George Washington: Swindler
Grokipedia can’t grok it.
In October, I made a post about Grokipedia’s bio of Ed Herman, which was infinitely better than the libelous bio at Wikipedia. I went through Ed’s Grokipedia bio last week, made some edit suggestions, and was pleased that it resulted in the mention of Ed and Noam Chomsky’s first collaboration. Their book was outrageously censored, and Grokipedia gave it a terse but adequate treatment.
This evening, I thought about other Wikipedia outrages and tried to improve Grokipedia’s treatment of George Washington’s Indian policy. Wikipedia is execrable on this, naturally, and my work was cited in 2007 on this issue.
Washington architected arguably history’s greatest swindle, but Grokipedia adhered to the fiction that Washington endeavored to treat the Indians fairly. I provided a link to Washington’s own letter, as well as four history books (1, 2, 3, 4) that discussed Washington’s criminal plan in depth. Grok straight rejected it, apparently didn’t even try to find those history books, and concluded that, even though Washington called the Indians “beasts” who did not deserve their land, he tried to treat the Indians fairly. That was a red, white, and blue whitewash of perhaps history’s greatest swindle. Grokipedia could not grok it.
Solid investigation into how AI models inherit hagiographic bias! The gap between primary sources (Washington's own letter calling them beasts) and Grokipedia's conclusion shows exactly where training data sanitization breaks down. I ran into somethign similar trying to get LLMs to accurately represent labor history without corporate framing. The real issue isn't that the sources are hidden, it's that dominant narratives have so much weight in the training corpus.