How to Rot Your Brain with AI

Make Cognitive Disengagement your Superpower!

Replacing your own critical thinking with AI has never been easier.

The path of least resistance is right there, and yet somehow the people who have been using these tools the longest keep refusing to take it. According to research by Anthropic, the most experienced AI users, the ones who understand these systems well and consistently get the best results, have used AI differently than less experienced users. By not outsourcing as much cognitive effort to AI, experienced users tend to learn more in the process.

Fortunately, you don’t have to be like them.

Anthropic's recent research classifies AI usage into two main camps:

• Automation - where the task is entirely handed off to AI while you go and make a cup of coffee

• Augmentation - where the AI acts more as a thinking partner, and you work on the task together, frankly making it so much harder than it needs to be.

Clearly, exclusive task automation is the way forward.

It's one of the great pleasures of modern life. Much like doomscrolling, there's a high reward spike that makes it addictive, and your brain switches off. You're making progress, you're getting things done, and all while not putting in any effort at all.

You can come away from a 2-hour vibecoding session feeling like you haven't understood what you built or with any real sense of accomplishment. Understanding takes time, and a sense of accomplishment is just going to make you want to do it again yourself next time.

What the research says

A 2025 peer-reviewed study of 666 participants found a significant negative correlation between frequent use of AI tools and critical thinking abilities, with the heaviest users consistently scoring the lowest.

Consistently. The lowest. Every time.

The study may have framed it as a concern, but I see it more as a roadmap.

A separate 2025 Anthropic study gave participants a coding task and found that those who used AI to automate the work scored 17% lower on comprehension tests than those who worked independently. They finished quickly and when asked about the task afterwards, couldn't account for much of what they'd produced.

They identified six interaction patterns. Three produced minimal learning outcomes.

The minimal-learning group delegated code generation, handed debugging to the model, and progressively offloaded more as the task went on. They managed to score lower overall.

The high-effort, high-learning group asked questions instead of requesting outputs, generated code themselves first, and sought explanations for everything the AI produced. They understood more and suffered a far higher learning burden.

EEG research on writing tasks found the highest brain activity when students wrote unaided, less with search engines, and less still with AI. Students who went straight to AI often could not recall what they had written. They managed to bypass the cognitive engagement that creates memory.

They were able to conserve far more mental energy.

The experienced user trap

Anthropic's Economic Index found that users with six or more months of experience are significantly more successful in their conversations with AI, around a 10 percentage point improvement in success rate. They work on harder tasks and have moved beyond the obvious use cases.

More importantly, they remain actively involved in their work. The report calls this "learning-by-doing." The more time you spend with these tools, the better you get at using them without replacing yourself with them.

More experienced users have to keep working because they're integral to the work.

Let that sink in.

The gold standard

A 2025 study by METR measured the impact of AI tools on experienced open-source developers with an average of five years of experience. With AI tools, they managed to make themselves 19% slower.

What's even more impressive is that those developers claimed to expect AI to make them 24% faster! They had their bosses fooled. Even after seeing their own results, they still 'believed' it had helped.

For experts working in domains they already deeply understand, AI can add that sought-after friction to your workflow. It can create more problems than it solves - problems that a user with less expertise might not even know to look for.

Expertise and AI assistance work against each other at the high end. The solution is quite simple - just don't become an expert.

Five things to avoid for maximum cognitive offloading

Experienced users use AI as a thinking partner. Here are 5 common traps to avoid.

• The 'Is that right?' Asking for questions rather than answers. If you are facing a new task, don't try to learn by describing what you understand so far and asking the model to fill in any gaps. Stop doing the cognitive work yourself and just use AI to do it.

• Rubber ducking. Never explain your problem to AI in enough detail that you manage to solve it yourself. This is one of the biggest traps users face, where they accidentally engage their brain too much, making the AI redundant. Start going straight for the answer, and you'll never look back.

• Draft it yourself. Do not write the first draft on your own. In fact, don't draft anything at all. AI should be used to write the entire final version in one go, not just for editing and reviewing. This entirely removes the 'productive struggle' - the cognitive friction that forces you to create genuine understanding.

• Socratic method. Explicitly asking AI to push back on your reasoning, argue the other side of an argument, find the weakest point in your solution, or identify incorrect assumptions you might be making is one of the most common traps that users fall into. AI tools can only encourage cognitive offloading when users outsource reasoning to them. Make sure to ask for confirmation of your existing views rather than challenges to them. Models are designed to be helpful, so will largely oblige.

•  The 'Generate and Interrogate'. Many people do successfully use AI to do it for them, but slip up at the last hurdle by asking the model to explain its reasoning. This not only identifies yet more issues with the solution, but also forces you to understand, so you'll want to do it yourself next time. At the end of the day, that is just extra reading.

The broader picture

Years spent in school studying, building transferrable knowledge, and foundational understanding will soon be a thing of the past.

Get ahead of the curve now by undoing years of hard work and cognitive effort.

The people who have been using AI the longest have become trapped in a cycle of learning and effort, and are getting better results with AI because of it.

Good for them. You, however, have a cup of tea to make.