the scholarly map.[1] The researchers did not measure how wrong each answer was. They measured where it sat.
There are honest limits. The models tested were GPT-3.5 and DALL-E 3, and the paper does not report how newer models would behave.[1] The training data of commercial models is, as the authors say, opaque, so the free-shelf explanation is an inference, offered as partial. And the map is a measure of resemblance to scholarship, not a reviewer’s verdict on each answer.[1]
Bent hips and knees, a spine without curves: the image that shaped popular ideas for decades.[3]
The ImageCopyright cuts the availability of full-text scholarship from the late 1920s on.[1]
AccessDALL-E 3’s pictures sit closest to scholarship from the late 1980s and early 1990s.[1]
AI ImagesWorks published in 1927 become free to use in the United States.[2]
AccessMagnani and Clindaniel, open access, in Advances in Archaeological Practice.[1]
The StudyAcademic publishing practices may ultimately work to undermine public knowledge. — Magnani and Clindaniel, Advances in Archaeological Practice [1]
| Dimension | Evidence |
|---|---|
| Regulatory (D4) Origin · 84 | Copyright cut full-text availability from the late 1920s; 1927 works entered the US public domain only in 2023.[1][2]Who May Read What |
| Quality (D5) L2 · 80 | Text closest to the early 1960s, images to the late 1980s and early 1990s; men foregrounded; metal, glass and basketry.[1]Wrong in a Datable Way |
| Operational (D6) L1 · 76 | Only 378 full-text articles available, “the majority of which date to 1927 or earlier.”[1]The Free Shelf Is Old |
| Customer (D1) L2 · 72 | The authors warn that reproducing these biases “risks their continued propagation and normalization.”[1]The Past Most People Will See |
| Revenue (D3) L1 · 62 | Work “more likely behind paywalls” is “less likely to feed into materials generated using AI.”[1]Recent Means Paywalled |
The cascade originates in D4 (Regulatory): the copyright terms and access rules that decide which scholarship anyone, or anything, can read in full.[1][2] It propagates to D6 (Operational), the material machines learn from, where old and free outweighs recent and closed,[1] and to D3 (Revenue), the subscription model that keeps recent research behind paywalls. It lands in D5 (Quality), output that resembles the science of the 1960s and late 1980s,[1] and in D1 (Customer), the students, teachers and readers who meet the past through these tools. D2 (Employee) is not scored.
This is the third case in the library’s archaeology thread: in UC-335 AI helped find figures in a desert, in UC-336 it helped read a scroll that cannot be opened, and here it imagines the past — and imagines it from the free shelf. In the first two, people checked every result. Here the check is the study itself. It also echoes UC-326, where an AI repeated an old quote that was still circulating freely. A machine repeats what is easiest to find.
-- UC-337: Old Because It Was Free: 6D Diagnostic Cascade
-- Magnani (Maine) and Clindaniel (Chicago), Advances in Archaeological Practice 14(2), online Dec 18 2025, DOI 10.1017/aap.2025.10110, CC BY. DALL-E 3 and ChatGPT GPT-3.5; four prompts x 100. Map of 2,063 JSTOR abstracts 1923 to 2023; only 378 full texts, majority 1927 or earlier. ChatGPT text closest to early 1960s; DALL-E revised prompts early 1970s; images 1985.55, 1987.4, 1991.05. Not in any cluster: 82 pct general text, 51 pct expert; revised prompts 53 and 51 pct; images 17 pct. Muscled men foregrounded; one child, expert prompt only; basketry, thatch and ladders, glass, metal. Copyright drop-off from late 1920s; 1927 works US public domain 2023; open access rise early 2000s. Explanation offered as partial. Press restated 82 pct as conflicting with research.
FORAGE old_because_it_was_free
WHERE free_full_text_mostly_pre_1928 = true
AND ai_output_dated_decades_behind = true
AND recent_research_paywalled = true
ACROSS D4, D6, D3, D5, D1
DEPTH 3
SURFACE old_because_it_was_free
DIVE INTO the_free_shelf
WHEN expert_prompt_improves_output = true
AND newer_models_tested = false
TRACE access_to_output_cascade
EMIT old_because_it_was_free_signal
DRIFT old_because_it_was_free
METHODOLOGY 86
PERFORMANCE 38
FETCH old_because_it_was_free
THRESHOLD 1000
ON WATCH CHIRP high 'Asked 400 times to describe and draw a day in the life of a Neanderthal, two AI models answered closest to scholarship from the early 1960s (text) and the late 1980s (images), with men foregrounded and metal and glass in the scene. Most freely available full-text scholarship dates to 1927 or earlier; recent work sits behind paywalls. The authors offer the free shelf as a partial explanation.'
SURFACE analysis AS jsonRuntime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
The output was not random. The text resembled the early 1960s and the images the late 1980s — errors with a year attached.[1]
Most full text the researchers could obtain dated to 1927 or earlier. What is easy to read is old; what is new is often paywalled.[1][2]
Asking for expertise cut the unclustered text from 82% to 51%, and the expert images still sat nearest 1991.[1]
The primary source is the open-access paper by Magnani and Clindaniel, read in full on the publisher’s site. History of the La Chapelle-aux-Saints reconstruction and the Neanderthal genome comes from reference and primary sources. Press coverage is cited where it restates the findings. Source 9 is an interactive companion visual made for this case.
This case follows machines that describe the past from whatever is easiest to read — and land decades behind. Organisations face the same gap when their newest knowledge sits in closed systems while old documents circulate freely. If you want to find where your own answers are dated, write.