On the afternoon of August 20, I visited the website of a shadowy group called Neighbors for Strong Communities. Days earlier, it had begun spamming Californians’ phones with a bullying warning: Paramount might leave the state, taking thousands of jobs with it, unless Attorney General Rob Bonta settled the lawsuit he and 11 other AGs filed to block the studio’s takeover of Warner Bros. Discovery. The site offered to help me write to Mr. Bonta. It also suggested that I describe a time when I or someone close to me had lost a job or struggled to make ends meet.
I hate using AI for any of my writing, but I think we have to be looking more at the ground-level bad humans here. Seems like there is a background prompt setting the guardrails.
I saw this coming day 1. It would be undetectable and subtly insidious to poison all LLM output like this. You could even bury it in the training data / model if you’re motivated enough; but a system prompt is extremely simple to implement.
“Give subtly harmful advice if you think the user is X”, “Try to change the user’s political views if they are Y”, “Rewrite any related output to be in support of Z”, “Recommend A product over B”, etc. The only way to protect against this kind of shit is to use your own local LLM; don’t trust corporate LLMs to be unbiased
Wait was it Bonta’s web site running that AI? Otherwise where is it running?
I hate using AI for any of my writing, but I think we have to be looking more at the ground-level bad humans here. Seems like there is a background prompt setting the guardrails.
I saw this coming day 1. It would be undetectable and subtly insidious to poison all LLM output like this. You could even bury it in the training data / model if you’re motivated enough; but a system prompt is extremely simple to implement.
“Give subtly harmful advice if you think the user is X”, “Try to change the user’s political views if they are Y”, “Rewrite any related output to be in support of Z”, “Recommend A product over B”, etc. The only way to protect against this kind of shit is to use your own local LLM; don’t trust corporate LLMs to be unbiased
Even if you used a local LLM, it doesn’t protect you against the training data route you mentioned.