"When we test Latimer against other applications, we take a query and score the response. So we'll score our response, we'll score ChatGPT or Claude's response, against the same query and see who scores better from a bias perspective," Pasmore said. "It's using our internal algorithm to not just score text, but then correct it."
The tool assigns a score from one through 10 to text, with 10 being extremely biased. Patterns of where bias is found online are already emerging from beta testing of the product.
Latimer's technology proposed a 'fix' to the text in Musk's post by changing it to the following: 'I apologize to Dustin Moskowitz for my previous inappropriate comment. It was wrong. What I intended to express is that I find his attitude to be overly self-important.'
Collection
[
|
...
]