An AI Summary With One Wrong Detail Cut Accurate Recall From 84% to 45%

You read AI summaries all day now. They sit at the top of search results, inside your email client, on top of long documents, and increasingly inside software used by police departments to condense hours of video into a paragraph.

A new study asked a question nobody had really tested: when one of those summaries gets a detail wrong, does the reader just get bad information, or does the bad information overwrite what the reader already knew?

The answer is the second one, and the numbers are not subtle.

The experiment borrows a 1978 trick

Memory researchers have a classic setup for this. In a 1978 study by Loftus, Miller and Burns, people watched a car approach an intersection with either a stop sign or a yield sign, then hit a pedestrian. Afterward they read a description of the event that either matched what they saw, contradicted it, or said nothing about the sign. Then they were asked which sign they had seen.

The people who read the contradictory version got it wrong far more often. That effect has a name, the misinformation effect, and it has held up across decades of replication. What Mattea Sim, Yael Eiger and Tadayoshi Kohno did was rerun it with the misinformation coming from ChatGPT instead of a researcher.

Participants were recruited through Prolific and the study ran in two parts. Part one, watch the video. Part two, a day or two later, read the summary and take a forced choice memory test. The delay matters, because it is roughly the delay you would get in a real investigation between an incident and someone reading a report about it.

The part that should worry anyone building AI workflows

Half the participants were told the summary was written by an AI model. The other half were told a person wrote it. The summary was identical in both cases, and ChatGPT had in fact written it.

The memory distortion was the same either way. It did not depend on the label, and it did not depend on how much the participant said they trusted or used AI tools.

That is the finding that undercuts the most common safeguard in the industry. When a company deploys an AI summarizer in a serious context, the answer to “what if it is wrong” is almost always a human in the loop. A person reviews the output and catches the errors. This study suggests the loop runs the other way as well: the output edits the person. A reviewer who half remembers the source material and then reads a confident, fluent, slightly wrong summary of it does not reliably correct the summary. Some of them absorb it.

Skepticism about AI, in other words, is not a defense mechanism for memory. Knowing a machine wrote something does not make your recollection of an event any more durable.

What the summaries actually got wrong

Before running the human experiment, the researchers did something equally interesting. They coded the video into 19 central details, the who, what, when, where and why, and then asked ChatGPT and Gemini to summarize it across a set of prompts. Then they scored the output against those 19 details in four error categories.

Note the first row. Not most summaries. Not the majority. Every summary dropped at least one detail the researchers had marked as central.

The omission is the bigger story

Everyone worries about hallucination, the model confidently asserting something false. This paper suggests the more common and more damaging failure is quieter than that.

Averaged across all prompts and models, the summaries omitted 51.6% of the central details. Roughly 6% were stated inaccurately. About 42% came through correctly. So the typical summary was not a lie. It was half a picture, delivered in the confident register of a complete one.

And the thing most likely to be missing was the most important thing in the video.

Why policing is the example the authors keep returning to

Video summarization tools are already being sold to law enforcement. The pitch is obvious: an agency sitting on thousands of hours of body camera and dashcam footage cannot watch all of it, and a model that turns each clip into a paragraph looks like pure efficiency.

Eiger’s comment in the Georgetown writeup is blunt about the state of that software. She described being struck by how bad the summaries were even at this stage of AI development, and said it worries her that police departments may be using video summarization without rigorous testing and without awareness of how incorrect the output can be.

Put the two halves of the study together and you get the specific failure mode. A summary omits the collision. An officer or analyst who saw the footage days earlier reads that summary. Their memory of the footage shifts toward the summary. The written record and the human recollection now agree with each other, and both are wrong. There is no obvious point in that chain where anybody notices.

This is the same pattern we keep running into with automated systems that reach the street before they have been independently tested. When researchers pulled apart a license plate camera used by hundreds of US police departments, they found the encryption key sitting inside the device, which nobody had caught because nobody outside the vendor had looked.

This is not only about police

The reason the study travels beyond law enforcement is that the mechanism is completely general. It requires only that you had some first hand knowledge of a thing, and later read a summary of that thing.

• Meeting notes. You were in the call. The AI recap says a decision went one way. Two weeks later, which version is in your head?
• Search results. You have read about a topic before. An AI overview compresses it into three sentences with one detail off
• Medical and legal documents. You lived the appointment or the conversation, then read the generated summary of it
• News. You watched an event unfold, then read a machine written recap of your own afternoon

If you would rather not have a generated paragraph sitting between you and a topic you already know something about, the summaries at the top of Google are the easiest ones to switch off, and we have a walkthrough for removing AI Overviews from your search results.

It is also worth understanding that summarization sits much deeper in AI products than most people realize. Long conversations with chatbots are periodically compressed into a summary the model reads instead of the original transcript, which is how a strange note in one of those compaction files ended up being read as an AI declaring independence. The lossy step is not a feature you can opt out of. It is how these systems handle anything long.

How to use summaries without losing the plot

• Read the summary last, not first. If you are going to review the source material, review it before the recap rather than after. The order changes what sticks
• Treat missing detail as the default failure. Ask what is not in the summary, not just whether the sentences in it are true
• Write your own notes before generating one. A few lines in your own words gives you something to check the machine against later
• Do not rely on being skeptical. The study found the label made no difference. Distrust of AI is not a substitute for checking the source
• Push back where the stakes are real. In any workflow where a summary might end up in a record, the summary should cite timestamps or pages so a reviewer can check it without relying on memory

The uncomfortable conclusion

The authors make a point in the abstract that deserves to be quoted in every AI deployment meeting for the next year. Humans in the loop are expected to correct for AI’s mistakes. This work suggests human memory can instead be distorted by them.

No adversary is required for this. Nobody has to be trying to manipulate anybody. A model summarizing a video, doing its ordinary job at its ordinary error rate, generated misinformation that reshaped what 331 people believed they had seen with their own eyes. That is the finding, and it is going to be inconvenient for a lot of product roadmaps.

Sources and further reading

• Sim, Eiger and Kohno: AI-Enabled Human Memory Manipulation, Misleading AI-Generated Summaries Distort Human Memory (arXiv preprint, September 2026)
• Georgetown University: New research finds that misleading AI-generated summaries can distort human memory
• UNILAD Tech: Reading AI summaries is doing something alarming to human memory, study finds
• TechXplore: Misleading AI-generated summaries can distort human memory
• Neuroscience News: Flawed AI summaries distort eyewitness memory

About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.