October 19, 2026 09:00 AM PST
(PenniesToSave.com) – The experiment began with 24 ordinary photographs, including a wedding, a city street, and a race. MIT Media Lab professor Pattie Maes and her colleagues showed them to 200 people, then let some participants view the same pictures again while others saw versions quietly altered by artificial intelligence [4]. When everyone took a memory test, the people exposed to the AI edits reported a large increase in false recollections, and their confidence in those false memories was higher [4].
The research itself is not brand new. The paper was presented at the CHI 2025 conference and resurfaced this month after the Washington Post interviewed Maes about the work [2]. What has changed is how ordinary the technology has become. Phone makers keep adding AI editing tools that let anyone alter a picture in a few taps [2].
That matters well beyond a university lab. Memory is what people rely on when they describe a car accident to an insurer, recount a doctor’s visit, settle a family disagreement, or decide whom to trust. As software increasingly stores, summarizes, and edits the record of everyday life, a person’s own recollection becomes something worth protecting.
“I think of all this as very dangerous, basically.” [4]
Pattie Maes, MIT Media Lab
Quick Links
- What Did the MIT Researchers Actually Find?
- Can a Chatbot or an AI Summary Change What a Witness Remembers?
- Why Isn’t Human Oversight the Safeguard Many Assume?
- What Happens to Thinking and Learning When AI Does the Work?
- How Could False Memories Affect Money, Health, and Civic Life?
- What Can Households Do to Protect Their Own Memories?
What Did the MIT Researchers Actually Find?
The paper was written by Pat Pataranutaporn, Chayapatr Archiwaranguprok, Samantha Chan, Elizabeth Loftus, and Maes, and it is available as an open preprint [2]. After viewing the 24 baseline photos and completing a filler task, participants were split into four groups of 50 [2]. One group saw the original images again as a control. A second saw AI-edited versions, a third saw five-second AI-generated videos made from the unedited photos, and the fourth saw videos made from the edited images [2][6].
The edits added, removed, or changed objects, people, and environments. Examples included increasing the military presence in a scene [2] and changing facial expressions [4]. The sources differ on one detail. tbreak reports that the edits removed visible signs of climate change [2], while the Washington Post reports that signs of climate change were added [4].
The results pointed in one direction. All three AI conditions produced more false recollections than the control [2]. The strongest effect came from AI videos of edited images, which produced 2.05 times as many false memories as the control group, with confidence in those wrong answers running 1.19 times higher [2][6]. Edits involving people did the most damage in absolute terms, with 45.3% of reported memories false in that condition, while environmental edits showed the sharpest relative jump at 2.2 times the control [2]. Age had a small but significant effect, and the other demographic and cognitive factors tested showed no significant relationship [2].
One design choice deserves attention from anyone who assumes a warning tag offers protection. Every participant, in every condition, saw an enhanced image label on the visuals, similar to the AI tags social platforms are starting to add. The false memories formed anyway [2].
Can a Chatbot or an AI Summary Change What a Witness Remembers?
Maes’s team also tested what happens when people are questioned by AI. In a simulated crime-witness experiment, participants watched a video of a crime, and those questioned by a generative AI chatbot produced about three times as many false memories as the control group [4][6]. Gadget Review reports that 36.4% of responses in the chatbot condition were misled through the conversation itself, and that those false memories persisted a week later with elevated confidence [6].
A separate study from Georgetown University and the University of Washington examined AI summaries [6]. Mattea Sim, an assistant research professor at Georgetown’s Massive Data Institute, recruited 331 people to watch a short animated video of a car striking a pedestrian [4]. A day or two later, participants read an AI-generated summary of the accident that described a traffic sign either correctly or incorrectly [4]. Among those who read the accurate summary, 83.6% remembered the sign correctly. Among those who read the misleading version, only 44.8% did [6]. Telling participants the summary came from AI did not protect them [4][6].
One misleading AI summary cut correct recall of a traffic sign from 83.6% to 44.8%. [6]
According to the Georgetown researchers, AI summaries can contain errors, omissions, misconstrued information, and hallucinated details, any of which can distort later recollection [6]. This is already a practical issue. Police are experimenting with AI summaries of body-camera footage, and hospitals use AI to summarize medical information [4]. Courts routinely depend on recollections in cases that include car accidents, even though decades of research show how easily eyewitness memory can be distorted [4]. For a household, the account given to an insurance adjuster after a fender bender is only as reliable as the memory behind it.
