A fluent answer that is not true — a name, a quote, a date, a source, a policy — delivered in the same calm tone as a true one.
The model is finishing a pattern. If the pattern of “helpful answer” is stronger than the pattern of “I don’t know,” it will often pick helpful. That is not malice. It is how the machine is built.
A mark you put on any claim that must be true before you send, publish, or act. Names, prices, dates, promises, medical or legal statements, “we can do X by Friday.”
A small habit
- Ask it to list facts separately from the draft.
- Tag each fact [CONFIRM] or drop it.
- Do not paste a reply until the tags are gone — because you checked, or because you cut the line.