Stop Letting AI Do Your Push-Ups"

It's Monday, February sixteenth. Happy President's Day. If you've got the day off, enjoy it. If you don't, we're sorry, and we'll try to make the next ten minutes worth it either way. So yesterday we talked about your brain's hidden problem-solving system — the one that gives you ideas in the shower, the one that pattern-matches while you're not paying attention. Today we're talking about what happens when you stop feeding it. Because a study published last month found that the more you use AI to do your thinking, the worse you get at the thinking itself. Not a little worse. Measurably, consistently worse — whether you've been coding for two years or twenty. And the uncomfortable part is that the same pattern has shown up before — with Google, with GPS, with calculators. We just weren't paying attention. One story today. It's the one nobody building these tools wants you to think too hard about. This is HigherIQ. Here's the study. Researchers at Anthropic — the company that makes Claude, one of the biggest AI assistants in the world — ran a randomized experiment with fifty-two professional software developers. Half were given access to an AI coding assistant while learning a new programming library. Half had to learn it the old-fashioned way: reading documentation, writing code, hitting errors, fixing them, hitting more errors. When both groups were quizzed afterward on what they'd actually learned, the AI group scored seventeen percent lower. Not because they were worse programmers. Because they'd offloaded the struggle to the machine, and the struggle was the part where learning happens. And here's the detail that should bother you: it didn't matter how experienced you were. Beginners, intermediate developers, experts — the gap was consistent across all of them. AI didn't just hurt novices who hadn't built the foundations yet. It hurt people who already knew how to code and were simply learning something new. The tool made the task feel easier, but feeling easier and being easier are not the same thing. The people who felt the most efficient understood the least. The researchers broke down how participants actually interacted with the AI, and it told a clear story. The ones who scored worst fell into what the researchers called AI delegation. They asked the AI to write the code, it worked, and they moved on. Task complete. Learning: almost zero. They finished fastest and understood least. Then there was a group the researchers called progressive AI reliance — people who started doing their own work but gradually handed more over as the task got harder. The slow slide. You start by asking one question, then another, and before you realize it you've stopped thinking and started managing a machine that's thinking for you. But the group that scored highest did something different. They tried the problem first. They got stuck. They sat with the discomfort of not knowing. And only then did they go to the AI — not to write the code, but to explain a concept they were struggling with. They used it as a tutor, not a ghostwriter. And that distinction — tutor versus ghostwriter — is the entire story. This isn't new. Psychologists have a name for it: cognitive offloading. It's the act of letting an external tool handle a mental task so your brain doesn't have to. And the research on it goes back decades, long before anyone was prompting ChatGPT. In two-thousand and eleven, a psychologist at Columbia named Betsy Sparrow published a study in Science that became known as the Google Effect. She found that when people believed they could look something up later, their brains simply didn't bother storing the information. They remembered where to find it instead of what it said. Your brain is ruthlessly efficient. If it thinks a tool has the job covered, it clocks out. That's not laziness — that's resource allocation. The problem is when the tool covers so much that your brain stops training the muscles it needs. GPS is the physical version of this. A study out of McGill tracked regular drivers and found that people with more lifetime GPS experience performed significantly worse on spatial memory tasks — drawing maps of routes they'd traveled, navigating new environments without help. And the researchers checked: it wasn't that people with bad navigation skills were compensating by using GPS more. People who used GPS more developed worse spatial skills because they outsourced the work. Their hippocampus — the part of the brain that builds mental maps — was getting less exercise. Like any muscle you stop using, it weakened. Now, we should be honest about something. Each of these studies — the AI coding experiment, the Google memory research, the GPS work — has its own limitations. Small samples. Specific populations. The Anthropic paper hasn't been peer-reviewed yet. And stacking three imperfect studies on top of each other doesn't make the evidence three times stronger. But the pattern is hard to ignore. Three different technologies, across two decades of research, all pointing the same direction: when you let a tool handle the cognitive work, the cognitive ability fades. No single study proves that. But if you're waiting for a single study to prove it, you're going to be waiting while it happens to you. Here's where this connects to yesterday. Remember Type one — the fast, unconscious, pattern-matching system? The one that gives you shower epiphanies? Type one doesn't generate patterns from nothing. It recombines things you've already struggled with, already encoded, already banged your head against. It needs raw material. When you offload the thinking to a tool — whether that's GPS, Google, or Claude — you're not just skipping a task. You're depriving Type one of the inputs it needs to make connections later. The shower epiphany only works if you've done the hard thinking first. If the AI did the hard thinking, there's nothing in the shower to recombine. Your lantern is still on, but there's nothing in the room to illuminate. So what does this actually mean for how you operate? Because the answer is obviously not to stop using AI. That ship has sailed. The Anthropic study itself found that certain interaction patterns didn't hurt learning at all. The difference was entirely in how you engaged. Think of it like the difference between riding in a car and driving one. Someone drives you across a city, you arrive, but you couldn't retrace the route. Drive yourself — even with a map — and you encode the turns, the landmarks, the feel of the route. Same destination. Completely different knowledge walking away. AI right now is the world's best chauffeur. And most people are sitting in the back seat. The practical version of this is simple enough to start today. Before you prompt anything — before you type the question, before you ask Copilot to write the function — spend five to ten minutes with the problem yourself. Even if you're bad at it. Especially if you're bad at it. Write the ugly first draft. Sketch the bad solution. Research on what psychologists call productive struggle shows that the discomfort of not knowing is actually the phase where your brain builds the most durable connections. The Anthropic study confirmed it: participants who wrestled with the code before asking for help retained significantly more than those who asked first. And when you do use AI, ask it to explain, not to perform. There's a world of difference between "write me a function that does X" and "I tried to write a function that does X and here's where I got stuck — what am I missing?" The first gets you a working function and an empty brain. The second gets you a working brain and a better function next time. The high scorers in the study weren't anti-AI. They were pro-understanding. This scales beyond coding. It scales to writing — drafting your own messy paragraph before asking AI to clean it up. To analysis — forming a hypothesis before letting AI crunch the numbers. To decisions — thinking through the tradeoffs before a tool summarizes your options. In every case, the principle is the same: do the thinking first, even badly, then use the tool to sharpen it. And here's the meta-question nobody in Silicon Valley is in a hurry to answer. Every major AI company is optimizing for the same thing: reducing friction, getting you to an answer faster. But if the friction is where the learning happens, then removing all of it isn't just an efficiency gain. It's a tradeoff. We're buying speed with skill, and doing it so seamlessly that most people don't feel the exchange. The GPS research showed us this already. People didn't notice their sense of direction declining because the tool always got them where they needed to go. The loss was invisible right up until the moment the tool wasn't available. The developers who fully delegated their coding in the Anthropic study finished fastest. They had the fewest errors. By every visible metric, they were the most productive people in the room. It was only the quiz afterward that revealed they'd learned almost nothing. And in the real world, there is no quiz. There's just the slow, quiet erosion of a skill you stopped practicing and a day when you need it and it's not there. Yesterday we told you your brain has a hidden problem-solver running in the background, connecting dots your conscious mind can't reach. Today's the other side of that coin. That system needs fuel, and the fuel is struggle. The messy drafts. The wrong turns. The five minutes of staring at a problem before you ask for help. Every tool we've ever invented creates the same bargain: convenience now, capability later. The question isn't whether you use the tool. It's whether you let it do the push-ups for you. Stay informed, stay curious, and we'll see you tomorrow.