Automatic summaries are reliably accurate and frequently useless, because the detail the reader wanted has been removed. The omission is systematic and predictable from how the summary is produced.
Importance is inferred from prominence
A summariser has no knowledge of your purpose and infers significance from the document itself.
Material that is repeated, positioned early, or stated at length reads as important, and material stated once in a subordinate clause does not.
Contracts, policies and technical documents place their most consequential content in exactly that briefly stated form, which is where summarisation fails hardest.
Exceptions are shorter than rules
A rule occupies a paragraph and its exception occupies half a sentence, so compression discards the exception and keeps the rule.
The output is then not merely incomplete but wrong, since an unqualified rule states something the document does not.
This is the most damaging summarisation error because it produces a confident statement rather than an obvious gap.
Numbers and conditions detach from their qualifiers
Figures survive compression well, and the conditions attached to them frequently do not.
A summary reporting a limit without the circumstances in which it applies has preserved the memorable half and dropped the operative half.
The same happens with dates, thresholds and eligibility criteria, all of which read as facts once separated from their conditions.
Readers rarely question a figure that appears in a summary, because a number carries an air of having been checked, and the qualifier that would have prompted the question is the part that was cut.
Stating the purpose changes the output
Summarising against a stated interest — what this says about termination, what obligations fall on the buyer — produces a different and far more useful result.
This works because the request supplies the criterion for importance that the document alone cannot provide.
Asking additionally for anything that qualifies or limits the answer recovers much of what a general summary would have removed.
Compression ratio is the setting nobody adjusts
Most tools summarise to a fixed proportion regardless of how dense the source is.
A discursive report survives heavy compression, while a document where every clause is operative loses something at any ratio.
The practical rule is that summarising is safe where the source is repetitive and unsafe where it is dense, and that judgement has to be made by whoever knows what kind of document it is, which is never the summariser.