Every human mind is a strange museum of moments we believe we lived. We walk around with images of childhood kitchens, school corridors, the face of a loved one at a crowded airport, the sound of laughter at a party. We trust these inner pictures with an almost instinctive certainty. But cognitive scientists have long known that memory is not a video recorder. It is a fragile, reconstructive act, influenced by what we hear, what we fear, and what we want to believe. Now, in the age of generative AI, a new actor has entered that private theater: synthetic media. A study from MIT’s Media Lab, reported by The Washington Post, has shown that AI-edited photographs and videos can do more than deceive people in the present. They can plant false memories about the past, and—unsettlingly—make people more confident in those invented recollections than they are in their own accurate memories. Pattie Maes, an MIT professor who led the work, puts it bluntly: “I think of all this as very dangerous, basically.” The danger is not just that we might be fooled once. It is that our memories—the very stories we tell ourselves about who we are—can be revised by software without our knowledge or consent. Think how often you scroll past images online that have been generated, edited, or animated by AI. A politician in an impossible setting, a celebrity saying something they never said, a familiar street scene altered with new buildings or erased crowds. Each one may seem like just another piece of digital content. But the MIT research suggests otherwise: these images can leave a mark on the mental record of real events. The problem is not limited to obvious propaganda or deepfakes. Even seemingly harmless photo edits—a changed sky, a removed person, an added object—can become part of what someone later remembers. In a world where cameras were once regarded as objective witnesses, the technology that once recorded reality now shapes, distorts, and invents it.
To test the power of AI-generated content over memory, Maes and her colleagues designed a careful, controlled experiment. They selected 24 ordinary, unremarkable photographs. The images depicted everyday scenes: a wedding with guests clustered around a couple, a busy city street with shoppers and traffic, a race with runners mid-stride. Nothing about these pictures was meant to be shocking; they looked like the sort of photos one might find in a family album or a news archive. Two hundred participants viewed each of the baseline photographs one by one, giving them a chance to form an initial memory. Then, after a filler task designed to create a delay, the participants were split into four groups of fifty. The first group, the control, simply viewed the original photographs again. The second group was shown AI-edited versions of the same photographs—images in which objects had been added, removed, or transformed, people added or erased, and background elements changed. The third group watched short five-second AI-generated videos that had been created from the original still photos. The fourth group watched videos that had been generated from the already AI-edited photographs. After this viewing, each participant answered one question about each photograph, testing their memory of the original image. The researchers measured not only whether answers were correct, but how confident participants were in their answers. The study was presented at the CHI 2025 conference, a venue where such findings could be shared with designers and technologists who shape AI tools. By using this four-way structure, the researchers could compare the impact of static edits with that of motion, and also see whether putting edits into motion made them even more powerful. The questions were not designed to be tricky; they asked about simple, visible details in the original photographs. Yet the results showed how easily a mind can be rewritten.
The results were striking, and they cut in a clear direction. Compared with the control group, all three groups exposed to AI-edited or AI-generated content produced more false memories. In other words, even a single exposure to an altered image was enough to nudge some participants to remember the photograph differently than it had been. The effect was strongest among participants who had watched AI-generated videos built from AI-edited stills. These participants produced 2.05 times as many false memories as the control group. Their confidence in those false answers was also inflated, coming in at 1.19 times the confidence level of the control group. What makes this second number especially troubling is the combination. People are not merely confused; they are certain. They do not experience a vague sense that something may have been different; they feel a clear, vivid conviction that the AI-influenced version is what they originally saw. The false memory has become, as far as their brain is concerned, a real memory. This effect has a kind of smoothness to it. A still photo can be studied. A video unfolds with the compelling logic of real time, carrying its own context and rhythm, and it is harder for the mind to question. When movement is added to an already altered scene, the brain seems to lower its defences still further. The result is a memory that feels not like an artificial construction but like an authentic experience. These numbers come from a controlled experiment, not from the chaos of social media, but that is exactly the point. In the lab, researchers had to alter photographs in obvious ways to test memory. In the wild, AI-generated content is often designed to be seamless and persuasive. The measured effect in a calm, attentive setting may be even larger in real-world conditions, where people scroll quickly, consume content without pause, and rarely compare an image to a reliable original. The MIT results are thus a kind of baseline warning: if false memories can be planted so easily in a study, what happens when the same tools are used at scale, all day, every day?
