正文 · AI 翻译
Thanks to IT Home reader Mie Mie Yang for the tip!
IT Home reported on October 19 that, according to The Washington Post, research has found that AI can alter human memories. The experiment began with 24 ordinary photographs showing weddings, city streets and races. Pattie Maes, a professor at the MIT Media Lab, and her colleagues showed the photographs to 200 participants.
After a short break, some participants saw the original images again; others saw versions altered by AI, with people's facial expressions changed, soldiers added to the scenes, or imagery related to climate change inserted.
The researchers then conducted a memory test.
Maes was shocked by the results: participants who viewed the AI-altered images were much more likely to form false memories. More troubling, she said, they were more confident in those false memories.
“Basically, I think all of this is very dangerous,” Maes said.
So far, public concern has largely focused on what some researchers call “cognitive surrender”: as machines take on more tasks that involve thinking, people become increasingly unwilling and less able to scrutinize the information machines provide. But another possibility has received far less attention: AI may be changing our memories themselves.
A growing body of research on perceptions of political events, crimes and even new films suggests that AI-generated images and conversations can implant false details in people's minds, leaving them firmly convinced that things that never happened were real.
Memory does more than preserve the past. It provides knowledge and experience that help us understand the present, anticipate the future, shape our sense of self, maintain relationships, and guide individual and collective decisions.
Yet memory is also a major weakness of the human mind: we remember only a small fraction of what we experience, details disappear, recollections become hazy, and memories themselves change each time we recall them.
AI seems well suited to make up for these shortcomings: it can store vast amounts of information, retrieve it in seconds, and recover details people might otherwise lose forever. But those same abilities carry potential risks. AI systems do more than store the past; they also interpret, summarize and reconstruct past information, and sometimes they make mistakes.
The consequences go far beyond misremembering a photograph or a conversation. Court cases, from car crashes to sexual assault, often rely on witnesses' recollections, even though decades of research have shown that eyewitness memories are highly susceptible to distortion. History itself depends on what individuals and societies choose to preserve, forget and ultimately believe about the past.
Human memory is inherently flawed
Maes is a pioneer in artificial intelligence. Early in her career, which has spanned more than 35 years, she helped develop some of the algorithms that laid the groundwork for the early internet. But she has long been more interested in another possibility for AI: expanding human intelligence rather than replacing it with machines.
For Maes, the issue is personal. As her mother grew older, she began forgetting parts of daily life.
“My mother would forget who had visited yesterday and what people had said to her,” Maes said, adding that her mother had to write herself notes: “‘My daughter said she will come in 2 weeks.’”
Maes began imagining a tool far more powerful than a notebook: a memory assistant that could quietly record the course of a person's life.
The issue is especially urgent for people with Alzheimer's disease and other forms of dementia. These conditions are characterized by memory loss and can rob patients of the ability to live independently.
Maes's research team has built several generations of external memory systems. Memoro and MemPal use AI together with wearable or portable devices to capture moments from daily life for later retrieval. Forgot what someone said in a conversation? You can ask the system. Can't remember where you put your keys? The system may be able to locate them.
The idea is to connect the aging brain to another store of memories instead of making it work harder to improve its own memory.
Early tests with older adults have shown promise. Maes said her mother, now 96, has experienced too much cognitive decline to benefit from the technology. But the scientist is highly optimistic that such systems will one day help many people maintain independent lives as their biological memory declines. Some comparable commercial products are already on the market, including interactive reminiscence tools that record a person's life story and personalized smart glasses that assist with daily activities such as brushing teeth.
But if machines become an extension of our memory, what happens to the memories we form ourselves?
Growing reliance on AI
Another study by Maes offers clear evidence bearing on this question.
In an experiment released last year and currently undergoing peer review, 54 participants were asked to write SAT-style essays on topics such as the meaning of happiness. One group used no online tools; a second could use Google Search but could not use AI-generated answers; a third could use ChatGPT.
During writing, electrodes on their scalps recorded electrical activity in their brains, revealing clear differences: participants who wrote independently had the strongest and most widespread neural connections among different brain regions; those who used Google fell in the middle; and those who used ChatGPT had the weakest brain connectivity.
