Reposted by Mukesh MakwanaSaurabh @saurabhr.bsky.social · 30/07/2026Most LLM eval tools ask "did it get the right answer." Cognitive psychology experiments ask "under which manipulated condition, with which controls, measured how." We open-sourced PsychScanner: an experimental harness for LLMs, not a benchmark, a lab. 🧵 172
Reposted by Mukesh MakwanaSaurabh @saurabhr.bsky.social · 30/07/2026Can an AI tell its own words from yours? This is "reality monitoring." In humans, failing at it is linked to hallucinations and delusions. In this work we investigated reality monitoring in 6 LLMs of different sizes in collaboration with @lazyneuron.bsky.social and @brianodegaard.bsky.social. 1147