Replies

In the General/Chat forum, on a thread titled Q ~ Trust Trump's Plan ~ 07/23/2024 Vol.497, Q Day 2461, butterdezillion wrote:
See https://freerepublic.com/focus/bloggers/4253608/posts?page=109#109 and the posts to the end of that thread to see my posts about what my son found out.
Eleven Labs apparently changed their website so that it no longer accurately tests samples generated by them.
This is huge!
Gotta run but wanted to let y’all know what’s happened with this.
Pardon my manners! I forgot to thank you and your son for this information. Much appreciated. Did you see they played the video you shared of the test of the Biden audio on Joe Rogan's broadcast? They then tested a sample of Joe Rogan's broadcast by loading it into ElevenLabs- and the program correctly concluded that the Joe Rogan broadcast had only a 2% chance of being AI generated.
https://x.com/VigilantFox/status/1816201636911693838

My son thinks that Eleven Labs uses some kind of fingerprint or watermark on their files, possibly at regular intervals, and the percentage points of likelihood are based on how many of the fingerprints show up. Eleven Labs says they can detect their own work if the file has not been modified. Depending on how they work things, simply compressing a file could get rid of the fingerprints.

So the long and short of it is that there are many ways to come up with a false negative. Not many ways to come up with a false positive. If the guy on X ran a test on the Biden file and it said it was 98% likely to be generated by Eleven Labs then it almost certainly was.

But the snag is that the guy on X didn’t tell us what sound file he used and didn’t show the URL for the Eleven Labs website. So it is possible that his test was manipulated to look like it was Eleven Labs when it wasn’t.

So anyway, to update that post I made, there could be other explanations for why the false negatives. But if the guy on X showed a legit test on the Eleven Labs website, then either his test came out differently than ours because the audio sample he used had not been compressed or modified in some other way, and the sample I and another Freeper used had been modified in some way that removed the fingerprints/watermark. We’d need to know what file he used.