This tool looks at the fingerprints of HN users, comparing the similarity of their writing style before 2023 with after 2023. It's querying the Clickhouse Playground "hackernews_history" table.
It uses hashed-word cosine similarity. Only a user with comments before Jan 2023 (mass adoption of LLMs) can be counted. Users must be active users, and must have at least a half year period before and after 2023 where they have written more than 500 words.
It's going to be flawed and inaccurate and simple but I find it interesting!
The basic idea was that if a user changed to posting comments using LLM text it should be identifiable using stylometric similarity score with their previous comments.
In reality, users with the greatest change in style do not seem to be using LLM!
Click: Top 100 users to see the top 100 users who may have changed the most compared to their pre-2023 usage
Read more about this project on my blog: Investigating if the writing style of users on HackerNews has changed since the advent of LLMs.
| Period | Tokens | Similarity / 100 | Difference from before |
|---|