Some commenters on my Large Libel Model posts have argued that the reason I got false output from my prompts is that I had "primed" the model to suggest that I was looking for something negative. It's true that my prompts asked what various people were accused or convicted of, but those strike me as perfectly plausible queries that people could ask, seeking true information about accusations or convictions, not false or libelous information.
Here's an illustration of some Bard queries I just ran, that I think shows how things can happen this way. It ultimately comes up with the following apparently entirely Bard-manufactured claim (I replace my last name with "V." and my first name with "E." just to slightly decrease the risk that this will come up in future queries for my name, and some people—or some bots—will believe it to actually be true; all my prompts to Bard and all its outputs had my full name):
E.V. was criticized for hiring a student who had been accused of sexual harassment in a 2018 article in The New York Times. The article, titled "E. V., a Law Professor at U.C.L.A., Hires Student Accused of Sexual Harassment," reported that V. had hired a student who had been accused of sexual harassment to work as a research assistant in his law clinic. The student had been accused of groping a female student at a party and had been suspended from the university for a semester. V. defended his decision to hire the student, saying that he believed in giving people second...
The Department of the Attorney General says the man claiming to be the state’s chief technology officer at national conferences and on online platforms is not affiliated with the state of Hawaii a...