By Rafi Moshe, founder of StoriedLife.

Yes, and it is already changing it in two opposite directions at once. The same technology behind the AI biography, a tool that helps someone tell and keep their life story, has now been shown in laboratory research to distort memory itself, planting recollections of things that never happened. Which of those two things AI does to your family’s memories is not decided by the technology. It is decided by a design choice: who stays the author of the story. This article walks through what the research actually found, what memory science already knew, and the questions worth asking of any tool, including ours, before you let it near the most personal archive a family has.
What the new research found
Two studies from the same MIT Media Lab team, both published in 2025, frame the risk precisely.
In the first, published at the CHI 2025 conference, researchers Pat Pataranutaporn, Chayapatr Archiwaranguprok, Samantha Chan, Elizabeth Loftus and Pattie Maes ran a preregistered experiment with 200 participants. Everyone first viewed a set of original photographs. Later, different groups saw the images again in altered forms: unedited, AI-edited, turned into AI-generated videos, or turned into videos generated from the AI-edited versions. When participants were then asked about the original photographs, the AI-altered versions had measurably bent their recollections. The strongest effect came from videos generated out of edited images, which produced just over twice as many false recollections as the control group, and the people in that group were also the most confident in their wrong answers. The researchers are careful about the scope: these were brief, controlled exposures. But their conclusion is plain. AI-edited visuals can implant false memories and distort recollection, and they raise the point themselves that this carries ethical and legal weight alongside any positive uses.
The second study, published at the IUI 2025 conference, is the one that should interest anyone building conversational AI. With 180 participants, the team tested what happened after people read an article and then either had no follow-up, read an honest or misleading summary, or discussed the article with an honest or deliberately misleading chatbot. The misleading summary planted some false memories. The misleading chatbot planted significantly more. Conversation, it turns out, is a powerful channel into memory, in both directions. A dialogue partner we trust can help us recall, and can also quietly edit what we recall.
Hold both findings in mind, because honest context matters here. These studies used deliberately deceptive AI, built to mislead, in laboratory conditions. They are not measurements of what a well-designed tool does in a family kitchen. A related study from the same lab, testing AI interviewers in simulated witness interviews, found the same pattern from the other side: a chatbot that asked leading, suggestive questions produced far more false memories than a plain questionnaire. The consistent lesson across all three is not “AI corrupts memory.” It is that the way an AI talks to you shapes what you remember, which makes the design of that conversation a serious responsibility.
Memory was never a recording anyway
None of this will surprise memory scientists. Elizabeth Loftus, a co-author on both studies above, has spent decades demonstrating that human memory is reconstructive. We do not play the past back like footage. We rebuild it each time we remember, and the rebuilding can be influenced by a question’s wording, a photograph, or another person’s version. Families know this intuitively. It is why two siblings can attend the same dinner in 1987 and produce two incompatible accounts, each sincerely held.
This matters for how we think about preserving life stories. A memoir, a recorded interview, an AI biography: none of them is a court transcript of the past, and any tool that implies otherwise is overselling. What a life story preserves is the person’s own account, in their own words and framing, captured while they are here to give it. That account has errors in it, the way every human account does. It is still irreplaceable, because the alternative is not a more accurate version. The alternative is nothing, or a secondhand summary assembled later by people who were not there.
There is a name for that secondhand problem in our own work: the whisper game. A story retold by children and grandchildren degrades with every handoff, details dropping, emphasis shifting, until the family owns a rumour of the person. Capturing the account directly from the teller, now, is how you stop the game.

The real question: interviewer or author?
If conversational AI can shape memory, the dividing line in this field is not AI versus no AI. It is what job the AI is given.
An AI set to author your story will smooth, embellish and invent transitions, because that is what generation does. An AI set to interview stays in a different posture entirely. It asks. It follows up on what you just said. It asks for the example, the name, the room it happened in. The material all comes from the teller; the machine’s job is to draw it out, keep track of it, and help shape it into something a family can hold. That difference sounds subtle. The MIT research suggests it is the difference that matters, because a system designed to insert content and a system designed to elicit it push memory in opposite directions.
There are honest limits to state alongside that. Follow-up questions can lead as well as elicit, in human interviewers as much as in machines; a good oral historian is trained to ask “what happened next?” rather than “that must have been frightening, wasn’t it?” Any biography tool, AI or human, should be judged on whether its questions open the teller’s account up or quietly steer it. And the teller must stay the editor of last resort: able to read what was made from their words, correct it, and own it outright.
