No, you cannot tell AI by use of emdashes and this common claim just makes people look even more illiterate than they already did. Real people use emdashes and we’re not going to stop.
Real people also like to use short sentences to make them more impactful.
You write with too much confidence, I think. I work on these models professionally so I don’t know what to tell you besides your surface level understanding of LLMs seems wrong. 🤷♀️
Models do use emdashes disproportionately often, along with shorter sentences than the average human, because they are trained with RLHF where humans rate the model outputs based on their preferences. This is different than copying human behavior as you suggested they do. With RLHF (and later more efficient methods), humans tended to rate models that used emdashes and that had shorter, matter of fact sentences higher because it projected more confidence. In fact, models that were confidently wrong were more likely to be rated higher than models that were unconfidently correct.
Respectfully, if you do the stereotypical armchair expert thing and double down, I will block you. I don’t have the patience to deal with people incorrectly mansplaining my own field to me. I just want to spread knowledge.
No, you cannot tell AI by use of emdashes and this common claim just makes people look even more illiterate than they already did. Real people use emdashes and we’re not going to stop.
Real people also like to use short sentences to make them more impactful.
The LLMs “learned” these from humans, after all.
You write with too much confidence, I think. I work on these models professionally so I don’t know what to tell you besides your surface level understanding of LLMs seems wrong. 🤷♀️
Models do use emdashes disproportionately often, along with shorter sentences than the average human, because they are trained with RLHF where humans rate the model outputs based on their preferences. This is different than copying human behavior as you suggested they do. With RLHF (and later more efficient methods), humans tended to rate models that used emdashes and that had shorter, matter of fact sentences higher because it projected more confidence. In fact, models that were confidently wrong were more likely to be rated higher than models that were unconfidently correct.
Sources cited:
Respectfully, if you do the stereotypical armchair expert thing and double down, I will block you. I don’t have the patience to deal with people incorrectly mansplaining my own field to me. I just want to spread knowledge.