AI document summaries: when they prepare your reading and when they replace it

A summary can prepare your reading or replace it, and which one it is doing depends on what the decision costs you.

The pitch for AI summaries is seductive because it is half true: paste the document, take the summary, skip the reading. What the pitch hides is that a summary does two different jobs, and only one of them is honest work at all times. A summary can prepare your reading, or it can substitute for it. The line between the two is not the length of the document and not the cleverness of the tool. The line is what the decision costs you if you get it wrong.

Take the long report you must anyway understand: the market study, the consultant's findings, the department's account of the quarter. Here the summary is scaffolding, and good scaffolding at that. It tells you where the load-bearing parts live before you climb, so you arrive at the real text already oriented, knowing which section holds the argument and which holds the appendices. The summary did not replace the reading. It arranged it, the way a map arranges a city you still have to walk.

Now take the contract, the regulation, the terms you are about to accept, anything binding. Here the summary is a different object entirely: a stranger's paraphrase of obligations you will be held to in full. The stranger is not careless, and is usually not wrong on the broad strokes. But nobody comes after you with the summary. When a dispute arrives, the parties open the document itself, and the document says what it says, including every part the summary found too small to keep.

And that is the mechanism of the failure: what summaries drop. Compression favors the assertive and the central; caveats, conditions, exceptions, and anything living in a footnote are the first to go. Those dropped rooms are exactly where binding documents keep their teeth. The obligation you remember was in the summary; the one that applies to you was in the exception the summary smoothed away. Fluency makes this worse, because confident prose reads as checked prose, and nothing about fluency implies that anyone checked.

The practice follows directly. Use the summary as the map, and read the territory wherever the map will be answered for: money clauses, liability, duration, termination, anything you would not want said about you in a room you are not in. Ask the tool where it is uncertain; a useful tool can name the passages it compressed hardest, and that question is worth more than another polished page. And treat a summary of anything binding as a reason to read, never as the having read.

This is where we sit with the tools we build: summaries for orientation, full reading for commitment, and the tool expected to say what it does not know. For the document that will bind you, the routing is not to another tool at all but to counsel who can read the original and answer for the reading. Guidance, not guarantees: a summary can get you to the right pages faster; it cannot sign, and it will not be the one held to the words.