alt Hacker NewsNo, it really isn't. If you think a 0.1% of writing being "poisoned" is going to trick AI, you're wildly underestimating AI's capabilities.
Replies
mid-kid • today at 7:02 PM
phoronixrly • today at 6:12 PM
Oh? People keep parroting that but the small detail missing is that there needs to be enough volume of correct information to counter your adversarial disinfo. Otherwise the LLM is one agentic call away from stating it as fact.
Research showed AI training could be poisoned by as few as five documents in a training set of millions.