Using AI is fine. Using it to replace your thinking isn't.
There's a version of AI adoption that looks like efficiency and is actually absence. You've stopped being the person doing the work. You've become the person approving outputs. The strange part is how good it feels while it's happening. The work gets faster, the results look competent, and nobody notices, including you, until the day you need the skill the tool was quietly doing on your behalf.
The Efficiency Trap
The pull toward maximum AI use makes sense. If a tool can draft something in seconds, why spend an hour doing it yourself? If it can summarize a document, why read it?
The logic holds right up until you notice that the thinking you're skipping isn't overhead. It's often the actual point.
Writing a first draft forces you to figure out what you believe. Reading a document means you bump into the one specific claim that changes your understanding. Debugging your own code means you'll recognize the same failure faster next time. The struggle is where the skill gets built. When you hand those tasks off from the start, you collect the output and quietly skip the part that was making you better.
This is the difference between a tool that extends you and a tool that stands in for you. A power drill extends a carpenter. It doesn't make decisions about where the cabinet goes. The moment AI starts making the decisions that used to be yours, the relationship has flipped, and you might not be the one in charge anymore.
Ethan Mollick, a Wharton professor who studies AI's effect on knowledge work, lands on a similar place in his 2024 book Co-Intelligence. He argues for treating AI as a capable collaborator you actively work with rather than a vending machine you pull finished answers from (Mollick, Co-Intelligence, 2024). One of his core rules is plain: be the human in the loop. Not because the model is dumb, but because you bring things it can't, including accountability, real context, and judgment about what actually matters in this specific situation.
What Kasparov Figured Out
Garry Kasparov has more reason than almost anyone to resent machines. In 1997 he lost a match to IBM's Deep Blue, the first time a reigning world champion fell to a computer under tournament conditions. His response was not to declare humans obsolete. It was to ask a better question: what happens when you put the human and the machine on the same side?
In 1998 he helped launch "Advanced Chess," where players used computers as real-time aids during their games. Years later, watching freestyle tournaments where any mix of humans and software could compete, he noticed something he never forgot. The winners weren't the grandmasters with the strongest engines. In his words, "weak human + machine + better process was superior to a strong computer alone and, more remarkably, superior to a strong human + machine + inferior process" (Kasparov, New York Review of Books, 2010).
Read that twice, because it inverts the assumption most people bring to AI. The decisive factor wasn't the strength of the human or the strength of the machine. It was the quality of the collaboration between them. Amateurs who knew how to work with their tools beat experts who didn't.
That's the model worth keeping. The human brings judgment about the problem and what the goal really is. The machine brings speed and pattern matching. Neither alone beats the pairing. But the pairing only works if the human is genuinely in the seat, steering, and not just rubber-stamping whatever the model proposes.
The failure mode was never using AI. It's using it passively.
The Cognitive Cost
Here's the part that's easy to wave away until it happens to you: leaning on a tool changes the brain behind it.
Nick Carr made this case in his 2010 book The Shallows, drawing on neuroscience to argue that the way we use tools reshapes how we think, not just what we get done (Carr, The Shallows, 2010). The brain is plastic. It strengthens what you practice and lets the rest fade. Outsource a mental task long enough and the underlying capability follows the use.
This isn't only a hunch. In 2011, researchers led by Betsy Sparrow published a study in Science showing what they called the Google effect: when people expect to be able to look something up later, they remember the fact itself less well and instead remember where to find it (Sparrow, Liu & Wegner, Science, 2011). We've started treating search as an external hard drive for memory. The general term for this is cognitive offloading, and researchers have spent the last decade documenting how readily we hand mental work to our devices (Risko & Gilbert, Trends in Cognitive Sciences, 2016). AI is cognitive offloading with the volume turned all the way up. It doesn't just store the answer. It produces the reasoning too.
A 2025 study from Microsoft Research and Carnegie Mellon put numbers on the worry. Surveying knowledge workers about how they actually use generative AI on the job, the researchers found that the more people trusted the AI, the less critical thinking they reported applying to its output. The people who kept thinking critically were the ones who trusted their own judgment, not the tool's (Lee et al., Microsoft Research and Carnegie Mellon, CHI 2025). Confidence in the machine and engagement of your own mind moved in opposite directions.
The everyday version is the GPS. It doesn't make you stupid. But if you never navigate without it, you stop building the mental map of the city that comes from occasionally getting a little lost. You arrive at the destination and couldn't redraw the route if your phone died.
