During a lecture, we were given a tip for reducing AI hallucinations: add a paragraph telling the model to stick to facts, admit uncertainty, and stop inventing sources.
I was sceptical. It sounded like advice from 2025. If one paragraph made such a difference, surely model providers would already use it? That was only an assumption, so I asked my agents to test it. Then I tested it again under conditions closer to everyday work.
The result was a distinction I had overlooked. The instruction made the agents more cautious. It did not reliably make them more correct.
