Every AI-generated meeting summary in our organisation carries a line stating that it was generated by AI and may contain errors. It’s the least sophisticated control we have, and it’s done more for trust than anything else we’ve built.
That still strikes me as slightly odd, so let me explain what I think is going on.
When we first enabled AI summaries, the anxiety wasn’t really about accuracy. It was about authority. A written summary has a way of becoming the official record. People worried that an AI’s version of a discussion, subtly wrong about who committed to what, would harden into fact before anyone checked it. That fear made people hesitant to use the capability at all.
The disclaimer changed the status of the document. It says: this is a draft. That one line moved the summary from record to working draft, and suddenly people could relax. They used the summaries, they corrected them, they shared them freely, because the artefact wasn’t pretending to be something it wasn’t.
Here’s the lesson I keep drawing from this. In the early stages of AI adoption, trust doesn’t come from performance. It comes from honesty about limits. A capability that claims perfection gets tested adversarially, people hunt for the failure that proves the claim wrong, and one failure destroys it. A capability that declares its fallibility invites people to engage with appropriate care, and every error found becomes confirmation the system is honest rather than evidence it’s broken.
Most organisations instinctively do the opposite. They oversell reliability to drive adoption, then haemorrhage trust at the first public failure.
Would your AI communications survive your most sceptical employee’s first bad experience? If not, you’re not building trust.