Experiments, findings, and notes from building unslop.
I ran the politicalcompass.org test on all major LLMs. They all landed in the lib-left quadrant. Grok as well, except when it didn't. There were two distinct personas that Grok assumed. This article aims to explain exactly how these results come about and what is happening with Grok specifically.
I scored the full text of 13,500 arXiv papers from 2021 to 2026 across ten different fields and found that 38% of papers sampled in August–October 2026 read as machine-written. There is also a clear upward trend. In this post, I try to give insight into what these results mean and what the limitations of my approach are.
What it was trained on, how it is calibrated, why it was built, and what makes it better than other detectors for academic writing.