The same dynamic runs through writing, code, and any creative work. AI that supplements your thinking is one thing. AI that replaces it is another. From the outside the gap is invisible, because the output can look identical either way. The only person who knows whether you actually thought about it is you.
There's also a skill-level wrinkle worth sitting with. The Brynjolfsson, Li, and Raymond study of customer-support agents found that AI assistance raised productivity most for newer and lower-skilled workers, while experienced top performers gained little (Brynjolfsson, Li & Raymond, NBER, 2023). AI is a powerful equalizer for beginners, and that's genuinely good. But flip it around: if you're experienced and you start leaning on the tool the way a novice does, you may be trading away the hard-won judgment that made you worth more than a novice in the first place.
The Ironies of Automation
None of this is new. Back in 1983, the engineering psychologist Lisanne Bainbridge wrote a short, sharp paper called "Ironies of Automation" that reads like it was written about AI last week (Bainbridge, Automatica, 1983).
Her argument went like this. When we automate a system, we hand the easy, routine parts to the machine and leave the human with two jobs: monitor the automation, and take over when it fails. But those two jobs are in tension. Watching a system run smoothly is boring, and humans are terrible at sustained vigilance. Worse, because the machine now handles all the routine work, the operator's own skills wither from disuse. So at the exact moment the automation hits something it can't handle and kicks control back to the human, that human is the least practiced they've ever been, facing the hardest situation the system can produce.
The more reliable the automation, the sharper the irony. The better it works, the more you trust it, and the more you trust it, the less you're really watching. That inattention is exactly what sinks you on the rare day it fails.
Aviation learned this the expensive way. In 2009, Air France Flight 447 fell out of the sky over the Atlantic after its airspeed sensors iced over and the autopilot handed control back to the crew. The pilots, suddenly flying manually in conditions the automation usually managed, made nose-up inputs that stalled the aircraft, and never recovered it. France's accident investigators, and the safety reviews that followed, raised an uncomfortable point: pilots now hand-fly modern jets so rarely that the basic stick-and-rudder skill they need in a crisis can quietly erode. The fix the industry reached for wasn't less capable autopilots. It was deliberately making pilots fly by hand often enough to stay sharp.
That's the lesson to carry into knowledge work. The danger isn't the tool. It's the slow erosion of the human underneath it, hidden right up until the day the tool can't help and you discover what you've forgotten how to do.
Staying the Author
So what does responsible use actually look like? Not refusing the tool. Using it on purpose.
Start with your own thinking before you ask for help. Sketch a rough outline. Form an opinion. Write the first ugly paragraph yourself. Then bring in AI to push against it, stretch it, or poke holes. Starting from your thinking and starting from the model's output produce very different results, because one sharpens an argument you own and the other quietly installs an argument you didn't.
Push back on what you get. Don't treat the first response as the verdict. Ask it to argue the opposite side. Ask why it chose what it chose. These systems are tuned to sound helpful and agreeable, which means they'll happily confirm a bad idea with total confidence. Your skepticism is the thing standing between fluent and correct.
Keep your name on the work in a way that means something. A simple gut check: if you'd be uncomfortable explaining exactly how something got made, pay attention to that feeling. Not because AI help is shameful. It isn't. But you should be able to stand behind your process, not only the polished thing at the end of it.
Do some of it the hard way on purpose. This is the part most people skip, and it's the one Bainbridge and the pilots would tell you matters most. Pick the skills you care about keeping and exercise them without the tool on a regular basis. Write something start to finish unassisted. Read the whole paper instead of the summary. Debug it yourself before you paste the error in. Treat it like keeping a language alive: use it or watch it fade.
The Human Isn't Just a Failsafe
There's a habit of framing "keep a human in the loop" as a safety measure, a way to catch the machine's mistakes before they cause damage. That framing is true but too small.
The human isn't there only to prevent errors. The human is there because the work needs a perspective that only comes from a person with real stakes in how it turns out. Someone has to know which question is worth asking, and notice when a technically correct answer is still the wrong thing to do. Someone has to actually care whether the result is any good. That's not a temporary capability gap the next model release closes. It's the part of the work that was always the point.
So use the tools. They're good, and they're getting better. Just be honest about which metaphor you're living by. Use AI like a calculator, something that takes a specific chunk of labor off your plate so you can spend your attention on the harder judgment it frees up. Don't use it like an autopilot you stop watching because the routine got boring. Flight 447 is what the second one looks like when the routine ends.
You're not here to review the machine's answers. You're here to be the person who knows whether it was even the right question.
