Answer
Can you trust AI to summarise something you have not read?
For orientation, largely. For anything you will act on, no, because the omissions are unannounced and you cannot see them.
For deciding whether something is worth reading, yes. As a basis for acting, no. A summary's omissions are invisible to someone who has not read the source, and the omitted part is frequently the qualification that mattered.
Summarisation is one of the more reliable things these systems do, and the reliability is about what it includes rather than what it leaves out. The content that appears is usually accurate. What is missing does not announce itself, and a reader who has not seen the original has no way to detect an absence. That asymmetry is the whole of the answer: summaries are trustworthy in the direction you can check and not in the direction you cannot.
The omissions are not random. Qualifications, conditions, exceptions, dissenting passages and specific figures are the classes most likely to be compressed away, because a summary favours the general shape of a document over its edges. Those are also the parts that most often matter for a decision — the condition attached to the obligation, the exception that covers your case, the figure the argument rests on.
That gives a workable rule. Use the summary to orient: what is this about, does it concern me, which section matters. Then read the section that matters. This is not a failure of the tool; it is the correct division of labour, and it saves most of the time that reading the whole document would have cost while keeping you in contact with the part you will act on.
There is a specific case where the answer is firmer. Where the document creates an obligation — a contract, a policy, a regulatory notice, a set of terms — a summary is not a substitute for reading it, because the operative content of such documents lives precisely in the qualifications that summarisation removes. This is also the case where people most want a summary, which is worth noticing about oneself.
One technique makes summaries considerably more useful: ask what it left out. Requesting the qualifications, the exceptions and the parts that complicate the main point produces a second pass over the same material with the opposite bias, and it frequently surfaces the thing that changes your reading. It is not a guarantee, and it is a substantial improvement over the summary alone.
Finally, be careful about a summary of a summary. Material that has been condensed twice has lost detail chosen by nobody, twice over, and the result reads as authoritative as the first. This happens more often than people notice — a briefing built from summaries, a report compiled from digests — and each layer removes the specifics that the layer above no longer knows are missing.
A summary tells you what is in a document; it cannot tell you what it left out, and you are not in a position to notice.
Siddharth Sharma, Context Theory
Related questions
Are summaries of meetings and calls different?
Same asymmetry with a different consequence: what gets dropped from a conversation summary is usually the tentative, the conditional and the thing somebody almost said. For a record of what was decided that is often acceptable; for understanding where someone actually stands, the summary systematically removes the signal.
How do you check a summary without reading the source?
You cannot, which is the point. What you can do is check the specifics it contains against the source, which is quick, and ask separately for what it omitted, which is imperfect. If neither is possible, treat the summary as a description of a document rather than as knowledge of one.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| AI-cited sources that also rank in the Google organic top 10 | 10% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
2026 generative engine citation study · fewer than · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Summary reliability is asymmetric: included content is generally accurate while omissions are unannounced, and a reader who has not seen the source cannot detect an absence, which makes summaries trustworthy only in the checkable direction. | Comparing a summary against its source and listing what did not survive. |
| Response | The classes most likely to be compressed away — qualifications, conditions, exceptions, dissent and specific figures — are the same classes that most often determine a decision. | Checking a set of summaries against their sources specifically for conditions and exceptions. |
| Constraint | For documents that create obligations, the operative content resides in the qualifications summarisation removes, which makes a summary an inadequate substitute precisely where the desire for one is strongest. | Comparing a summary of a contract or policy against the document's conditional provisions. |
| Software | Twice-condensed material has lost detail chosen by nobody at each stage while reading as authoritatively as the first pass, and each layer removes specifics the next layer does not know are absent. | Summarising a summary and comparing the result against the original source. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one