Opus 4.6 regression symptom #11: confident hallucination of dates/facts in legal documents — liability risk

Status Fixed / completed

Maintainer reply None cached

Activity 0 comments · opened Sep 23, 2026 · closed Sep 23, 2026

Follow-up to #96308 — Additional regression symptom

Symptom #11: Confident hallucination of dates/facts in legal documents

Since ~Sep 16, 2026, the model fabricates specific dates and facts with full confidence in formal/legal documents (complaints, reports, formal letters). Before this date, the model consistently verified data before including it.

Real-world risk

A user who trusts the model output and submits a legal document containing hallucinated dates could face charges of document falsification (e.g., Article 390 of the Spanish Penal Code — 6 months to 3 years imprisonment). The user is legally responsible for any document they sign and submit, even if the false data was AI-generated.

Before vs After regression

| Aspect | Before ~Sep 16 | After ~Sep 16 |
|--------|---------------|---------------|
| Date accuracy | Searched and verified before including | Fabricates with zero hedging |
| Uncertainty | Flagged when unsure | Presents invented data as fact |
| Sources | Cited references | No citations, full confidence |

Impact

This is not just a quality issue — it is a legal liability risk for every user who relies on the model for document drafting. Combined with symptom #2 (hallucinations with confidence), the model has gone from a reliable legal drafting tool to a risk factor.

Environment

Model: claude-opus-4-6 (Opus 4.6)
Platform: claude.ai chat + Claude Code
Regression window: ~September 16, 2026 onwards

View original on GitHub ↗