When the Machine Remembers for You
How synthetic memories are changing the relationship between systems, selves and the stories we tell about who we are
In the 1995 film Ghost in the Shell, a refuse collector discovers that his memories are not his own.
He believes he has a wife and a daughter. He remembers his marriage, their separation and the family life he has lost. He even carries a photograph he believes is of his daughter. But when investigators examine his story, they find that none of it happened. He has lived alone for years. There was no marriage. There is no child. His memories were fabricated and implanted to steer his behaviour.
The scene is disturbing because the man does not merely believe a lie. He remembers a life he never lived. And when the deception is revealed, he cannot simply discard those memories. They have already become part of who he understands himself to be.
Thirty years later, researchers at MIT are investigating something that brings this science-fiction premise rather closer to reality. Their findings suggest that artificial intelligence can induce false recollections, increase people's confidence in events that never occurred, and alter their understanding of information they previously encountered.
The technology does not need to enter the brain, as it does in Ghost in the Shell. It can work through images, videos and conversations.
Which raises a more profound question than whether we can trust AI-generated content. What happens when we can no longer be certain that the experiences we remember as our own have not been reconstructed by a machine?
The memory that never happened
In 2025, researchers at the MIT Media Lab, including Pat Pataranutaporn, Pattie Maes (guests of my podcast in 2024) and the psychologist Elizabeth Loftus, published a study titled Synthetic Human Memories: AI-Edited Images and Videos Can Implant False Memories and Distort Recollection.
The team recruited 200 participants and divided them into four experimental conditions. Everyone first viewed a set of original photographs. They were then shown either unedited images, AI-edited images, AI-generated videos, or videos generated from AI-edited images.
The final condition was the most revealing. Participants exposed to AI-generated videos made from manipulated photographs reported roughly twice as many false recollections as the control group, and they were more confident in those inaccurate memories.
In other words, the synthetic evidence did not simply change what people said they remembered. It changed how certain they felt about it.
A second MIT study, Slip Through the Chat, examined 180 participants who read an article and then either read an AI-generated summary of it or discussed it with a chatbot. Those who talked to a deliberately misleading chatbot produced 2.92 times as many false recollections as the control group.
The conversation mattered. A machine that engaged people in dialogue while quietly introducing misinformation was more effective at producing false recollections than a static misleading summary.
An earlier MIT study of AI-assisted witness interviews, released in 2024, found that misleading questions from a generative chatbot produced more than three times as many immediate false recollections as the control condition. A week later, the number of false memories had not fallen, and participants remained more confident in them than the control group.
None of this establishes that AI can wholesale rewrite a person's autobiographical identity. The researchers were studying specific recollections under experimental conditions. But the findings do demonstrate something important: an artificial system can influence not only the information we encounter, but what we subsequently experience as memory.
That distinction matters, because remembering is not simply another form of knowing. Memory is personal. It carries the sense of having been there, of having witnessed something, of having lived through it.
A false statement can be challenged by evidence But a false recollection produces what feels like its own evidence: the subjective certainty of experience. And once that recollection becomes part of our personal narrative, its implications can reach well beyond the original error.
The database of the self
There is a reason this matters so much for identity.
Human memory has never been a perfect recording of the past. For decades, psychologists have shown that remembering is reconstructive. We do not retrieve complete experiences from storage. We assemble recollections from fragments of experience, associations, emotional significance, contextual cues and our present understanding of ourselves.
The psychologist Martin Conway, with Christopher Pleydell-Pearce, developed a model known as the Self-Memory System, which describes the reciprocal relationship between autobiographical memory and what he calls the working self. Our memories help shape our identity. But our current goals, concerns and sense of who we are also influence which memories we retrieve and how we interpret them.
We reconstruct the past partly through the person we have become. This is not a defect in human cognition. It is one of the ways we create continuity across a changing life.
But something new is now entering that process.
Until recently, the materials from which we reconstructed our autobiographical memories came largely from lived experience, personal testimony, photographs, recordings and other forms of documentation. None of these was ever entirely reliable. Families tell competing stories. Photographs exclude as much as they include. Diaries reinterpret events even as they record them.
Generative AI introduces a different capacity into this already imperfect process: evidence itself can now be synthetically altered, extended, animated or generated. And increasingly, the systems that organise our personal archives can also interpret them for us.
A photograph of an event can be turned into a moving image. A conversation can be summarised into a record that leaves out important details. A personal history can be reconstructed from scattered digital traces and presented back to us as a coherent narrative. We may then encounter those reconstructions as evidence of what happened.
The database no longer simply stores representations of the self. It supplies material from which the self is reconstructed.
This is an important development in what I have called the Database Condition. The system is no longer merely recording an identity that exists independently of it. It is becoming involved in the ongoing construction of that identity. And that involvement can happen without anyone intending to manipulate.
When the summary becomes the memory
In September 2026, researchers from Georgetown University and the University of Washington published AI-Enabled Human Memory Manipulation: Misleading AI-Generated Summaries Distort Human Memory.
Participants watched a video of a traffic incident involving a car and a pedestrian. Later, they read a summary of what they had seen. Some received an accurate description. Others were given a summary containing misleading information about a road sign.
Among participants who read the accurate summary, 83.6 per cent correctly recalled the sign. Among those who read the misleading version, accuracy fell to 44.8 per cent. Crucially, it made little difference whether participants were told the summary had been written by AI or by a human transcriber. Knowing the source did not protect them.
The study also tested the summaries themselves. When popular AI models were asked to summarise the videos, they omitted roughly half of the central details on average.
This adds another dimension to the problem. Synthetic memory manipulation does not require a malicious agent deliberately inventing a false past. It can emerge from the ordinary operation of systems designed to make information more convenient.
