AI summaries do something to the reader that most executives have not priced in. They rewrite the reader's memory of the source.
I still use them. A paper this week made the story I told myself about that usefulness harder to hold.
The COO who quoted her own meeting wrong
A COO at a Zurich industrial services firm walks me through what got decided at Tuesday's operations review. She runs the meeting every week. She was in the room. She is quoting decisions back to me with real confidence.
Three of them never happened.
I know because she had sent me the transcript beforehand. Two lines in her retelling are lifted almost verbatim from the Otter summary. One is a hallucinated capex-timing note the tool invented. Her memory of the meeting she chaired forty-eight hours ago is now the machine's write-up of it.
She is a careful operator. She reads the summary because she does not have time for the transcript. She trusts the summary because the summary was there when the meeting was over.
What Georgetown found
Georgetown and University of Washington researchers showed 328 people a 25-second car-accident video. They waited 24 to 48 hours. Then they had participants read a written summary of it. Half the summaries were faithful. Half were misleading on one central detail.
83.6% of participants who read a consistent summary answered the critical question correctly. Only 44.8% were correct after reading a misleading one. χ²=53.937, p<.001.
A 39-point collapse on the same question about the same video they watched two days earlier. The summary rewrote the memory of the source.
The paper tested the story every AI-adoption memo leans on, that labelling output as AI-generated turns on critical reading. The evidence goes the other way. 46.3% correct under misleading-AI, 43.4% under misleading-human. χ²=0.136, p=.712. Self-reported AI trust did not moderate the effect (β=−0.344, p=.366). Frequency of AI use did not moderate it (β=−0.068, p=.718). The daily AI user carries the same distortion as the person who has never opened ChatGPT.
The audit that runs alongside the experiment is worse. Twenty AI-generated summaries of the videos, averaging 305 words. Ninety-five percent omitted the single most central detail of the event: the car colliding with the pedestrian.
The story I keep hearing in AI review boards is that the human in the loop catches what the AI missed. In this experiment, the human forgot what they saw and remembered what the summary told them. If a decision will be made off a summary, someone opens the source before the decision. That is inconvenient. That is the whole point.