2 ms·
These watermarking/"detect if someone is AI" techniques are trash, and OpenAI even admits it on their post about this. No proposed or implemented technique outs
by Der_Einzige 4y ago
These watermarking/"detect if someone is AI" techniques are trash, and OpenAI even admits it on their post about this. No proposed or implemented technique outside of this is even close to robust, reliable, or accurate.
It is trivial to get around even the best of these techniques. This is a cat and mouse game with a fat, lazy cat and really motivated, olympic tier mice.
- corobo 4y agoI expect this from Reddit but surely Hacker News isn't judging future tech based on current prototypes? Haha None of this existed at all 5 years ago. Let it cook.
- sebzim4500 4y agoYeah, within a few years there will probably be tools that can reliably detect if a (long) passage was generated by chatGPT or not. I doubt they will be able to tell if the text was generated by whatever will replace chatGPT by that time.
- corobo 4y agoWell then we wait another half year for the 'whatever will replace chatgpt' detector. It's always been a to and fro between spam and ham. Always will be. Unless you're releasing updates to your paper each time a new AI model comes out you're at risk of being caught out if anyone decides to do a "We found 80% of xyz to be AI generated" blog post.
- WolfOliver 4y agojust add "make some common spelling mistakes" at the end of your prompt :D