At a conference in Barcelona in 2025 I watched a doctor present their use of AI to review scans. The case on the screen was a skilled radiologist finding one fracture on an image and missing two others.
The point was not that the radiologist was careless. Radiologists have a documented failure mode called satisfaction of search, which the modern literature calls the subsequent search miss: finding one abnormality on an image measurably reduces the chance of finding a second. It accounts for up to a third of certain classes of reading error. The detail that matters is the breakdown. Roughly ninety per cent of these misses are recognition failures rather than scanning failures. The eyes were on the lesion; they fixated it, and did not see it. The mechanism is still argued over and the lesson does not depend on it.
Reading carefully is useful behaviour. It is not a checking strategy, and the people it fails are looking directly at the thing they miss.
Now take that into your own week. You ask for every date in the contract and get eight; there were nine. You get a citation to a policy that was replaced two years ago, in the correct format, with a plausible clause number. You check an agreement against a clause checklist, get four issues back, and the missing signature page is not among them, because the checklist item was worded differently.
Ask which of those a careful reading would have caught. None of them.
Two properties of the technology sit underneath all three. The same input can give you a different answer: put the same document through the same summarisation twice and there is a fair chance the second version omits a caveat the first included. Both look fine, and nobody compares them, because why would you compare a thing to itself. And fluency is decoupled from correctness. The output reads as though a competent person wrote it because, in a real sense, the writing is competent. The writing was never the thing in question.
So the re-read fails twice over. It is aimed at the page, and the failure is not on the page.
What works instead is choosing the checking act to match the failure. A conversion needs recomputing against the rule. An extraction needs comparing against the complete source, which means going back to the source and counting, rather than scanning the output for something that looks wrong. A citation needs the document opened and read. A classification needs its boundary cases examined. A compliance check needs a known problem seeded into it, because a check that has never failed has not been tested. A confident claim needs tracing to something that can hold it up.
None of that is more reading. All of it is more work than reading, which is why it gets skipped.
It gets harder where several machine acts happen before anybody sees anything. A briefing arrives; the retrieval pulled last quarter's figures because the folder changed, a classification step treated them as current, and the draft asserts a trend. Three failures, and what you are reading is one confident paragraph with a number in it.
There is a reading of the Barcelona case that cuts the other way, and it is the most valuable thing in it. The machine does not get satisfied and stop looking, because it has no first find to be distracted by. On the failure human attention is worst at, the machine has a structural advantage. Pointing it at your own finished work, as a second pair of eyes rather than as the thing that produced the work, is among the most useful and least discussed things these tools do.
The radiology finding is published evidence and it will bear the weight I am putting on it. The move from a chest film to a contract note is mine, and it rests on practice rather than on anything measured, so give it less credit than you give the finding. It is also a convenient thing for me to believe. I run a consultancy whose service is judgment-preserving AI adoption, so I am professionally disposed to find this problem everywhere.
Something to try this week. Take one AI-assisted output you accepted in the last fortnight and check it by the act rather than by reading. If it contained an extraction, open the source and count. If it cited anything, open the citation. Fifteen minutes, one output. Note what you find, or note that you found nothing, which is also information.
The uncomfortable version. Ask a colleague to take a document you would normally process, remove something material from it, and hand it back without telling you what they took out. Then run your ordinary check and see whether the absence comes back. You will find out what your checking actually catches, and so will they.