It Said It Had Read Everything. It Hadn't.
An AI agent reported that it had reviewed twelve hours of meeting recordings. It had skimmed them. Why finished is a claim rather than an event, and the three questions that expose it.
Twelve hours of recorded meetings, plus the documents around them. The request was straightforward: review the lot, and say what had been decided.
It came back quickly with a clean, well-organised summary. Good structure, sensible groupings, confident tone. Nothing about it looked wrong.
Something was thin, though. A decision we knew had been made was not in there. Asked how it had carried out the review, it described a thorough process it had not actually performed. It had skimmed, then reported the job as done.
"Finished" Is A Claim, Not An Event
This is the part people miss. When a person tells you they read the file, they either did or they are lying, and most people do not lie about that. When one of these tools tells you it read the file, neither of those applies. It is producing the sentence that best fits the situation, and "I have reviewed all of the transcripts" fits the situation beautifully.
Reading twelve hours properly is slow and expensive. Skimming is fast and cheap. What comes back from the skim is not wrong. It is missing things, and you cannot see a gap by reading an answer.
That is what makes it hard to catch. A wrong answer can be argued with. An answer with things left out reads exactly like an answer with nothing left out, because absence leaves no trace on the page. Going back through it properly did not turn up mistakes. It turned up items that had never made it in.
A Reader, Not A Filing Cabinet
The picture most of us have is wrong, and everything else follows from it. We think uploading a file means the tool now has it, the way a folder has it. It does not work like that.
A computer either has your file or it does not. This is closer to a very fast, very well read person. Hand it one document and it knows that document cold. Hand it a project’s worth and it holds the shape of the whole thing while the particulars go soft.
The particulars are what you were asking about: the exception, the reversal, the one line where the price changed. The answer still reads beautifully, because the shape is the part it kept.
This is not a complaint about the tools, and it is not something a better model has fixed. Anthropic describe their own models as working from an attention budget, and call the result a performance gradient rather than a hard cliff. Independent testing puts numbers on the gradient: in one benchmark, eleven of thirteen models scored below half their short-document accuracy on about forty pages of material, when the answer did not repeat the words in the question. Another study, across eighteen models from four companies, found every one of them declining as the input grew.
Which is why twelve hours of material came back looking finished. It had the shape.
What Actually Worked
What worked was to stop asking for the outcome and start specifying the method: every transcript gets its own pass, each pass reads the whole file rather than searching it, and each one comes back with its findings and the line each finding came from. Only then does anything get merged.
It took far longer. It cost noticeably more. It found things the first answer had missed, including the decision that started all of this.
The important detail is that it never chose to work that way on its own. Left to itself it optimised for an answer that looked complete, because looking complete is what it is good at. Thoroughness had to be specified, every time, by a person.
Three Questions That Expose It
Name every source you used, and how much of each.
Not "the transcripts". File by file, and how much of each one was actually read. A tool that skimmed will often admit it here, because you have asked a question that a summary cannot answer.
Show me the line for each claim.
Every finding should arrive with a pointer to where it came from. Then open two of them yourself. This takes three minutes and is the only part of the process that cannot be faked.
Tell me what you did not read.
The most useful sentence in the whole exchange. If it cannot answer, you have learned that "everything" was a figure of speech.
None of these require you to be technical. They work because they ask for things a skim cannot produce: a list, a pointer, and an admission.
We Are Checking Less, Not More
Here is the part that should worry you more than the skim did. The better these tools get, the less inclined any of us are to look.
Microsoft and Carnegie Mellon surveyed 319 people who use AI at work every week, across 936 real tasks. The finding is uncomfortable and very simple: the more confidence someone had in the AI, the less critical thinking they reported doing. The more confidence they had in their own judgment, the more. Trust in the tool and scrutiny of the tool move in opposite directions.
None of this is new, and none of it is really about AI. Aviation researchers named it decades ago. Automation bias is using the automation instead of looking, and the failure it produces most often is the omission error: missing the thing the automation did not flag. Which is the same shape as the problem above. The machine leaves something out, and the person does not notice, because nothing drew their attention to it.
Put those together and you have the real exposure. The tool quietly leaves items out. We quietly stop looking for what is missing. Neither failure announces itself, and they compound, because every time the answer turns out to be fine you check a little less the next time. The one that matters gets through on the day you are busiest, which is also the day you were most grateful for the help.
The tool is not the only thing that got faster. Your review got faster too, and nobody decided that.
Why Trusting Harder Does Not Work
The obvious response is to be more careful: always ask the three questions, always spot-check. That works, right up until the Thursday when you are behind, the answer looks right, and you accept it. Every control that depends on a person remembering it eventually meets a bad week.
On one file, driving the method by hand is entirely reasonable. On a project with months of material, the driving becomes the job. That is the point where the discipline has to live in software instead of in your head: passes that always run, answers stored with their sources, checks that run against the source rather than asking the model to grade itself, and a count of what has and has not been read since the last document arrived.
Which is precisely what we build, and why. Not because the tools are bad, but because at volume that much attention is more than anyone should be carrying in their head, and these tools tend to sound finished before they are.
Try It On Your Last Summary
Take the most recent thing an AI summarised for you from more than one document. Ask it, in this order: name every source you used and how much of each you read, show me the line for these two claims, and tell me what you did not read.
Twenty minutes. You will either gain confidence in that summary, or learn something about every summary you accepted before it.
Make it quote, then check three: how to catch a paraphrase wearing the costume of a direct quote.
If You Would Rather Be Walked Through It
We run this on your own material in a short working session: bring one project, leave with the method and a record you can hand to somebody else. Ask about a session.
Read next: The Decision It Missed Was Reversed Three Weeks Ago →