Beyond the immediate findings, the experiment points to a broader question about the future of human cognition. Memory is not a passive clipboard. It is built from a series of actions: paying attention, making an effort to encode the experience, later retrieving it, and connecting it to other ideas. Each of these actions requires mental work, and that work keeps our minds healthy and flexible. Generative AI, by its very nature, can perform many of these functions for us. It can generate an image, summarize a document, recall a fact, or produce a plausible version of an event without any effort on our part. The convenience is real, and so is the risk. Researchers at the University of Pennsylvania, including Gideon Nave, an associate professor of marketing at the Wharton School, and Steven Shaw, are studying how AI is changing human cognition. Nave is not primarily worried that AI will become so powerful that it takes control of the world, the way dystopian movies often imagine. Instead, he fears the slow, quiet decay of the human mind as we outsource the very processes that keep it sharp. “What we are more likely to see is humans starting to decay instead of AI taking over,” he told The Washington Post. He added that he fears a world in which “thinking becomes optional.” This is a deeply humanist worry. We are not losing our memories to a robot rebellion; we are losing them to convenience. The less we practice remembering, the less skilled we become at remembering. The less we work to distinguish real from synthetic, the more vulnerable we are to accepting the synthetic without a second thought. Our mental muscles, like any other muscles, weaken when they are not used.
The implications reach far beyond the laboratory and far beyond personal cognition. Our shared understanding of the world depends on trust in our memories and in the records we create. Courtrooms rely on eyewitness testimony, but if a witness has seen an AI-edited video of the crime before testifying, their memory may no longer be their own. Journalists use photographs and video to document events, but if those images have been algorithmically altered, the historical record itself becomes unstable. People argue about what happened at a family gathering, a political rally, or the scene of an accident; now those arguments can be influenced by images that never existed in the original event. More intimately, our relationships are built on shared memories. A couple remembers the day they met; a family remembers a childhood vacation. If someone has been exposed to AI-generated photographs from that period, the collective memory of the family can be rewritten in subtle but persistent ways. The new memories may be happier, sadder, or merely different—but they will feel just as real as the old ones. Maes’s word “dangerous” seems almost too mild. The danger is not that people will be lied to, though they will be. The danger is that they will no longer be able to tell the difference between a lie and a memory. We have always known that memory can be shaped by suggestion; the MIT study shows that technology can automate the suggestion, directing it at scale, with customised precision. The result could be a society in which the past is not simply rewritten by historians or governments, but by the algorithms of every smartphone, in a continuous and invisible flow.
Where does this leave us, as ordinary people trying to live with technology? It would be easy to take these findings as a reason to despair, or to throw up our hands and accept that our memories are doomed. But the same science that reveals the problem also suggests a path forward. If memory can be shaped by exposure to false media, it can also be protected by awareness and intentional behaviour. We can pause before sharing an image. We can ask where a video came from and whether it matches other evidence. We can make it a habit to recall important moments from our own lives before looking at someone else’s version of them. We can keep the mental muscles of attention, effort, and connection alive, even in a world where AI offers to do the work for us. This is not about rejecting technology; it is about using it with our eyes open. When we see an AI-generated image, we can label it in our minds as a possibility, not a fact. When we are uncertain about a memory, we can reflect before letting a synthetic image fill the gap. Above all, we can remember that our memories are not just data. They are the architecture of our identities, the source of our stories, the ground we stand on. If we let them be silently revised by software, we lose something essential about ourselves. The AI-powered tools will only improve; the question is whether we improve alongside them. The MIT experiment is a warning, but it is also an invitation: to be more conscious, more curious, and more protective of the fragile and beautiful record inside our heads. In an age when thinking can become optional, choosing to think is its own quiet act of resistance.