Summarising looks like the safest AI task. It is the one that fails most quietly.
Here is the failure. You paste in a 40-page contract, ask for a summary, and get eight clean bullet points. Everything in them is accurate. You read it, feel informed, and move on. Three weeks later you discover clause 14.3 — the auto-renewal with the 90-day notice window. It was not in the summary. Nothing was wrong. Something was *absent*, and absence leaves no trace to notice.
Summaries are lossy, and you choose the loss
Any summary throws away most of the text. The only question is what gets thrown away. If you do not say, the model decides using a generic sense of importance — which is roughly "what would a general reader find notable." That is almost never your criterion.
So say your criterion. This is the whole skill.
Summarise this tenancy agreement for someone deciding whether to sign. I care about: money I must pay, dates I must act by, things that let them end it early, and anything that survives after I move out. Quote the exact clause number and wording for each. If a category has nothing, say so explicitly.
That last instruction is doing heavy lifting. "Say so explicitly" turns a silent gap into a visible line: *No early-termination clause found.* Now you know the difference between "there is nothing" and "it did not look."
Ask for the quote, not the paraphrase
For anything consequential, demand the original wording alongside the summary. Two reasons. First, you can check it in ten seconds with Ctrl-F. Second, paraphrase is where meaning drifts. "The tenant may request an extension" and "the tenant is entitled to an extension" summarise to the same bullet and mean different things in a dispute.
Three summary shapes worth knowing
- Decision summary. "I have to decide X. Pull out only what bears on that decision, and tell me what is missing that I would need."
- Change summary. "Here are version 1 and version 2. List only what changed, and flag anything that shifts obligation or cost." This is one of the genuinely great uses — humans are terrible at diffing prose.
- Adversarial summary. "You are advising the other side. What in this document works in their favour?" It surfaces things a neutral summary buries.
The check that costs 90 seconds
After reading the summary, before you act on it, ask one question: what would have to be in this document for my decision to change? Then search the document for those words yourself. Penalty. Terminate. Renew. Indemnity. Exclusive. Whatever your version is.
You are not re-reading the document. You are checking the four or five places where being wrong would be expensive. That is what the summary bought you: not the absence of reading, but the ability to read only where it counts.
And for anything long, know that quality drops in the middle. Models attend well to the start and end of a long text and less well to the centre. If a 90-page report has its worst news on page 47, that is exactly where a summary is weakest. Summarise long documents in sections rather than in one gulp.
Before you move on