Mnemosyne tests 23 claimed LLM writing tells against 800 pre-ChatGPT human replies
Start here
Mnemosyne Research Labs tested 23 commonly cited "AI writing tells" against human text written before ChatGPT existed. It gave seven 2026 models the same 800 Hacker News comments and Stack Exchange questions that people answered between 2019 and 2022, then counted how often each tell appeared under a pre-registered plan covering 600 human replies and 4,199 model replies. The setup is meant to check whether the tells actually separate machine text from human text, since the human samples predate the models being tested. The reported comparison rests on that pre-registered count; the results themselves are not stated in the available summary.
Generated from reports · updated 41 min ago
Report timeline
Follow the reports to see the event from different sides.
- Hacker News · AIMnemosyne tests 23 claimed LLM writing tells against 800 pre-ChatGPT human replies
Mnemosyne Research Labs gave seven 2026 models the same 800 Hacker News comments and Stack Exchange questions that people answered between 2019 and 2022, then counted 23 claimed AI-writing tells under a pre-registered plan covering 600 human and 4,199 model replies.
Heat for this event
Not enough continuous observations yet to draw a trend.