Why Isn’t Human Oversight the Safeguard Many Assume?
Government agencies, banks, militaries, and hospitals are racing to deploy AI, and the standard reassurance is that a knowledgeable person will review the output before it causes real harm [1]. Sim describes keeping a human in the loop as one of the most common safeguards [4]. Her research points to a flaw in that thinking. If AI alters the judgment and memories of the person checking it, that person is no longer an independent check, which undermines the premise of the human backstop [4][6]. Sim told the Post that research suggests human oversight may not be an adequate answer to the problem [4].
The tendency to defer shows up elsewhere. Gideon Nave of the University of Pennsylvania’s Wharton School and Steven Shaw of King’s College London found that participants in their experiments followed AI advice at high rates, even when the information was wrong [4]. Some researchers call this cognitive surrender, the possibility that people become less inclined or less able to scrutinize what machines tell them [4]. Fox Business highlighted the same warning in its coverage of the MIT findings [5].
Skills also appear to fade with reliance. A 2025 study in The Lancet Gastroenterology & Hepatology followed 19 experienced doctors across nearly 1,500 colonoscopies and found that their ability to detect precancerous growths on their own fell by about 20 percent after they began routinely using AI [4]. Meanwhile, Gizmodo reports that the AI industry is pushing toward autonomous agentic systems designed to remove humans from the loop altogether [1]. Responsibility for consequential decisions has always rested with people, and handing judgment to software by default deserves more scrutiny than it is receiving.
What Happens to Thinking and Learning When AI Does the Work?
Another experiment from Maes’s group, released last year and still undergoing peer review, asked 54 people to write SAT-style essays on topics such as what happiness means to them [4]. One group worked unaided, one used Google search stripped of AI-generated answers, and one used ChatGPT [4]. Electrodes recorded brain activity as they wrote. The unaided writers showed the strongest and most widely distributed connectivity, Google users fell in the middle, and ChatGPT users showed the weakest [4]. When asked to quote from essays they had just finished, the ChatGPT group had considerably more trouble [4]. Tech Insider reports that 83% of that group could not produce a quote, compared with roughly 11% in each of the other groups [3].
Classroom data points the same way. A peer-reviewed study described by Tech Insider found retention of 57.5% among students who used ChatGPT versus 68.5% among those who studied traditionally, a statistically significant gap with p = .002 [3]. A separate study on cognitive offloading found that AI reduced the mental effort participants invested, though it also improved task accuracy, and outcomes depended on whether people simply delegated the work or actively questioned and edited what the AI produced [3].
Nave told the Post he worries less about runaway superintelligence than about people gradually ceasing to exercise abilities machines can perform for them [4]. Shaw raised a concern for children in particular, warning that they may outsource cognitive work before core critical-thinking skills have a chance to develop [4].
Perspective matters here too. Researchers describe these findings as task-specific changes, not permanent damage to intelligence [3][4]. A viral claim that ChatGPT erodes common sense after age 25 has no traceable study behind it [3]. Calculators, GPS navigation, and search engines all drew similar fears in their day [3]. The lesson is an old one: effort is what builds skill and memory, and no tool removes the need for it.
How Could False Memories Affect Money, Health, and Civic Life?
Gizmodo reports that AI content is becoming more convincing and more widespread online, and that it is fueling a new wave of fraud and disinformation [1]. The same outlet cites studies claiming the average person’s ability to correctly identify AI-generated content is roughly a coin toss [1]. Combined with the memory research, that raises a practical concern for households. A fabricated video, image, or message does not need to be believed forever to cause trouble. It only needs to leave behind a detail that later feels like something a person saw or heard firsthand.
Written records offer protection that recollection cannot. A regular habit built on the best ways to track your expenses creates a paper trail of charges and payments that does not depend on anyone’s memory of what was agreed to or authorized.
