Less than two weeks later, I was mildly surprised when I performed a Google search and found its AI quoting my words nearly verbatim, and presenting them with the confidence of fact: https://i.imgur.com/pgImAWh.png
I took the opportunity to play with search terms, to see just how generic a query could lead to a response lifted from my post. At the moment, googling "Asus Pro WS WRX90E-SAGE SE diagram has errata" does it [2], and to some extent even just "Does ASUS block diagram have errata" [3].
In this case, the comment is (mostly) correct - in fact, ASUS is working on fixing the diagram.
But what if I had made up the post, or worse, were a competitor trying to sabotage their reputation?
I'm not sure whether I should be impressed or scared. I'm amazed how quickly one obscure comment can be surfaced to look authoritative. The AI further amplified my sentiments in its responses by referring to "users" or "technicians" reporting these particular mistakes, when it's really just one person - me. While I know a human with good technical instincts could quickly gauge my comment and recognize its legitimacy, I'm not as confident the AI isn't just naively parroting.
I'm curious, are such tactics being used for "AI poisoning" attacks in the wild, and is there anything publishers can do to protect themselves?
[1]: https://www.reddit.com/r/threadripper/comments/18obxsd/comment/nses469/
[2]: https://i.imgur.com/Qf3pGnF.png
[3]: https://i.imgur.com/FvHiF5h.png
gus_massa•1mo ago
I quote people here many times, I add a link to the original comment, and I only do when <allcalps>I</allcaps> think the comment is correct. [1]
If it's correct, it's not poisoning.
[1] Or maybe when it's wrong and I explicitly say it's wrong. But that is usually not nice, so I don't do that too often, or perhaps I never did that.