Ars TechnicaAn Iowa school district is using ChatGPT to decide which books to ban
Official: "It is simply not feasible to read every book" for depictions of sex.
Amy Dentata@arstechnica “The question of whether a book contains a description of depiction of a sex act can only be accurately answered by a model that has seen the full text of the book”
Incorrect! The model can and will produce wrong answers regardless of what is in its training set. It will produce wrong answers even if it has been fed the text directly. LLMs do not produce accurate results, period, as they are just text generators, not answer machines.
This is stupid on top of stupid on top of stupid.
Dan Shick@amydentata @arstechnica yes, but of course they don’t care about facts. they’re just happy to blame their moral panic on computers. it’s not like the general public understands well enough to push back!
Jared White (ResistanceNet ✊)@amydentata @arstechnica The idea that an LLM is "accurate" is ludicrous. That's not how the technology works. 😡
Atomic FoxI don't understand why people aren't getting this. ChatGPT and its relatives are statistical models for generating plausible strings, to all intents and purposes super-duper Markov chains. They don't know anything, they don't answer questions.
Aargh. But there are people out there betting billions of dollars, and in some cases human lives, that people will choose to pull the wool over their own eyes. And experience suggests that they will, at least for a while.
Pippin@amydentata @arstechnica And by the way they work, picking the most likely next word to continue the text, they are designed to produce plausible-LOOKING output, regardless of accuracy or anything else. LLMs are literally automated con-artists.