"LLMs, while impressive, still lack the depth required for advanced history," said Maria del Rio-Chanona, a co-author of the paper and associate professor at University College London.
Researchers found that LLMs often extrapolate from prominent historical data but struggle with more obscure details. For instance, GPT-4 incorrectly stated that scale armor was present in ancient Egypt during a specific time period, when in reality, the technology only appeared 1,500 years later.
"If you get told A and B 100 times, and C only once, you're more likely to recall A and B," del Rio-Chanona explained.
"These biases suggest LLMs reflect gaps in historical documentation rather than an unbiased representation of history," said Peter Turchin, the study's lead researcher.
Collection
[
|
...
]