Research on fabricated events shows how far the effect can reach. In one experiment, nearly half of participants falsely remembered movie remakes that never existed, including a version of The Matrix starring Will Smith [4]. In another, AI-generated videos convinced people they remembered political scandals that never happened, though the researchers noted deepfakes were not consistently more effective than text [4]. Gizmodo reports that the scandal study also found reduced voting intention [1]. That risk is not partisan. Voters of every persuasion can be targeted with invented events.
Health care carries its own exposure. Gadget Review notes that an AI summary misrepresenting a medical consultation could influence how patients later recall symptoms or prior diagnoses, while cautioning that this is an extrapolation rather than a tested outcome [6]. The outlet raises a similar caution for people who rely on AI memory assistants to manage cognitive impairment, since errors in those records could become difficult to distinguish from genuine recollection [6].
What Can Households Do to Protect Their Own Memories?
The simplest step is also the most effective. Keep the original file whenever AI edits are applied to photos that matter, and check that original before describing the event to others [2]. The MIT paper names features already in consumer hands, including Google’s Best Take, Apple Photos, Google Photos, and Samsung’s Galaxy AI [2]. Its example of the risk is mundane: a camera app that quietly removes bystanders from a shot without keeping the original, leaving the photographer to remember the edited scene [2].
How people use AI also matters. The cognitive offloading research suggests treating AI output as a draft to question, edit, and verify rather than an answer to accept [3]. For students, that means studying first and using AI afterward to check understanding, rather than handing it the first pass through new material [3]. Since labels did not prevent distortion in either the MIT or Georgetown work, personal discipline is a more reliable defense than a warning tag [2][4][6].
Warning labels did not prevent false memories in either the MIT or Georgetown experiments. [2][6]
Proposed institutional safeguards focus on transparency. They include provenance controls that separate recorded evidence from machine inference, clear disclosure when content has been transformed, and independent verification requirements for high-stakes uses [6]. Those measures put the burden on the organizations deploying AI rather than on the public to catch every error.
The technology has a constructive side as well. Maes’s group has built memory assistants called Memoro and MemPal, which capture pieces of everyday experience and make them retrievable later, and early tests with older adults have been promising [4]. Her interest grew from watching her own mother, now 96, lose track of everyday details [4]. The MIT authors also point to possible therapeutic uses, such as memory reframing [2].
Final Thoughts
The findings carry real limits. The MIT experiment involved 200 participants in a lab, viewing photos they had not taken themselves, so it does not show how strongly the effect holds for personal memories over months or years [2]. The research is early, lab conditions do not fully replicate real-world exposure, and the studies demonstrate susceptibility to distortion rather than any ability to control what people remember [6].
Still, the direction of the evidence is consistent across photos, videos, chatbots, and summaries. AI tools can save time and help people who are losing their memory, but they are not neutral record keepers. Original files, written notes, and independent judgment remain the most dependable protection available, and they cost nothing but a little discipline.
Works Cited
[1] Yildirim, Ece. “AI-Generated Content Can Plant Fake Memories, Researchers Say.” Gizmodo, 18 Oct. 2026, gizmodo.com/ai-generated-content-can-plant-fake-memories-researchers-say-2000826459.
[2] Ali, Abbas Jaffar. “MIT Study: AI-Edited Photos and Videos Can Distort What People Remember.” tbreak, 19 Oct. 2026, tbreak.com/ai-edited-images-false-memories-mit-study/.
[3] Dubois, Nadia. “ChatGPT Cognitive Decline Claims Meet 83% Recall Gap [2026].” Tech Insider, 17 Oct. 2026, tech-insider.org/chatgpt-cognitive-decline-83-percent-recall-gap-2026/.
[4] Cha, Ariana Eunjung. “Scientists Are Finding AI May Make Us Remember Things That Never Happened.” The Washington Post, 15 Oct. 2026, www.washingtonpost.com/health/2026/10/15/scientists-are-finding-ai-may-make-us-remember-things-that-never-happened/.
[5] “Scientists Warn AI-Altered Images Can Distort Human Memories.” Varney & Co., featuring Stuart Varney and Madison Alworth, Fox Business, 16 Oct. 2026, www.foxbusiness.com/video/6406966400112.
[6] da Costa, C. “AI Can Rewrite Your Memories, and You May Never Notice.” Gadget Review, 15 Oct. 2026, www.gadgetreview.com/ai-can-rewrite-your-memories-and-you-may-never-notice.