The bot gave a wrong answer, but the log looked clean. The right article was retrieved. The confidence score was high. The content was correct. The answer was coherent.
It addressed the wrong problem entirely.
This is context contamination. The system injected incorrect metadata into the prompt based on a wrong channel assumption. The model answered correctly for the wrong context.
The diagnostic signature
Imagine a support system built around a mobile app. Device signals such as operating system, app version, connectivity, and SIM presence are injected when a customer contacts support through the app. When the customer arrives by email, the system may interpret missing signals as "no SIM detected" or fall back to the mobile adapter.
The retrieved article is correct for a mobile user. The answer is polished. But the customer is on email and has a different problem. No hallucination. No keyword gap. No content error. The context was wrong.
Treat this as its own root cause category: CONTEXT_CONTAMINATION.
Useful defences include provenance scoring for every injected signal, a post-generation channel consistency check, extracting context from what the customer actually wrote, and contradiction detection. System signals should be treated as hypotheses to validate, not facts to consume.
The most dangerous failures are the ones that look like successes until someone reads the actual message.
This is part of the Parenting Your AI series, a practitioner's guide to building AI skills that are safe, effective, and worth trusting. Written from inside enterprise AI systems by someone who has spent years diagnosing what goes wrong when AI meets real work at scale.
Read the full series at KnowledgeManagement.ie