At what point does a technological aid become a crutch—or even a handicap?

That question lies at the heart of several recent research efforts into the interactions between human minds and AI chatbots. So far, the findings of these experiments suggest that relying on AI to lighten our cognitive workload may potentially undercut our own capabilities.

In a study published in April, researchers in the U.S. and U.K. found that when people spent just 10 or so minutes using AI to help them solve math or reading-comprehension problems, their own unaided performance on the same types of problems diminished. The people who received help from AI not only fared worse than a second group who had worked without AI assistance, but they also gave up on challenging problems more quickly.

According to the authors of that study, their research provides “causative evidence” that relying on an AI for help reduces persistence and impairs unassisted performance. “People do not merely become worse at tasks, but they also stop trying,” they wrote. If these short-term effects persist with long-term use, then current AI systems “risk eroding the very human capabilities they are meant to support,” they added.

A related study, published last year and led by researchers at the Massachusetts Institute of Technology, found that people who used ChatGPT to help them write an essay “consistently underperformed at neural, linguistic, and behavioral levels” compared to people who wrote the essay without AI. Within a few minutes of completing the task, the people who had worked with an LLM (large language model) struggled to remember what they’d just written. They also did worse than their study counterparts when asked to write an essay without AI’s help.

While these sorts of short-term thinking problems are worrying, some researchers say the greater concern is that, over time, people who rely too heavily on LLMs to do their thinking for them may gradually lose some of their ability to think deeply and critically for themselves.

“When you’re completing a task like writing an essay, even if you consider the task to be not important, you’re training yourself to sift and pick out important pieces of information, and to build an argument and develop a structured chain of thought—all of these important skills that improve with practice,” says Nataliya Kosmyna, a research scientist at MIT and one of the authors of the essay-writing study. “If you skip all that work by using an LLM, you’ve going to start losing those capabilities.”

‘Offloading’ may underlie AI’s cognitive effects  

These findings come at a time when policymakers are considering the potential benefits and harms of AI adoption.

Even before the advent of powerful LLMs, widespread declines in math and reading scores among U.S. students—as well as similarly troubling trends among American adults—had raised concerns about the role modern information technologies are having on our ability to think and reason. As AI-powered tools have swiftly flooded U.S. classrooms and companies, these concerns have intensified.

Many of the possible pitfalls of AI overreliance revolve around a phenomenon neuroscientists call “cognitive offloading,” which refers to the use of external tools to support reasoning, remembering, and other mental processes.

“There is some evidence, at least among adults, that excessive cognitive offloading can lead to a short-term decline in corresponding abilities,” says Kristy Armitage, a research fellow at the University of Queensland in Australia.

While that evidence is far from conclusive, especially when it comes to the latest AI tools, Armitage says caution is warranted. “It feels as if we're entering a qualitatively different era with AI that seems more concerning than other digital thinking tools,” she says. “[AI] fosters the sense that it can replace the need for independent thinking, and there are clear indications that people are using it in this way—to bypass the need to develop new skill sets, write assignments or reports, make complex decisions, and so on.”

Some of Armitage’s work has examined cognitive offloading among kids. “This is mainly speculation, but I think it's reasonable to be concerned about children,” she said. “If excessive offloading is interfering with unaided cognitive abilities, then these negative consequences may be far worse for children, as many of these abilities are still developing.”

Not all researchers are so concerned.

“I’m skeptical we’re harming our abilities with artificial intelligence,” says Sam Gilbert, a senior research fellow at University College London. “While there definitely is evidence that when people use cognitive tools they disengage their own cognitive processes, there’s also evidence that we re-engage these processes elsewhere.” For example, someone who relies on AI to compose an email or look up a piece of information may then devote that time and brainpower to some other, more meaningful task. “So it may be more of a rebalancing than a net loss."

Gilbert says that some mental abilities function like muscles in the sense that if we don’t exercise them, they weaken. “If you always offload one particular skill to AI, you’re likely to become deskilled with regards to that specific task,” he explains. “But even if that's happening, I think we may be gaining new skills at the same time, like learning to critically evaluate the output from tools like ChatGPT.” 

He points out that many older forms of technology-supported cognitive offloading—such as relying on a calculator to do math work—also had people worried. But at least in the case of the calculator, research eventually revealed that the technology could actually augment learning and skill development—provided it was introduced in the right ways. “With calculators, we learned that you need an independent phase where you learn first without them,” he says. “There’s an argument that you need to build foundational skills before bringing in technology, and I think that might be similar with AI.”

Understanding how best to use these technologies

While some researchers are considering the effects of artificial intelligence in hypothetical terms, others say the way people are already using AI in real-world settings paints a different and more alarming picture.

In his new book The Convenience Trap, the SBS Swiss Business School professor Michael Gerlich describes how many of the university students he teaches, as well as the business professionals he trains, now turn to AI chatbots for “anything and everything.”

“It’s a different form of offloading than we’ve had before,” says Gerlich, who has also published research on our use of AI. Unlike older online search engines, which would lead users to pages of potentially relevant but also imprecise information—information that users would often have to parse and collate themselves to formulate a useful answer—the latest AIs do all the work for us. “This makes generative AI quite different from other forms of information technology,” he says. “It can lead to what I call the AI-trust spiral, where the more I use AI, the more I trust it, the more work I offload to it, and the less critical I become of what it tells me over time.”

Like other researchers who have published work on AI, Gerlich says the costs and benefits of these tools likely depend on how people use them.

“If you’re including AI at the end of your own thinking process—asking it what you’ve left out, or to give you opposing opinions—that could elevate your own thinking,” he says. But in his experience, this isn’t how most people use these technologies. “People say, ‘Well, I’ll ask the AI first, and then start thinking for myself when I get the results,’ but we know it’s very hard to really think outside the box once you have this well-reasoned, nice sounding answer from an AI.”

Kosmyna, the MIT scientist, echoed many of his sentiments. “Before we all start using LLMs for everything—especially in educational contexts—we need to take more time to understand the best ways to use them,” she says.

The stakes couldn’t be higher. 

“By introducing these tools, you’re toying with the brains of people, and really with the future,” she says. “You are messing with some of the most precious things we have, and to do that without understanding the effects is very dangerous.”