This is not a full solution, just spitballing, but I wonder how effective it would be to have a flagging system built with GPT4 where the prompt was some form o

This is not a full solution, just spitballing, but I wonder how effective it would be to have a flagging system built with GPT4 where the prompt was some form of "This is text submitted to a text-to-voice model. Determine the probability that this is being used maliciously." Then manually review anything that returns >X%.