Getting AI to summarise a long document without betraying it

A 40-page report, a contract, a study: you do not have time to read it all, so you ask AI for a summary. The result is fluent, well written, convincing — and that is exactly the problem. A text can be pleasant to read while quietly erasing the nuance that mattered. Here is how to get a reliable summary, and how to spot one that is not.

The real risk is not the glaring error

We picture the danger as a visible hallucination, an invented figure that jumps out. The most frequent risk is more discreet: a nuance that disappears, a caveat in the source text that becomes a categorical claim, a secondary result that takes on the air of a conclusion. Nothing that jumps out unless you reread the original — which is precisely what a summary is supposed to save you from doing.

The four-step method

1. Do not ask for a summary too short for a document too dense. An 80-page report compressed into ten lines forces the AI to decide between what stays and what disappears. First ask for a summary per section or per chapter, then a synthesis of those summaries: Anthropic’s documentation calls this a meta-summary, and it often captures details a direct compression would have lost.

2. Place the document before the question. Counter-intuitive but documented: pasting the long text before your instructions, rather than after, clearly improves the quality of the answer — by up to 30% in some tests, according to Claude’s technical documentation. The question comes last, once the model has “read” all the context.

3. Specify what you want to keep. “Summarise this report” is a poor instruction. “Summarise this report, keeping the key figures, the decisions made and the points of disagreement between the parties” is a good one: you tell the AI what, for you, must not disappear.

4. Check the passages that matter, not the whole document. You do not have to reread everything — that would give up the time saving. But for a contract or a report that shapes a decision, go back to the original on the clauses or figures that truly weigh. That is where a lost nuance costs the most.

A template instruction for a reliable summary

These four steps fit into a reusable instruction. Adapt the brackets to your document:

Here is a [report / contract / study]. Summarise it in [format: max 10 bullets].
Keep, without fail: the key figures, the decisions or obligations,
and the points of disagreement or caveats expressed.
Flag with "[TO CHECK]" any point where the text is ambiguous.
Never turn a hypothesis into a fact or a caveat into a certainty.

The last line is the most important: it explicitly names the risk you want to avoid. A model asked to preserve the caveats erases them far less often.

Three signals that should raise a flag

• The summary is sharper than the source document was (a hypothesis becomes a fact).
• A figure appears that you cannot find in the original text.
• The document contained contradictory views and the summary keeps only one.

In all three cases: ask again, explicitly specifying that nuances and points of disagreement must be kept.

Before / after: the nuance that fades

A typical case of flattening, from a sentence in a study:

• Source: “The results suggest a possible link between the tool and the productivity gain, subject to confirmation on a larger sample.”
• Unfaithful summary: “The tool improves productivity.”
• Faithful summary: “The study suggests a possible link with productivity, still to be confirmed on a larger sample.”

The unfaithful summary is not false “at face value”: it has simply become too sure of itself. That is exactly the kind of drift your instruction must forbid, and your verification must catch.

The risk changes with the document

Not all documents betray themselves the same way. This table helps you know where to focus your attention:

Type of document	What risks disappearing	The precaution
Report / study	Caveats, margins of uncertainty	Require keeping the conditions and hypotheses
Contract / clause	Exceptions, conditions, deadlines	Reread the committing clauses at the source
Meeting minutes	Disagreements, undecided points	Ask it to list the open points
Long email thread	Who said what, the order of decisions	Ask for a timeline, not a synthesis

When not to delegate at all

A document scanned as an image rather than selectable text degrades reading quality before the AI even summarises anything. And for any document with legal or financial stakes — contract, confidentiality clause, amendment — the summary is a starting point for your reading, never a substitute for it. It is the same reflex we detail in the confidentiality habits before pasting your data into an AI: AI prepares, you decide.

Key takeaway

A good AI summary is built, not requested in a single sentence. Split dense documents, place the text before the question, say explicitly what must not disappear, and reserve your verification for the passages that really matter. The time saving stays real — provided you keep control of what you keep.