Attention Thieves
Attention and time are today’s currency. Straight Arrow explores how technology is built to capture both, using human psychology to keep us hooked.
Two eighth-graders who were dating got into a fight. That’s not unusual in middle school, except for this: Neither said a word to the other.
Instead, each opened ChatGPT, typed out their side of the story and let the chatbot speak for them, trading messages back and forth through the chatbot. The chatbot didn’t solve anything and the conflict escalated.
“It ended up going from like a tinder to a fire, and they couldn’t really figure out what was wrong,” said Amanda Bickerstaff, CEO of AI for Education, who heard the story from a school leader.
Eventually, a school administrator stepped in and did something rather simple. They put the two kids in the same room and made them actually talk. It worked, and the fighting stopped.
Download the Straight Arrow app today to get the stories that matter free from manipulation, bias or agenda.™
Point phone camera here
It’s a small story. But it’s a concrete glimpse of a much bigger question that researchers, educators and technologists are increasingly asking: As artificial intelligence gets better at thinking, writing and deciding for people, as it increasingly takes our attention away from the world and the people around us, are humans losing the ability, or the will, to do those things by themselves?
The answer, according to the people studying it, is messier, more contested and more interesting than headlines might suggest.
What the research actually shows
Researchers have a term for what happened to those two eighth-graders: cognitive offloading.
“We cognitively offload when we use an external resource — a calculator, GPS, an AI tool — to do some of our cognitive work as we pursue our goals,” said Evan Risko, a psychology professor at the University of Waterloo in Canada. Along with University College London’s Sam Gilbert, Risko coined the term in a landmark 2016 paper.
The concern that new technology might weaken human thinking isn’t new. In 2008, tech writer Nicholas Carr asked in The Atlantic whether Google was “making us stupid.” Risko told Straight Arrow that this essay inspired the title of his newest paper on the subject, led by Trent Cash and co-authored with Purdue University’s Brooke Macnamara: “Is AI Making Us Stupid?”
What’s actually different about AI, Risko argues, isn’t the concept — it’s the scope.
“Current AI tools are more ‘general’ than tools we’ve used to offload in the past,” he said. “They’re much more a one-stop shop for offloading cognitive work.”
That scope is what a 2025 study by Microsoft Research and Carnegie Mellon set out to measure. Researchers surveyed 319 knowledge workers — professionals who use critical thinking and problem-solving skills in their jobs — about 936 real instances of their use of generative AI at work. The results found that trust cuts both ways. Workers who trusted the AI more reported doing less critical thinking, while workers who had more trust in their own expertise reported doing more.
“The more you have confidence in your own work, the higher the chances you can get the best out of AI collaboration,” said Hank Lee, a doctoral candidate and one of the study’s authors, “because you have more ability and confidence to actually have meaningful collaboration with AI, but also because you have the expertise to make that happen.”
The study also found that generative AI doesn’t necessarily reduce effort so much as relocate it, shifting workers’ attention away from gathering information and solving problems and toward verifying AI’s output and integrating it into their own work.
Does AI create a feedback loop?
A new narrative is emerging in the public conversation about AI and thinking. Use it enough, and your underlying skills start to erode. This pushes users to lean on it even more, in a kind of downward spiral.
It’s an intuitive story — but also, according to researchers, one that’s getting ahead of evidence.
Risko said he was aware of no proof of that kind of feedback loop with AI tools. But he pointed to a related, previously documented pattern: an “Internet Fixation Effect.” This is when people who relied on the internet to answer trivia quiz questions were more likely to keep relying on it later, even when memory alone would suffice. It’s an example of how technology hooks users, presenting itself as indispensable as it usurps their attention — and their personal data.
“The basic idea that reliance can turn into a kind of dependence, though, is sound,” Risko said.
Tina Grotzer, the principal research scientist in education at Harvard University’s Graduate School of Education, landed in almost the same place, independently.
“There are emerging studies that support this narrative, but they are just that,” she said. “It is too early to say … but the speculation is useful in prompting us to do those studies.”
Camille Carlton, senior director of strategy and impact at the Center for Humane Technology, is less cautious.
“I don’t know if we’ve seen a study that specifically looks at this,” she told Straight Arrow. “That said, we’re hearing from a lot of folks, and seeing studies pointing to this being a real concern.”
She compares the brain to a muscle. It weakens with disuse, she said, but that decline is reversible with real effort.
Lee’s research at Carnegie Mellon offers one clue about where the concern might be most warranted — not in AI use broadly, but in a newer, more fully automated style of use.
