evidence EV-AI-COMM-2026-D120

Automatic factuality metrics do not reliably measure all factuality failures

Identifier: EV-AI-COMM-2026-D120
Research Area: ai-mediated-communication
Status: draft
Version: 0.1
Confidence: 0.9
Priority: medium
Tags: factuality, evaluation, ai
Updated: 2026-07-22
Source: research/evidence/EV-AI-COMM-2026-D120--factuality-metrics-limited.md

Evidence Record

Stress tests show automatic factuality metrics capture incomplete and sometimes misleading proxies. An AI-generated handoff cannot be accepted merely because a second model scores it highly.