tested whether AI models can recognize their own writing in a blind lineup. grok went 0 for 9. it wrote something, then a minute later insisted someone else wrote it