Your Brain on AI: Cognitive Offloading, Debt, and Atrophy

Source: u_91c4jx8lri / Pixabay

I cheated in school as a kid. Sort of.

When I was in junior high, I gave other students the answers to homework and test questions. I did it because they asked and because, as someone lying somewhere on the nerd spectrum, I wanted to be liked. Of course, giving out answers wasn’t helping the other kids learn, so you could argue that I was doing them a disservice by enabling them to cheat.

As a parent now, I make my elementary school child do extra-curricular homework. I choose work that challenges him beyond the classroom, and I help him through it, but although he might ask for the easy way out, I never do the work for him. That would, after all, defeat the whole purpose of having him do it in the first place.

Kids today probably don’t ask other students for the answers anymore. Or at least not as much. Why would they, when they can just ask an artificial intelligence (AI) chatbot instead?

Cognitive Debt, Laziness, and Atrophy

AI technology is increasingly being incorporated into education, from elementary school through college. Often, this is happening with an “integrate first” philosophy that reserves questions about safety and utility for later.^1,2 The underlying justification—put forth by the companies with a financial stake in the success of AI and which are signing multi-million dollar contracts with schools to implement their products—is that AI dominance is an inevitability, so kids need to get familiar with it so they don’t get left behind. And yet, both students and parents now worry about whether AI is really helping or actually hurting kids. Some are saying that they even hate AI, both because it’s being forced upon them and because of fears that it might be counter-productive to learning by breeding cognitive laziness.²

What’s the evidence that this is true? One of the first such claims came from an as-of-yet unpublished study from the MIT Media Lab that found lower “cognitive engagement” and “cognitive load” as measured by electroencephalography (EEG) when students used AI chatbots to complete writing tasks compared to those who used Google searches or no such aids.³ Personally, I take this finding with a grain of salt, in part because I’m not convinced that EEG is the best measure of cognitive engagement, but also because the conclusions are completely unsurprising. Of course, there’s lower cognitive engagement when students have chatbots write their essays for them instead of tapping into their own creativity!

This isn’t necessarily a bad thing. People often use chatbots—just like they use calculators—for “cognitive offloading” to save time. Still, the MIT study warned that using AI caused an “accumulation of cognitive debt” in students, suggesting that cognitive offloading comes at a price.

Subsequent studies have clarified what kind of cognitive debt we might incur when using AI. One unpublished study reviewed a series of randomized controlled trials (RCTs) and found causal evidence that the use of AI reduced persistence on a variety of math and reading tasks and impaired subsequent unassisted task performance.⁴ The authors concluded that AI assistance can improve performance in the short-term, but can impair long-term learning by denying people the “experience of working through challenges on their own.” Another published study likewise found that relying on AI chatbot summaries for answers to questions led to “shallower knowledge” based on less effort compared to using Google searches that require “actively discovering and synthesizing information sources themselves.”⁵ This reduction in longer-term performance and depth of knowledge has been described as “cognitive atrophy”³ or "cognitive surrender."⁶

Two other studies produced similar conclusions. One peer-reviewed, survey-based study found a negative association between the use of “AI tools” and self-reported critical thinking skills (based on responses to questions like “How often do you critically evaluate the sources of information you encounter?”).⁷ The other RCT compared the effects of ChatGPT as a study aid to traditional non-AI methods on knowledge retention, tested 45 days later, and found that AI chatbot use impaired knowledge retention, presumably by “reducing the cognitive effort that supports durable memory,” consistent with cognitive laziness and atrophy.⁸ The authors conclude that AI chatbots are a kind of “cognitive crutch” that provides immediate support, but weakens rather than strengthens learning.

To be fair, this summary of existing research isn’t exhaustive, and several of the studies, as mentioned, are unpublished and non-peer-reviewed preprints. But the findings are fairly consistent in supporting that the potential cognitive offloading benefits of AI in education may come at the cost of limiting depth of knowledge, knowledge retention, and critical thinking.

AI Literacy

Returning to the comparison of AI chatbots to calculators, we might think of the dilemma as follows: There’s no doubt that calculators can save us time and potentially reduce error, but if we don’t know or remember how to do long division or trigonometry, then calculators won’t help us learn or reinforce those skills. We won’t know if the calculator is right or wrong if we don’t know how to double-check it. And if our calculator battery dies, we’re left stranded.

So it is with AI, especially for students who need to learn. You can't cognitively offload if you never onloaded in the first place.¹

Here's something else to consider: A recent survey found that 65 percent of young adults aged 18 to 28 years used AI chatbots as a replacement for Google searches.⁹ When users search Google and get an AI summary, they’re half as likely to click on any of the search results.¹⁰ Meanwhile, Google’s AI overviews are inaccurate anywhere from 10 to 28 percent of the time.¹¹ That might not seem like a lot, but since there’s an estimated 5 trillion Google searches a year, we’re talking about a potentially massive amount of misinformation being spread by AI.¹² So, beyond its effects on cognitive atrophy, it can lead people to accept patently false answers at face value.

Artificial Intelligence Essential Reads

Medical AI Just Lost to a General Model

Is AI Slop the New Spam?

A new Congressional bill was recently introduced, awarding National Science Foundation grants to integrate “AI literacy” into K-12 curricula.¹³ The bill defines literacy as “having age-appropriate knowledge and ability to use [AI] effectively, to critically interpret outputs, to solve problems in an AI-enabled world, and to mitigate potential risks.”

The definition of AI literacy seems woefully inadequate to me. How can we teach people—and kids—to “critically interpret outputs” and recognize information if we’re talking AI-distilled answers at face value and using AI as a cognitive crutch that leaves our brains hobbled?

AI literacy should include a firm understanding of how AI chatbots really work, their potential to churn out misinformation and “hallucinations,” and their risk of causing harms related to learning and mental health, including suicide, psychosis, and even violent behavior.

AI Enables Cheating

Some students say they’re not trying to use AI to cheat; they’re just trying to “streamline their studies, become more efficient, and boost their learning.”¹⁴ But others openly admit that cheating is, in fact, rampant in school—even at top universities like Columbia and Yale.^15,16

Much like I did back in junior high, AI chatbots enable children—and adults—to take the easy way out by doing the work for them and giving them an answer. But in doing so, the chatbots are cheating, and by accepting their use as an inevitability, we're cheating ourselves and our kids. And as one college professor reminds us, rather than preparing us for the future, if we let AI do our jobs, then AI may very well take our jobs.¹⁶

For the record, when I help my son with his homework, I do sometimes use a calculator to check his math. But when he gets something wrong, I go through it with him, writing out the problem by hand. Sometimes I have to reacquaint myself with math I last used almost a half-century ago. It takes time, but I think both he and I are the better for it. After all, the effort is part of the point.