A summarizer is trained to produce text that a reader will accept. That is a different objective from producing text that is complete, and the gap between the two is invisible by construction: whatever was dropped does not appear in the output, and nothing in the interface marks its absence. The reader's only signal is fluency, which the model supplies in full regardless of how much it discarded.

The failure has a documented name in human-factors research — automation bias — and two distinct shapes. Commission errors happen when the tool says something wrong and the human follows it; those get caught eventually, because wrong claims collide with reality. Omission errors happen when the tool stays silent about something that mattered and the human never asks; those are almost never caught, because there is no collision, only an absence. Summarization produces almost entirely the second kind. A design squad reading a one-page digest of forty user interviews receives three clean themes and no indication that four interviewees raised a fourth concern that did not recur often enough to rank. The sprint ships against three themes. The fourth arrives later as a support queue.

What makes this durable rather than a one-time mistake is the metric. Time saved is easy to count and gets reported; the omission is uncountable and never gets reported, so the practice keeps returning positive evidence for itself. Each cycle the team trusts the digest slightly more and spot-checks slightly less, which is exactly the behavior that would follow if the tool were improving. Nothing about the loop distinguishes a summarizer getting better from a reader getting less willing to check.