One of the most immediately useful things an AI assistant does is summarise: turn a long report, a dense contract, a thread of emails or a wall of meeting notes into a few readable sentences. It saves real time, and the summaries are usually good. That very reliability, though, breeds a habit worth resisting - treating the summary as a faithful substitute for the source, acting on it directly, and never opening the original at all. Most of the time that works. The trouble is concentrated in the times it does not, and those are exactly the times that matter.
A summary is a compression, and compression loses things. It can flatten an important nuance, drop a condition or exception that changes the meaning, emphasise what the model judged salient over what actually matters to you, or - occasionally - state something the source does not quite support. When the stakes are low, none of this is worth worrying about; a slightly imperfect summary of an article you were curious about costs nothing. But when the summary concerns a contract clause you are about to agree to, a figure you are about to report, an instruction you are about to follow, or a commitment you are about to make, the gap between the summary and the source stops being harmless and becomes the thing that trips you up.
The discriminating habit is not to distrust summaries in general - that would throw away most of their value - but to know which parts you are about to rely on, and to check those specific parts against the original before acting. If a summary of a contract says the notice period is one length, and you are about to plan around that, open the contract and read the clause. If a summary reports a number you are about to put in front of a client, confirm the number in the source. The check is quick because it is targeted: you are not re-reading the whole document, only verifying the specific claim your decision rests on.
This is really a particular case of a broader discipline for using AI well - lean on it for speed, but verify before you commit, in proportion to what a mistake would cost. A summary you are reading to orient yourself needs no checking. A summary you are about to act on in a way that would be costly to get wrong needs the one relevant fact confirmed at the source. Building that reflex - pause at the point of reliance, confirm the load-bearing detail against the original - is what lets you enjoy the genuine time savings of AI summarisation without occasionally being led confidently to the wrong place by a summary that was ninety-five per cent right and wrong in exactly the part you needed.