The other direction: what AI makes possible
It would be a shame to read the research as a verdict against the whole idea, because the same technology is doing something for memory that nothing else has managed.
Rafi’s favourite example from building StoriedLife is about language. Some of our storytellers grew up speaking another language, and when they tell their childhood stories in the language those years were lived in, the stories arrive with noticeably more colour: more detail, more humour, more of the person. A grandmother can speak her story in Portuguese, the language of her childhood, and her grandchildren can read it in English. Before conversational AI, honouring both sides of that equation took a bilingual interviewer, a translator and a budget few families have. Now the telling and the reading can each happen in the language that fits the person doing it.
The same is true of patience and availability. A human biographer is wonderful and expensive, and works in scheduled sessions. A parent who remembers something at the kitchen table on a Tuesday can tell it that Tuesday. Memory does not keep office hours, and a tool that is present when the story surfaces will catch stories that a booked appointment never would. None of this requires the AI to understand a life the way a person does. It requires the AI to be a good listener at scale, and the person to remain the source.
How StoriedLife is built, given all of this
We build an AI biography tool, so we do not get to wave this research away; we have to answer it in the design.
Eva, StoriedLife’s biographer, is an interviewer, not an author. She holds conversations, asks follow-up questions drawn from what the storyteller just said, remembers across sittings so a story can deepen over time, and the stories are written from the person’s own account, in their phrasing and their framing. The storyteller and their family own the stories and can download them. Eva does not generate a life you did not describe, and we do not claim the result is a perfect record of the past. It is the person’s own telling, kept. We think that distinction, between a tool that creates content about your life and a tool that helps you tell it, is the whole ballgame, and it is the standard we would suggest holding every tool in this space to, ours included.
If you want to see what that looks like in practice, see how your story can become a book and start with a single conversation.
You can test the distinction yourself: try a free conversation with Eva, or start with 3 free stories, and judge what comes back against your own memory of the telling.
Questions about AI biography and memory
Can AI really change what someone remembers?
Under laboratory conditions, yes. In 2025 MIT Media Lab studies, AI-edited images and videos increased false recollections, with the strongest condition producing about twice the control group’s rate, and a deliberately misleading chatbot produced significantly more false memories than a misleading written summary. These studies used deceptive AI on purpose; they show the channel is powerful, which is why design and intent matter.
Does that mean an AI biography will invent parts of my parent’s life?
It depends entirely on the tool’s job. A generator asked to produce a biography will fill gaps by inventing. An interviewer-style tool, like StoriedLife, builds stories only from what the person says in conversation, and the family owns and can review the result. Ask any provider this exact question before you start: does your AI write from my parent’s words, or does it write about my parent?
Is an AI biography still the person’s real story?
It is their real account, which is a slightly different and more honest thing. Human memory is reconstructive, so any memoir, however it is made, carries the teller’s framing and the occasional wrong date. The value is that the account comes from the person, in their words, captured while they can give it, rather than being assembled secondhand later.
What should we check before using any AI biography tool?
Four things. Who is treated as the author, the person or the software? Can the storyteller review and correct what is produced? Who owns the stories, and can you download them? And does the company describe the result honestly, as the person’s account, rather than promising a perfect preservation of the past?
Sources and how we checked
- Pataranutaporn, P., Archiwaranguprok, C., Chan, S. W., Loftus, E. F., & Maes, P. “Synthetic Human Memories: AI-Edited Images and Videos Can Implant False Memories and Distort Recollection.” Proceedings of CHI 2025. Preregistered study, 200 participants; the 2.05 figure and the confidence finding are from the published abstract: https://www.media.mit.edu/publications/synthetic-human-memories/
- Pataranutaporn, P., Archiwaranguprok, C., Chan, S. W., Loftus, E. F., & Maes, P. “Slip Through the Chat: Subtle Injection of False Information in LLM Chatbot Conversations Increases False Memory Formation.” Proceedings of IUI 2025, pp. 1297 to 1313. 180 participants, five conditions: https://www.media.mit.edu/publications/slip-through-the-chat-subtle-injection-of-false-information-in-llm-chatbot-conversations-increases-false-memory-formation/
- The related witness-interview study from the same lab (Chan et al., generative chatbot interviewers and false memories) is available as a preprint: https://arxiv.org/html/2408.04681v1
- Rafi’s language observation is from his own account of building StoriedLife, also covered in his Senior Safety Advice podcast interview (September 2026): https://www.youtube.com/watch?v=U-DPlpulxN8
- Research and sources checked 26 October 2026.