Ben Grace

I mapped my AI brain fry

AI doesn't reduce work, it intensifies it. I mapped where the cognitive load lives so I could manage my own AI brain fry and enjoy the tools again.

You’ve probably seen all the articles these days about ‘AI brain fry’. There is research from Berkeley and they found: AI tools don’t reduce work, they consistently intensify it. As the authors put it, “AI makes it easier to do more, but harder to stop.” That problem is exactly what this post is about.

There’s a related concept called cognitive debt. The gap between how fast you can produce with AI and how fast you can actually understand what you’ve made.

“The paradox: the more capability you have, the more you feel compelled to use it. The more you use it, the more fragmented your attention becomes. The more fragmented your attention, the less you actually ship.” ~ Francesco Bonacci, Fortune

The map

I’ve been mapping where the stress actually lives. This is a work in progress, but the first step is awareness. This awareness has helped me take some steps to limit my AI brain fry.

Discovery: You can’t keep up. It’s impossible

What’s emerging that you should care about? Is your workflow and AI knowledge already outdated? A field moving this fast creates real imposter syndrome pressure. The thing is, we can’t keep up. It’s impossible. So, even the most advanced feel like they’re falling behind.

Related: We also risk moving fast for movement’s sake. Intentionality here is more important than ever.

Ingestion: Processing signal you found

Taking AI output and actually comprehending it. Not just reading it. This is the human-speed bottleneck that no model bypasses. Someone has to read it, internalize it, and decide if it’s right. As your role expands (more on that below), you’re increasingly ingesting output from domains where you have less expertise, and that makes it harder still. Not to mention, much of this happens in a little chat window.

This only gets worse when you’re not just ingesting the outputs from your agents, but you’re ingesting the slop that others are outputting and you have to judge if it’s accurate, good, etc.

System orientation: Understanding spatial structure

Holding complex architecture in your head like codebases, data flows, how components connect. This is already difficult. What makes it harder with AI: models are extraordinary at reshaping entire systems in seconds. Refactor a codebase, restructure a sea of markdown files, change the schema. The human is left trying to understand the implications of changes they didn’t make and can’t fully see. AI pivots fast. Human comprehension of what just changed does not.

This can be exhausting when working with AI and the agent suggests what seems like a simple change, but you don’t realize it changes the entire system. You don’t even realize this until three days later when your agent points out you approved the change to the plan.

Communication to AI: How you express intent, interpret responses, and communicate naturally

Giving AI enough context to be useful without a long back-and-forth. We get stuck in these little chat windows all day… just one more bit of polish to the page, the prototype, the code. You realize you’ve been trying to convey something that is common sense to another person. Or, something that a hand gesture or a quick sketch would convey clearly to another human.

I’ve been thinking about this problem a lot lately. I’ve been thinking about how I work with humans, and how I LIKE to work. This led me to two solutions that help me a lot:

  • Loomersidian. We all make a lot of Loom videos. They’re rich in context, easy for humans to digest. Why not AI?

  • Sketching with Excalidraw. I love quick sketches. TLDR: you can share screenshots with annotations to quickly convey context. Also, I’ve connected Excalidraw within Obsidian and LLMs can sketch concepts to me via a sketching skill.

Persistent Memory: Remembering context across time

It can be exhausting to re-teach and re-tell LLMs what you said in another chat. Not losing context between sessions and remembering why a decision was made, what you tried, where you left off. Without this, you re-ingest the same things repeatedly.

Then, at work there are five Slack threads, 12 docs, a handful of Jira tickets: where’s the source of truth? The cost of not having one is invisible until you’re starting over for the third time. AI can compound this.

Expansion without limits

Limits of your role, your time, your attention. We’re only human. This is what Berkeley research names most clearly. Two things are happening at once.

First, task expansion: AI makes adjacent work feel newly accessible, so you absorb responsibilities that used to belong to others. Designers writing code, researchers doing engineering work. The scope of what feels possible and therefore obligatory quietly widens.

Second, boundary blur: because AI makes starting a task so frictionless, work becomes ambient. Just one more prompt before lunch. A quick check before bed. “Just one more prompt” is the same impulse as absorbing three new job functions. “AI makes it easier to do more — but harder to stop.”

Why do this

I’m sharing this because it’s fundamentally changed my relationship with AI. Identifying when and where I have “AI brain fry” has helped me manage my cognitive load and actually enjoy using AI more.