AI detection software is an impossible joke. Offline AI or anyone with knowledge of the API tools for Textgen can beat this. All I have to do is declare “in the style of” with a few parameters and ban a few common tokens that cause the structures the detection software is looking for. That is just what is possible with low effort. I can train a small Low Rank Adaptors layer and no detection software will be able to find the patterns. This is as easy as training on an entire novel’s text. It will have very little effect on the knowledge or accuracy but it will greatly impact the overall style and word choice with a large overall deviation that no software will pick up on. It is possible to train such a LoRA in an hour on consumer grade hardware. Indeed, my favorite has been a LoRA I trained on God Emperor of Dune.
We need an undergraduate student union. And I mean a real one, not a department called “The Student Union” that is just a normal part of the administration which handles student life.
Add comment