When later asked to recount the essays they had just written, participants who relied on ChatGPT struggled to remember their own words.
As IT Home understands it, memory is not simply a matter of storing information in the brain. Attention, deliberate effort, retrieving information and making connections all work together to strengthen memories. Generative AI takes away some of that mental work.
Gideon Nave, an associate professor of marketing at the University of Pennsylvania's Wharton School, and Steven Shaw, an assistant professor at King's College London, study AI's effects on human cognition. They have raised a deeper concern. Nave is less worried about a Hollywood-style nightmare in which superintelligent AI seizes control from humanity than about a subtler risk: people may gradually stop exercising abilities that machines can perform for them.
“What is more likely to happen is a decline in people's own abilities, rather than AI taking over the world,” he added, expressing concern that in the future “thinking will become optional.”
Signs of what researchers call “deskilling” have already appeared. Shaw cited several studies showing that people accustomed to using AI at work, such as doctors who rely on AI to diagnose cancer, perform worse once they lose access to that assistance.
A study published in 2025 in The Lancet Gastroenterology & Hepatology revealed one such potential effect. Researchers observed 19 experienced doctors assessing nearly 1,500 colonoscopies and found that, after the doctors began using AI routinely, their ability to detect precancerous lesions independently fell by about 20%.
“They're no longer exercising that mental ability as they used to,” he said. “Use it or lose it.”
For children, he worries about a more fundamental problem: they may outsource mental tasks to AI before their core cognitive abilities have fully developed. “They may never develop essential critical-thinking skills in the first place.”
Of course, this does not mean that using AI tools such as ChatGPT will inevitably diminish people's abilities. Recent research has not established that ordinary AI use necessarily causes cognitive decline or permanent memory damage. Research in this area is still at an early stage, and people have long used tools to offload memory and thinking, including abacuses, clocks, dictionaries and calculators.
But generative AI introduces a new factor. We increasingly rely on it to interpret the information we retrieve, and people readily accept what AI tells them. In experiments by Nave and Shaw, participants followed AI advice at a high rate even when they were given incorrect information.
What happens when AI makes a mistake?
This week, researchers in AI ethics and society are holding their annual conference in Malmö, Sweden, where many are focused on the same question: what happens when artificial intelligence begins to interfere with human memory?
A growing body of research suggests that generative AI can interfere with memory in remarkably lasting ways: distorting memories, blurring the line between people's own ideas and machine-generated content, and even implanting memories of events that never happened.
In one study, researchers had participants watch a video of a crime and then answer questions from a generative AI chatbot. The participants formed 3 times as many false memories as a control group. In another experiment, participants came up with creative ideas either with or without AI assistance; a week later, they struggled to tell which ideas came from the machine and which were their own.
Other experiments show that AI-generated false information can make people remember things that never existed. In one experiment, nearly half of participants falsely remembered remakes of films that were never made, including The Matrix starring Will Smith. In another, an AI-generated video made people believe they remembered a political scandal that never happened. Researchers noted, however, that deepfake videos were not consistently more effective than plain text at creating false memories.
The latest research evidence suggests that this problem is harder to solve than researchers expected.
Companies and government agencies are racing to deploy AI: police are trying to use it to summarize body-camera footage, while hospitals use it to organize medical records. Mattea Sim, an assistant research professor at Georgetown University's Massive Data Institute, said the most common safeguard is to keep humans involved to check for machine errors.
But that approach assumes people can independently verify the machine's output. Sim's research suggests the influence may run the other way: AI can change the judgments and memories people use to spot AI errors, fundamentally undermining the safeguard of human oversight.
In a recent experiment, Sim recruited 331 people to watch a short animated video of a car hitting a pedestrian. A day or 2 later, participants read an AI-generated summary of the accident. Some received a summary that accurately described the traffic sign in the video; others received one in which AI confidently gave incorrect information about the sign.
Among participants who received the accurate summary, nearly 84% correctly remembered the sign; among those who received the misleading summary, only about 45% did. Telling participants in advance that the summary was AI-generated did not prevent it from influencing them.
“People generally know that AI makes mistakes,” Sim said, but the research shows that “human oversight may not be a good solution to this problem.”