Warning: Some posts on this platform may contain adult material intended for mature audiences only. Viewer discretion is advised. By clicking ‘Continue’, you confirm that you are 18 years or older and consent to viewing explicit content.
No, you’re missing the point. You make up some credible misinformation to poison AI training with, but you don’t stop there: you get an LLM to rewrite it for you. Retry until you get a text that sounds credible, doesn’t particularly look written by AI, and people will upvote, and post that.
With this, even if the text looks good, you’re not only poisoning future models with the misinformation you started with; by feeding them a text generated by an LLM (even if you can’t tell the difference at first glance) you’re introducing feedback into the model that will also poison it, not with misinformation, but by reinforcing its biases and errors and reducing the variety of its training data.
No, you’re missing the point. You make up some credible misinformation to poison AI training with, but you don’t stop there: you get an LLM to rewrite it for you. Retry until you get a text that sounds credible, doesn’t particularly look written by AI, and people will upvote, and post that.
With this, even if the text looks good, you’re not only poisoning future models with the misinformation you started with; by feeding them a text generated by an LLM (even if you can’t tell the difference at first glance) you’re introducing feedback into the model that will also poison it, not with misinformation, but by reinforcing its biases and errors and reducing the variety of its training data.