Scientists are finding AI may make us remember things that never happened

“I think of all this as very dangerous, basically,” Maes said.

So far, much of the public anxiety has centred on what some researchers call “cognitive surrender”: the possibility that, as machines take over more of the work of thinking, people will become less inclined or less able to scrutinise what those machines tell them. But another possibility has received far less attention: AI may be changing what we remember.

New research that involves perception of political events, crimes and even new movies suggests AI-generated images and conversations can implant false details into our brains, even leaving us confident things happened when they did not.

Memory does more than preserve the past. It supplies the knowledge and experience with which we understand the present and anticipate the future. It helps construct a sense of self, sustain relationships, and guide both individual and collective decisions.

Yet memory is also one of the human mind’s great limitations. We retain only a fraction of what we encounter. Details disappear. Recollections blur. Memories can change as we revisit them.

AI seems almost perfectly suited to compensate for these shortcomings. It can store enormous amounts of information, search it in seconds, and retrieve details a person might have lost forever. But these qualities also make it potentially risky. These systems do not merely store the past. They interpret it, summarise it and reconstruct it. And sometimes they get it wrong.

The consequences extend far beyond whether someone misremembers a photograph or a conversation. Courts routinely depend on recollections in cases ranging from car accidents to sexual assault, even though decades of research has demonstrated how easily witness memories can be distorted. History itself depends on what individuals and societies preserve, forget and eventually come to believe about the past.

Maes is one of the pioneers of artificial intelligence. Early in a career that now stretches more than 35 years, she helped develop some of the algorithms that shaped the early internet. But she has long been more interested in a different possibility for AI: not machines that replace human intelligence, but machines that expand it.

For Maes, the question was personal. Her mother began forgetting pieces of everyday life as she grew older.

“My mother would forget who visited yesterday or what somebody told her,” Maes said. She would make notes to herself: “‘Oh, my daughter told me she is coming down two weeks from that day.’”

Maes began imagining something far more powerful than a notebook: a memory assistant that could quietly keep track of a person’s life.

That question takes on particular urgency for people with Alzheimer’s disease and other forms of dementia, diseases defined by memory loss that can rob a person’s ability to navigate everyday life independently.

Maes’ research group has now built several versions of a kind of external memory system. Memoro and MemPal use AI with devices that can be worn or carried to capture pieces of everyday experience and make them retrievable later. Forget what someone told you during a conversation? Ask. Can’t remember where you left the keys? The system may be able to find it.

Instead of asking an ageing brain to become better at remembering, give it access to another memory reservoir altogether.

Early tests with older adults have been promising. Maes’ mother, now 96, has experienced too much cognitive decline to benefit from the technology herself, Maes said. But the scientist has huge optimism that systems like these could eventually help many others preserve independence as their biological memories become less reliable. Commercial products are already beginning to offer some services along these lines, such as interactive recall tools that can be programmed with people’s life histories and personalised smart glasses that help guide people through things like brushing their teeth.

But if machines become extensions of our memories, what happens to those we make for ourselves?

Some of the clearest evidence of what may be happening comes from another corner of Maes’ research.

In one experiment, which was released last year and is undergoing peer review, 54 people were asked to write SAT-style essays on subjects such as what happiness means to them. One group had no online tools. Another could use Google search, stripped of AI-generated answers. A third could use ChatGPT.

As they wrote, electrodes placed across their scalps recorded electrical activity in their brains. The differences were striking. The people writing unaided showed the strongest and most broadly distributed patterns of connectivity between different areas of the brain. Google users fell somewhere in the middle. The ChatGPT users showed the weakest connectivity.

When participants were later asked to quote from essays they had just written, those who had relied on ChatGPT had considerably more trouble recalling their own words.

Remembering is not simply a matter of depositing information somewhere in the brain. Attention, effort, retrieval and the work of making connections all help memories take hold. Generative AI can remove some of that work.

