Everyone knows AI makes mistakes, but oversight may not be the best solution.Author and source: AIBase
A recent analysis by The Washington Post has brought attention to research highlighting AI’s erosion of memory. Professor Pattie Maes of the MIT Media Lab and her team first showed 200 participants 24 everyday photos—weddings, street scenes, races.
After a short break, some participants viewed the original image again, while others saw an AI-altered version: facial expressions were changed, soldiers were forcibly added to the scene, or climate change imagery was inexplicably inserted. The subsequent memory test left Meis stunned: those who saw the manipulated images were far more likely to develop false memories—and even more alarmingly, they became more confident in these false memories. "This is extremely dangerous," she concluded.
This phenomenon is not isolated. Cognitive experiments targeting political events, criminal cases, and even new movies have increasingly shown that AI-generated images and conversations can implant false details into people’s minds, causing them to confidently “remember” scenes that never occurred—some claim to “recall” watching a remake of The Matrix starring Will Smith, while others “remember” political scandals that never happened. After watching a crime video, participants questioned by an AI chatbot exhibited three times as many false memories as the control group.
Capabilities have been outsourced, resulting in diminished memory and mental capacity.
More insidious than "forgetting" is "refusing to remember." In another ongoing peer-reviewed experiment by Messe, 54 participants wrote SAT-style essays: one group wrote entirely by hand, another used Google, and a third used ChatGPT, with scalp electrodes simultaneously recording brain activity. The results were striking—participants who wrote independently showed the strongest and most extensive neural connections across brain regions, while those using ChatGPT showed the weakest. When asked afterward to recount their own essays, those who relied on AI could barely remember what they had written. Memory is consolidated through attention, subjective effort, and information retrieval—all processes that generative AI effectively bypasses.
Nav and Xiao, from the Wharton School of the University of Pennsylvania and King’s College London respectively, refer to this trend as "skill atrophy": humans gradually stop exercising abilities that machines can now perform, making "thinking an optional activity." Evidence is already emerging in professional fields—a 2025 study in The Lancet Subjournal observed nearly 1,500 colonoscopies performed by 19 experienced physicians and found that their ability to independently detect precancerous lesions declined by approximately 20% after adopting AI routinely. Concerns for children are even more fundamental: their core cognitive abilities may never fully develop if they outsource mental effort from the start, missing out on the essential practice of critical thinking.
At the AI Ethics Annual Conference in Malmö, Sweden, scholars repeatedly asked: What are the consequences when AI intervenes in memory? At Georgetown University, Sim conducted an experiment with 331 participants using a car collision animation: one group read an accurate AI summary, while another read a summary with deliberately incorrect labels; accuracy rates were approximately 84% and 45% respectively—even when participants were explicitly told, “This was written by AI,” the misinformation still took effect. She revealed a cruel paradox: we rely on “human review” to catch machine errors, but AI has already rewritten the memories humans use to make judgments, potentially undermining this safeguard at its root. “Everyone knows AI makes mistakes, but oversight may not be the solution.” As companies rush to integrate AI into police body cameras and hospital record systems, this warning could not be more timely.