It sounds like a promise about the future. It is an instruction about the present.

Think about how you actually formed your opinion of these tools. Most people ran one test, early, on something that mattered — a clause in a contract, a piece of code, a question inside their own field. It fumbled. It invented a citation, or missed the obvious, or handed back something confident and wrong. You closed the tab and filed it: not for serious work. You have been running off that verdict ever since.

The trouble is the verdict expires in weeks. The model that embarrassed itself in front of you has already been retired. The one that replaced it is quietly better at the exact thing you gave up on — and you have no idea, because you stopped looking. Mollick's line is not cheerleading. It is a warning against building a permanent opinion out of a temporary sample.

This bites hardest for the people who tested earliest and most carefully — the ones who decided, with evidence, that AI "can't do our domain." That evidence was real. It is also two versions old. The confidence you earned from a fair test in spring is the same confidence that now keeps you from noticing the tool grew up.

So today, open the one task you concluded these tools couldn't handle. Hand it to the version in front of you, not the one in your memory. You are not testing the AI. You are testing how old your opinion is.