A set of recent studies compiled by The Washington Post have brought the erosion of memory caused by AI into the spotlight. Patti Meece, a professor at the Media Lab at MIT, and her team first showed 24 everyday photos—weddings, street views, races—to 200 participants.
After a short rest, some people re-examined the original image, while others saw a version altered by AI: facial expressions were changed, soldiers were forced into the scene, or climate change scenarios were added out of nowhere. The subsequent memory test shocked Meyes herself: those who saw the altered image had a significantly higher chance of forming incorrect memories, and even more terrifying, they became more convinced of these false memories. “It’s very dangerous,” she concluded.
Such phenomena are not isolated. Studies on political events, criminal cases, and even new movies show that images and dialogues generated by AI can embed details that never happened into the brain, making people confidently “remember” imaginary scenes—some people “remember” watching a remake of *The Matrix* starring Will Smith, while others “remember” a political scandal that never occurred. After watching a criminal video, the number of false memories among participants who were questioned by an AI chatbot was three times that of the control group.
Abilities are outsourced, leading to a reduction in memory and mental capacity
What is more hidden than “misremembering” is “not wanting to remember”. In another experiment still under peer review by Meis, 54 subjects wrote SAT-style essays: one group wrote purely by hand, one group using Google, and one group using ChatGPT. Scalp electrodes recorded brain activity synchronously. The results were distinct—independent writers had the strongest and widest connections in various brain regions, while those using ChatGPT had the weakest connections; when asked to retell their essays later, those relying on AI could hardly remember what they had written. Memory is consolidated through attention, subjective effort, and information extraction. Generative AI eliminates this part of mental effort.
Nave from the Wharton School of the University of Pennsylvania and Shaw from King’s College London call this trend “skill degradation”: humans gradually stop exercising the abilities that machines can perform, and “thinking becomes an optional choice”. Evidence has emerged in professional fields—a study published in the 2025 Lancet journal observed 19 experienced doctors performing nearly 1,500 colonoscopies. After using AI routinely, their ability to independently detect precancerous lesions decreased by about 20%. The concern for children is more fundamental: core cognitive abilities have not yet developed, so they may outsource their mental effort, lacking critical thinking training from the start.
At the AI Ethics Conference in Malmö, Sweden, scholars repeatedly asked: What are the consequences when AI intervenes in memory? Sim from Georgetown University conducted an animation experiment with 331 people. One group read accurate AI summaries, while another group read summaries with intentionally incorrect labels. The accuracy rates were approximately 84% and 45% respectively—even after explicitly stating “this is AI-generated,” the misinformation still had an effect. She exposed a cruel paradox: we expect “manual review” to prevent machine errors, but AI first rewrites the memories used by humans for judgment, and this defense may fail from the root. “Everyone knows that AI can make mistakes, but oversight may not be a good solution.” When companies rush to integrate AI into police body cameras and hospital medical records, this warning is precisely timely.