The Man Who Has Read the Report

“On the following pages, respond to the questions based on your own recollection of the original video.”

A reasonable instruction. I object to when it arrives.

It comes after the witness has read the account he is supposed to remember independently of. First the event, then the words about the event, then a request to reach past the words and bring back the event. Nothing up my sleeves. Except we have just put something in the witness.

I like to think I would be difficult to fool. Not because I am particularly bright. I have cultivated a disagreeable manner and would hate to discover it had no practical use. I can doubt a sentence before its author has finished typing it. There are men whose entire public careers could have been shortened if somebody had hired me to sit nearby and look unconvinced.

Now there is research suggesting that looking unconvinced may leave the rest of the head available.

Ariana Eunjung Cha reports in the Washington Post on October 15 about AI and false memories. I read the article in a copy archived that morning. One study she describes, by Mattea Sim, Yael Eiger, and Tadayoshi Kohno, gives a particularly unpleasant job to a familiar sentence: keep a human in the loop. The human is meant to catch the machine’s mistakes. The study asks what happens if the mistakes get to the human first.

Their paper, posted on September 23, describes an online experiment. People watched a twenty-five-second animated video of a car hitting a pedestrian. A day or two later they read a summary built from ChatGPT output. The researchers selected and edited the text, deliberately arranging for its traffic sign to agree or disagree with the video. Stop or yield. This wasn’t a machine spontaneously ruining a witness in the wild. It was a controlled test of what misleading words could do.

Then came that instruction. Remember the video. Not the summary.

Of the 328 people included in the analysis, 83.6 percent who read an accurate sign description picked the right sign in a two-choice test. Among those given the misleading description, 44.8 percent did. These are two groups, not a measurement of each person’s memory getting worse. There wasn’t a no-summary group in this experiment.

I would like to volunteer an objection before somebody assigns me to one of the groups. The thing was written by AI. Surely a little contempt would help.

The researchers tried a version of that distinction. Some participants were told the summary came from AI software; others were told it came from a professional human transcriber. In fact, they read the same AI-generated text, apart from the sign manipulation. The supposed source made no statistically significant difference on the traffic-sign question. Nor did the researchers find a significant protective relationship with how much people said they trusted AI or used it. They caution against treating those null results as the last word.

But there goes my first qualification. I had brought an attitude to a memory problem.

This is a small experiment with brief animated accidents, not proof that actual police witnesses always behave this way, or that an AI summary permanently destroys a person’s memory. The paper leaves those real-world questions open. It doesn’t establish that every warning or every way of checking must fail. It does take away the easy assurance that a person sitting there will make the arrangement safe.

Sitting there is the part I could have guaranteed.

What irritates me is how readily I had accepted the promotion. Human judgment. My department. After all the talk about machines writing, reasoning, doing the difficult work, here was something reserved for the older animal: tell whether the thing was right. I wouldn’t have to make the first draft. I would supply the important frown afterward.

But a writer ought to recognize the difficulty. Give me an account of an incident and I begin seeing what the sentences describe. That is part of what words are for. A good paragraph doesn’t politely remain a set of marks while I complete an investigation into its origins. It gets on with the job. I can dislike the writer and still see the room.

The researchers aren’t claiming to have discovered that human beings only became suggestible when ChatGPT arrived. Their experiment extends a line of memory research going back decades. A human can supply the misleading words too. The new offer is to have a system produce accounts at speed, then depend on people to correct them from memories those accounts may already have reached.

I cannot sneer my way outside that sequence. Nor can I return to the earlier version of myself by announcing that I am now being careful. I can look for the recording. I can compare actual details instead of relying on the force with which I remember them. But if the only check available is my recollection after reading, calling it human judgment doesn’t add another witness.

There is work I can do. There are errors I might catch. I am not resigning from thought because somebody has made it less flattering. I am refusing the little extra claim, the one that arrives without a wage attached: that my presence guarantees a second, untouched account.

You asked for the man who saw what happened.

You have the man who has read the report.

Primary research: AI-Enabled Human Memory Manipulation: Misleading AI-Generated Summaries Distort Human Memory (September 23, 2026).

Source: Scientists are finding AI may make us remember things that never happened (October 15, 2026; inspected copy archived October 15, 2026, 11:56:30 UTC).