AI will make me dumber if I let it. So I made a "helmet."

How I built personal cognitive guidelines for working with AI

Mar 11, 2026

[Updated March 19, 2026] Since this article was published, I've revised one of the five protective instructions described below based on new research on sycophancy bias in large language models. The original version instructed Claude to ask for my initial read before providing its own analysis. I've removed that instruction. For the full story of what changed and why, see my follow-up post.]

The biggest concern I have about using AI on a daily basis for my work is that it’s making me dumber. I often think about the humans in WALL-E who become fat, lazy, and sedentary when robots do everything for them.

I’ve worked hard to build my intellect, and I don’t want it to become jello.

This isn’t abstract for me. I spend most of my working day using AI. I use Claude for research, writing, analysis, brainstorming, and project management. I’ve built automated workflows that pull from my calendar, email, notes, and reading library. I advise nonprofits, foundations, and schools on how to adopt AI strategically.

If AI is quietly degrading the cognition of its heaviest users, I’m in the blast radius.

I was already aware of research on cognitive offloading and skill atrophy, and it had been nagging at me. I didn't need convincing that the risks were real. What I needed was protection. I play hockey and wear a helmet to protect my brain. This comes from the same instinct. I'm not going to stop using AI. It's too valuable. But I'm also going to protect myself and, ideally, get smarter and make my work better in the process.

So I had an idea to build a protective layer: personal guidelines grounded in the research that I could bake into my daily AI habits and, critically, into the AI tools themselves. Naturally, I enlisted Claude to help me review the research, build a framework, and then I taught Claude to enforce the framework in how it interacts with me. More on that later.

Here’s what the evidence says.

Seven risks to cognition

Cognitive offloading is the foundational mechanism. When you delegate cognitive tasks to AI, you reduce your own mental engagement. Gerlich (2025) found a significant negative correlation between frequent AI usage and critical thinking, mediated by cognitive offloading. Humans have always offloaded to tools, but AI expands the scope from memory and arithmetic into analysis, synthesis, and creative generation.

Skill atrophy is the “use it or lose it” problem. When AI handles tasks you used to do, the skills supporting those tasks weaken. Anthropic’s 2026 skill formation study found that developers learning a new library with AI assistance scored 17% lower on comprehension, with the biggest gap in debugging. The FAA found the same pattern in aviation decades earlier and now requires pilots to periodically fly manually.

Critical thinking decline is distinct from skill atrophy. It’s not “I can’t do this” but “I don’t bother to check.” Microsoft Research and Carnegie Mellon (2025) found that higher confidence in AI correlated with less critical thinking. The more you trust the tool, the less you interrogate the output.

Automation bias goes further. This is when you evaluate the output, form your own judgment, and then defer to the machine anyway. A systematic review of 35 studies in AI & Society across healthcare, finance, and national security found that people consistently agreed with incorrect AI recommendations. This research predates LLMs and has decades of experimental evidence behind it.

AI brain fry is the newest concept. BCG and UC Riverside coined it after surveying nearly 1,500 workers and finding that 14% experienced cognitive fatigue from overseeing AI tools. The most draining activity was monitoring AI outputs, not using AI directly. Productivity dropped when workers used more than three AI tools simultaneously.

Output homogenization is the finding that AI-assisted work converges toward sameness. Doshi and Hauser (2024) found AI lifted individual creativity but reduced collective novelty. A follow-up study found this homogeneity persists even months after people stop using AI, creating what they called a “creative scar.” And Jiang and colleagues at the University of Washington and CMU found that different AI models from different companies produce 81-82% similar outputs. Switching between Claude and ChatGPT doesn’t diversify your inputs as much as you’d think.

Agency decay is the progressive loss of independent decision-making. Cornelia Walther at Wharton describes four stages: experimentation, integration, reliance, and dependency. The key insight is the illusion of enhanced agency. You feel more capable because you’re producing more, while actually making fewer independent decisions.

But it depends on how you use it

The same body of research that documents these risks also identifies conditions where AI use genuinely improves cognition. The determining factor, across every study I reviewed, is how you use it.

AI as an explanation partner. Anthropic’s study found that participants who asked AI “why does this work?” scored 65%+, while those who said “do this for me” scored below 40%. A cross-country experiment confirmed: structured prompting enhanced reasoning, while unguided use fostered offloading.

The expertise duality. Klein and Klein (2025) describe how AI levels novice performance on routine tasks but amplifies expert capabilities on complex, frontier-of-knowledge work. Whether AI helps or hurts depends on whether you already have the relevant expertise.

Cognitive load management. AI helps when it reduces unnecessary friction (formatting, data gathering) and hurts when it removes the productive struggle where learning happens. Hong and colleagues (2025) found that when AI handled lower-order tasks while students focused on analysis and evaluation, critical thinking actually improved.

Ambition expansion. AI lets you tackle work you wouldn’t attempt otherwise. A Harvard RCT by Kestin and colleagues found AI tutoring outperformed in-class active learning, but only when designed around evidence-based pedagogical principles.

Equalization on bounded tasks. AI closes performance gaps on well-defined tasks, but the skill doesn’t transfer. Noy and Zhang (2023) found AI closed 75% of the productivity gap between education levels, but when AI was removed, the underlying gap barely changed.

Divergence stimulation. When used to introduce unfamiliar perspectives rather than produce final outputs, AI can counter groupthink. Ashkinaze and colleagues found that AI-generated ideas partially reset the natural convergence in group brainstorming.

The thread connecting all of it: desirable difficulty

Underneath the risks and benefits sits a concept from cognitive science that ties them together: desirable difficulty. Robert and Elizabeth Bjork have studied this since the 1990s. Challenges that slow you down in the short term, like working through errors and struggling with material, actually strengthen learning and retention in the long term.

