They killed the tool, because it was doomed to fail to begin with. You can't "detect" AI content. It's content MADE to be like human content, that's what it was trained to be.
Like we can use neural networks to detect content by specific authors (if we have enough content to sample), we can detect content produced by specific versions of specific AI models. But that doesn't tell us if it's AI or not in general. A model is just another "personality" with little biases. And if the sample is article-sized, and not book-sized, then it's also not big enough to detect properly.
4 comments
[ 568 ms ] story [ 2180 ms ] thread> Look at the attached chart. It tells a scary story that will not be limited to StackOverflow.
> Right now, detecting AI-generated content is impossible.
> Last week, OpenAI shut down the tool they created for this purpose. They launched it in January, and it's dead today, less than seven months later.
[continues at article URL]
Like we can use neural networks to detect content by specific authors (if we have enough content to sample), we can detect content produced by specific versions of specific AI models. But that doesn't tell us if it's AI or not in general. A model is just another "personality" with little biases. And if the sample is article-sized, and not book-sized, then it's also not big enough to detect properly.
Discussed at length 2 days ago, 768 comments: <https://news.ycombinator.com/item?id=36855516>