Gotchas, Limits & Anti-Patterns
Gotchas, Limits & Anti-Patterns
Cowork Will Confidently Make Things Up. Here's How to Catch It.
Every user gets burned by hallucination once. Most get burned again.
You ask Cowork for the exact wording of a clause in a policy you half-remember. It answers immediately: a clean paragraph, a section number, a confident tone. You paste it into an email and send.
Two days later someone replies that there is no such section. The clause doesn't exist. Cowork didn't hedge, didn't stumble, didn't say "I'm not sure." It produced a plausible answer to a question it had no source for, and you couldn't tell the difference from a real one.
This is hallucination (or, more precisely, confabulation) and it's not a sign the tool is broken. It's a known, documented behavior of how these systems generate text. The first time it burns you, it feels like betrayal. The reason most people get burned again is they treat it as a fluke instead of a category.
The fix isn't to distrust everything. It's to learn the small set of moves that make a confident answer auditable.