“Those looping-type use cases are newer,” he said, describing systems in which “the user is pulled off the loop entirely.”
What matters, he argues, is intention — whether a person is still steering or has been removed from the equation altogether. That distinction points to a harder question underneath the first one. How much of this is about individual choices, and how much is about how the tools themselves are built?
Carlton points to a small, familiar design pattern in which the AI, after answering, asks whether the user would like it to proceed to the next step.
“You’re always kind of prompted to spend more time on the platform,” Carlton said.
For Bickerstaff, whose company AI for Education promotes AI literacy, it’s not a design quirk — it’s a design choice.
“The tools are designed to critically offload your thinking, and that is not OK,” she said. “It has to be something the companies themselves take responsibility for.”
It matters how AI is used
If there’s any consensus, it’s that the relationship between AI and thinking isn’t fixed. It depends heavily on how someone is actually leaning on it.
“Yes, the ‘how’ will be critical,” Risko said. He pointed to a 2025 study in which students who used an AI that simply handed them answers showed real learning impairments, while students who used an AI designed to support their thinking — without doing it for them — performed about the same as students who used no AI at all.
His own advice follows from that distinction: don’t let AI take over a task entirely. Risko said people should write the email themselves and then ask AI for feedback, rather than generating something wholesale and hitting send. Even reading what the AI produces, instead of just copying it, helps keep a person in the loop.
Lee’s research points to something similar, with a sharper edge. The real risk, he found, isn’t AI use broadly — it’s AI use in areas in which someone lacks the expertise to know whether the output is actually correct.
“AI could still give you an output that sounds about right,” he said, “and if you don’t have expertise, you can’t tell.”
But that cuts both ways. Experts who understand a task well enough to evaluate AI’s work don’t lose that expertise by using the tool.
“You still need a knowledge worker to have enough expertise to get the best outcome from AI,” he said.
Grotzer, the education researcher from Harvard, frames the skill itself in similar terms.
She encourages students to practice what she calls “parallel metacognition” — understanding what their own mind is good at, alongside what AI is good at, and deliberately deciding which to reach for. It’s less a rule than a habit of mind: people knowing their own strengths well enough to know when they’re the better tool for the job.
What about the children?
For kids, the stakes are arguably higher and harder to see. Eighty-five percent of kids who use AI say they’ve used it for help with schoolwork or homework, according to a Common Sense Media Census of more than 1,200 kids ages 9 to 17 that Bickerstaff cited, and they got there fast. It took social media about three years to reach nearly three-quarters of teens — generative AI hit a similar mark in under two. This left even less time for parents, schools and children themselves to catch up before it had already claimed a piece of their daily attention.
Bickerstaff points to research showing students are getting higher grades on homework and lower grades on exams. Homework, which students can now do with AI’s help, measures something different than an exam, something students generally can’t use AI for. The gap between the two is what’s widening.
The deeper problem, she argues, isn’t that students are offloading too much. It’s that they don’t yet have the judgment to know what’s safe to offload in the first place.
“Young people don’t always know what’s a foundational skill-building activity and what isn’t,” she said. “So they can’t make decisions about what to offload, if anything.”
None of this means avoiding AI. It means using it the way Grotzer describes teaching her own child to shop for a skateboard.
“If my kid wanted to purchase a skateboard, I would encourage them to think about their personal needs, then gather information about different boards, then do the critical thinking about it based upon their criteria, AND consider anything that AI might not be including — the exceptions and edge cases,” she said.
Some of that responsibility is shifting toward the tools themselves, too. Following his research, Lee said product teams at Microsoft have begun experimenting with intentional friction — moments that deliberately slow users down and make them think about their AI use and maybe turn their attention elsewhere — rather than simply automating every step of a task. It’s a quiet acknowledgment that ease and attention protection aren’t the same thing and that sometimes protecting the second one means making a product a little less frictionless.
Whether that becomes the norm or stays the exception may be the real question underneath all the others.
“It’s never been a question of humans versus AI,” said Carlton, of the Center for Humane Technology. “The question is how AI is designed, developed, deployed, and regulated in service of our human needs.”
Round out your reading
- All your questions about napping, answered.
- Trump’s $5,000 promise echoes past payouts that never materialized.
- Inside the effort to make data centers pay their share of electricity costs.
- Why did the Feds seize the ’largest Martian meteorite on Earth’?
- Photos and video show exactly where and when Trump visited Ground Zero after 9/11.