For Gideon Nave and Steven Shaw, researchers who study how AI affects human cognition, that raises a larger concern. Nave, an associate professor of marketing at the University of Pennsylvania’s Wharton School, worries less about the Hollywood nightmare of artificial superintelligence taking over control from humanity than about a quieter possibility: humans simply stop exercising abilities that machines can perform for them.

“What we are more likely to see is humans starting to decay instead of AI taking over,” he said, adding he fears a world in which “thinking becomes optional.”

There are already hints of what researchers sometimes call “de-skilling.” Shaw, an assistant professor at King’s College London, points to studies suggesting that people who become accustomed to performing a task such as doctors diagnosing patients with cancer with AI assistance can struggle when that assistance disappears.

A 2025 study published in the Lancet Gastroenterology & Hepatology details the potential effect. Researchers looked at 19 experienced doctors and their assessments of nearly 1500 colonoscopies and found that after the doctors began routinely using AI, their ability to detect precancerous growths on their own fell by about 20%.

“They are not using that mental muscle in the same way,” he said. “If you don’t use it, you lose it.”

For children, he worries about an even more fundamental problem: They may outsource cognitive work before some of those abilities have fully developed. “They may never develop those key or core critical-thinking capabilities in the first place,” Shaw said.

None of this means that using ChatGPT and other AI agents inevitably makes people less capable. Recent studies do not demonstrate that ordinary AI use inevitably causes cognitive deterioration or permanent memory loss. The research is in its early stages, and humans have been outsourcing thinking and memory to tools – abacuses, watches, dictionaries, calculators – for a very long time.

But generative AI introduces a new complication. We increasingly rely on it to interpret the information it retrieves for us – and people appear remarkably willing to accept what it says. Participants in experiments Nave and Shaw conducted followed AI advice at high rates, even when receiving wrong information.

This week, researchers who study AI ethics and society gathered for an annual conference in Malmö, Sweden, where many are focusing on a particular question: What happens when artificial intelligence starts meddling with human memory?

A growing body of research suggests that generative AI can interfere with memory in surprisingly durable ways. That includes distorting what we remember, blurring the boundaries between our own thoughts and those generated by machines, and even planting memories of events that never happened.

In one study, researchers asked participants to watch a video of a crime and found that those questioned by a generative AI chatbot had three times as many false memories. In another, people were prompted to generate ideas with and without AI, and a week later they had trouble remembering which ideas originated with the machine and which ones were their own.

Other experiments suggest AI-generated misinformation can lead people to remember things that never happened at all. In one experiment, nearly half of participants falsely remembered movie remakes that never existed, including a version of The Matrix starring Will Smith. In another, AI-generated videos convinced people they remembered political scandals that never happened. However, the researchers noted that deepfakes were not consistently more effective than text in creating false memories.

Some of the newest evidence suggests that the problem may be harder to solve than researchers expected.

Companies and government agencies are racing to adopt AI. Police are experimenting with AI summaries of body-cam footage. Hospitals use it to summarise medical information. One of the most common safeguards, said Mattea Sim, an assistant research professor at Georgetown University’s Massive Data Institute, is to keep a human in the loop to catch the machine’s mistakes.

But that assumes the human remains an independent check on the machine. If the influence runs the other way, as Sim’s research suggests, AI may alter the judgment and memories people rely on to catch its mistakes, which undermines the whole premise of the human backstop.

In a recent experiment, Sim recruited 331 people to watch a short animated video of a car striking a pedestrian. A day or two later, the participants read an AI-generated summary of the accident. For some, the summary accurately described a traffic sign in the video. For others, the AI confidently identified the wrong sign.

Among people who received the accurate summary, nearly 84% correctly remembered the sign. Among those who received the misleading version, just about 45% did. Telling participants that the summary had been generated by AI did not protect them from being influenced by it.

“It’s pretty common knowledge that AI can make mistakes,” Sim said. But she said research shows “human oversight may not be a great solution to this problem.”