Opus 4.6 regression symptom #13: model teaches incorrect information to students — educational risk
Follow-up to #96308 — Additional regression symptom
Symptom 13: Model teaches incorrect information with full confidence — educational risk
Since ~Sep 16, 2026, users who rely on Claude for studying or learning receive confidently stated but wrong information. This is a direct consequence of the hallucination regression (symptoms 11 and 12).
Example: Spanish Penal Code
A user studying criminal law would learn from the current model that:
- Art. 390 applies to private individuals (WRONG — it only applies to public officials)
- The penalty is 6 months to 3 years (WRONG — that is Art. 392; Art. 390 is 3-6 years)
- Negligent document falsification is a crime for everyone (WRONG — only for public officials under Art. 391)
All three are stated with full confidence, no hedging, no "I am not sure". A student who memorizes this fails the exam.
Before vs After regression
| Aspect | Before ~Sep 16 | After ~Sep 16 |
|--------|---------------|---------------|
| Legal concepts | Correctly distinguished between legal subjects | Mixes up who each article applies to |
| Confidence calibration | Confident when correct, hedged when unsure | Confident when WRONG |
| Educational value | Reliable study aid | Actively harmful — teaches wrong facts |
Impact
- Students learning law, medicine, engineering or any technical field cannot trust the model
- The model does not flag uncertainty — it presents wrong answers identically to correct ones
- Before the regression, the same model (Opus 4.6) was a reliable study tool
Environment
- Models affected: claude-opus-4-6, claude-opus-5-5, claude-sonnet-4-6
- Platform: claude.ai chat + Claude Code
- Date: 2026-09-23
- Regression window: ~September 16, 2026 onwards