Consider an AI assistant summarising a meeting, an automated system describing police body-camera footage, or a personal AI searching through years of correspondence to reconstruct a relationship.
A summary is already an interpretation. It selects what matters, imposes emphasis and produces an account of what took place. If that account becomes the version we revisit, quote, share or use to settle a disagreement, it may gradually displace our original understanding of events. The summary becomes part of the evidence from which the memory is rebuilt.
And the more we rely on such systems to organise our lives, the more opportunities there are for their interpretations to become ours.
The researchers do not show that AI summaries are uniquely capable of producing these effects. Human misinformation distorts memory too. What changes with AI is the prospect of automated, personalised and continuously available systems producing such reconstructions across an expanding range of everyday experience.
The significance is structural. The systems helping us remember may also be changing what we remember.
Between the remembered past and the imagined future
There is a telling symmetry within MIT's wider research programme.
The same research community investigating AI-induced false memories has also developed Future You, an experimental system that lets people converse with an AI-generated version of their older self. It builds a plausible future biography from a person's present circumstances, goals and aspirations, and pairs it with an age-progressed image, so that participants can discuss the life they might go on to live.
The researchers found that this kind of interaction could strengthen people's sense of connection to their future selves and reduce reported anxiety. Here, the synthetic narrative is not intended to deceive. It is explicitly presented as a possible future, a tool for reflection and imagination.
Yet the relationship between these projects is philosophically significant. One line of research shows how synthetic media can influence our recollections of the past. Another shows how synthetic narratives can influence our relationship with a possible future. And between them stands the present self.
We have long understood ourselves through a relationship between memory and anticipation. The person we believe we have been helps shape the person we imagine we might become. So what happens when both ends of that relationship are increasingly mediated by artificial systems? One machine may help generate the future we imagine, while another influences the past we remember.
Neither intervention necessarily determines who we become, and technological mediation is not inherently harmful. A carefully constructed imagined future may enlarge our sense of possibility. The distinction lies in whether the system supports our capacity to interpret our own lives, or begins to substitute its interpretations for ours.
There is a difference between using AI to imagine a possible self and allowing AI to prescribe an apparently inevitable one. There is a difference between using AI to recover evidence of the past and allowing it to manufacture certainty about what happened.
The risk is not simply counterfeit memory. It is counterfeit continuity: a coherent story of the self assembled from experiences, interpretations and possibilities whose origins we may no longer recognise.
The question of authorship
This is where the discussion needs to move beyond misinformation.
Most concerns about synthetic content are framed around truth and authenticity. Is the photograph genuine? Was the video manipulated? Did the person really say those words? These questions remain essential. But when synthetic information enters autobiographical memory, the issue becomes more personal.
Who has taken part in constructing the story I tell about myself? Who selected the evidence? Who decided what was significant? Which elements came from my lived experience, which were inferred, and which were generated? And how would I know?
This is not simply a problem of technological accuracy and authenticity. It is a problem of identity and agency.
A person can be entirely sincere about a recollection that is inaccurate. Their conviction tells us nothing about where the memory came from. Authenticity of feeling is not the same as authenticity of provenance.
There are important questions of consent here too. We may knowingly ask an AI to restore an old photograph or help us make sense of a difficult period in our lives. But the terms of that reconstruction should remain visible. The distinction between original evidence, interpretation and invention matters, and the problem becomes far more troubling when that distinction disappears. A machine-generated narrative may be emotionally persuasive precisely because it is personalised, conversational and delivered in the familiar language of our own lives.
This is why cognitive security cannot be confined to defending the brain against direct technological intrusion. There are more ordinary routes by which systems influence cognition: by shaping the information we encounter, the records we keep, the explanations we accept and the recollections we go on to form.
Protecting the self in this environment requires more than detecting manipulated images. It requires preserving our ability to distinguish lived experience from synthetic reconstruction, evidence from interpretation, a possible story from an actual past. And, perhaps most importantly, it requires preserving the individual's authority to question the accounts that systems produce about their own life.
What remains ours?
There is an irony here. Human memory has always been fallible, selective and deeply social. Other people have always influenced what we remember. Our personal histories have never been entirely self-authored.
So the answer cannot be to retreat into a fantasy of perfect, independent memory. Nor should it be to reject technologies that might genuinely help us recover, organise or understand our experiences. AI-assisted recollection could have real value. It might help people navigate vast personal archives, reconnect fragmented histories, or explore alternative readings of difficult events.
But usefulness is not the same as authority. The question is whether these systems make us more capable of examining our own memories, or more dependent on the versions they produce. That is the distinction between assistance and substitution.
Which brings us back to Ghost in the Shell.
The refuse collector's tragedy is not simply that someone has given him a fictional family. It is that the fictional family has become emotionally real to him. The memories retain their meaning even after their factual foundation has been destroyed. His story reveals a troubling asymmetry: a memory can be fabricated, yet the feelings attached to it can be entirely genuine.
Today, researchers are beginning to show how external technological systems can produce smaller, experimentally observable versions of that problem. The consequences for identity remain an open question. But the direction of the inquiry is clear. As machines become more deeply involved in recording, summarising and reconstructing our experiences, we will need to think differently about the relationship between memory and selfhood.
The question is no longer only whether machines can remember us. It is whether they can influence what we remember about ourselves.
If every future is written between the system and the self, we may now need to ask who is helping to write the past. Because the future of you may depend on who gets to reconstruct the person you remember being.
AIES: AI-Enabled Human Memory Manipulation
The Neuron on the Georgetown study
The National on the Georgetown study
UCI on Synthetic Human Memories