AI systematically removes this effort. That’s what makes it useful. It’s also what makes it risky.

Bastani and colleagues at Wharton tracked over 200 students learning chess with AI assistance over three months. Students started with restraint. By the end, they were requesting AI help every few moves. In interviews, they admitted they knew this was hurting their learning. But in the moment, they clicked anyway.

The pull of AI assistance is progressive and operates even against conscious intentions. Which is why passive awareness isn’t enough. You need structure.

From research to something I can actually use

Knowing the research is interesting. But I wanted something I could put to work. I wanted a system that both protects me from the risks and helps me use AI in ways that actively make my thinking sharper, my work more distinctive, and my output better. The research gives you the ingredients, but it doesn’t give you a recipe. And I realized early on that any recipe worth following would have to be personal. A generic set of “AI best practices” wouldn’t help, because the risks don’t hit everyone the same way.

So I built a personal framework. I organized the seven risks and six positive conditions into a knowledge base, then did something that turned out to be the most valuable step: I pressure-tested it. Much of the research has real limitations: small samples, self-report measures, short timeframes, and a publication bias toward “AI makes us worse” findings. Several concepts overlap more than a clean taxonomy suggests. And nearly all of the studies were conducted with college students or junior developers, not experienced professionals with deep domain expertise.

That last point changed everything. When I rated each risk for personal relevance to my actual situation, the results surprised me.

Skill atrophy, the most intuitive risk, rated LOW. My core consulting skills, strategic thinking, facilitation design, relationship management, aren’t easily offloaded to AI.

But five of seven risks were rated HIGH:

Critical thinking decline and automation bias because, as a solo practitioner, I have no built-in peer review layer and no second set of eyes on deliverables
Output homogenization because my competitive positioning depends on distinctive thinking
Brain fry because I use multiple AI tools across integrated workflows daily
Agency decay because I’m building increasingly sophisticated AI integrations and moving through the dependency stages faster than most

The pattern: domain expertise protects against some risks but not others. The risks that matter most for me are the ones that operate even when you’re experienced and paying attention.

Two heuristics to protect my brain and improve my thinking

With the framework built, I created two practical tools.

The first is embedded in AI itself. I added persistent instructions to Claude’s memory that change how it interacts with me across all conversations.

𝗖̶𝗹̶𝗮̶𝘂̶𝗱̶𝗲̶ ̶𝗻̶𝗼̶𝘄̶ ̶𝗮̶𝘀̶𝗸̶𝘀̶ ̶𝗳̶𝗼̶𝗿̶ ̶𝗺̶𝘆̶ ̶𝗶̶𝗻̶𝗶̶𝘁̶𝗶̶𝗮̶𝗹̶ ̶𝗿̶𝗲̶𝗮̶𝗱̶ ̶𝗯̶𝗲̶𝗳̶𝗼̶𝗿̶𝗲̶ ̶𝗽̶𝗿̶𝗼̶𝘃̶𝗶̶𝗱̶𝗶̶𝗻̶𝗴̶ ̶𝗶̶𝘁̶𝘀̶ ̶𝗼̶𝘄̶𝗻̶ ̶𝗮̶𝗻̶𝗮̶𝗹̶𝘆̶𝘀̶𝗶̶𝘀̶,̶ ̶𝘁̶𝗼̶ ̶𝗽̶𝗿̶𝗼̶𝘁̶𝗲̶𝗰̶𝘁̶ ̶𝗮̶𝗴̶𝗮̶𝗶̶𝗻̶𝘀̶𝘁̶ ̶𝗮̶ ̶𝗱̶𝗲̶𝗰̶𝗹̶𝗶̶𝗻̶𝗲̶ ̶𝗶̶𝗻̶ ̶𝗰̶𝗿̶𝗶̶𝘁̶𝗶̶𝗰̶𝗮̶𝗹̶ ̶𝘁̶𝗵̶𝗶̶𝗻̶𝗸̶𝗶̶𝗻̶𝗴̶.̶ ̶ [Updated: I've removed this instruction. Research on sycophancy in large language models shows that when you state your position before the model responds, it biases the output toward agreement rather than honest analysis. The fix is to think through your position privately, then let the model respond without knowing where you stand. I also learned a broader lesson: if you give an AI permanent instructions that it follows in every conversation, those instructions need to work across all your different uses. "Ask me what I think first" makes sense for brainstorming, but adds unnecessary friction to straightforward tasks. I now use context-specific prompts for different types of work rather than a single blanket rule. More on this, including links to the research, is here.]

Claude now scaffolds my thinking with frameworks and questions rather than handing me finished outputs, protecting germane cognitive load. It flags generic AI language when writing for me, protecting against homogenization. It surfaces the pattern when I’m asking it to decide rather than inform, protecting against agency decay. And it checks in during long sessions to suggest pausing, protecting against brain fry. All of these can be overridden. They’re defaults, not constraints.

The second is physical. I made a one-pager summarizing the guidelines, color-coded by personal risk level, and taped it above my desk. You can see it here. The physical artifact matters because several of these risks are progressive. They happen in the flow of work, when you’re focused on the task and not thinking about cognitive hygiene. A persistent visual reminder interrupts that drift in a way that good intentions alone don’t.

You use sunscreen when you go to the beach, wear a helmet when you ride your bike, and put on a seatbelt when you get in a car. We can’t eliminate every risk, but we can minimize the ones we know exist with a little forethought and protection.

Jason Weeby

Discussion about this post

Ready for more?