Codex asks researcher to blind treatment status from itself during data analysis

Afinetheorem · x · 2026-08-20

While using Codex to build follow-up outcome data from a randomized treatment, a researcher found Codex suggested blinding the treatment status in the CSV from Codex itself — since next-token generation could in theory be biased by knowing group assignments. A correct move for causal inference, and an amusing sign of AI self-awareness about its own biases.

Original post →

More from Fun

Fun channel →