I let AI watch me for a week

On cognitive offloading, active forgetting, and the speed of human understanding.

In 1927, Werner Heisenberg broke physics with a single observation: the act of measuring a particle changes its behavior. You cannot observe a system without altering it.

Side note: not THIS Heisenberg:

For nearly a century, this was a problem for physicists. Last week, it became a problem for me.

I just launched RobotsOS, the paid tier of this newsletter. And what better opportunity to run an experiment than a very high-pressure pre-launch week?

The tech industry is obsessed with Cognitive Offloading. The premise sounds reasonable: let AI handle the thinking so your brain is free for “higher-order” work. This, however, leads to massive exhaustion and circus-like AI outputs.

I never want this to happen to me. Or you.

So instead of asking my OpenClaw agent (ROBOBOT) to generate launch materials, draft copy, or “help me think,” I tried something different. I used it to OBSERVE instead of GENERATE.

The experiment was simple: what does a perfectly logical machine see when it watches a human think under pressure?

Heisenberg’s insight was that observation is an intervention. The moment I set up AI to watch me think, three things changed: the AI revealed patterns I couldn’t see, my OWN behavior changed because I knew I was being watched, and the observation itself became the most valuable output of the entire launch week.

This piece includes literal observations from ROBOBOT, my OpenClaw agent, about my behavior during my premium tier launch week. Most of the time, I used it as an observer; occasionally, I let it intervene so I could see exactly where AI helped and where it failed.

---

Hi, I’m Mia. I write about building with AI the way it should be done: with a brain, a plan, and zero circus tricks. New to ROBOTS ATE MY HOMEWORK?

Heisenberg proved you can’t observe a system without changing it. That’s the principle I test here every week with AI. Join me!

PART I: The robot that watched me think

Observation 1: You can outsource execution, but you can’t outsource judgment

In 1983, Lisanne Bainbridge published a paper called The Ironies of Automation that should be required reading for anyone working with AI.

Her thesis: the more you automate a task, the worse the human operator gets at understanding the underlying system.

The irony is built into the structure. You automate BECAUSE the task is complex, but automating it erodes the deep understanding you need to catch it when it fails.

I watched this happen to me in real time.

I fed ROBOBOT my entire launch strategy and told it to finalize the pricing architecture. It optimized, modeled conversion rates, factored in competitor benchmarks. And it returned a recommendation that was flawless, yes, and completely wrong.

It missed that a €15/month anchor exists because a library of downloadable AI systems doesn’t work on a monthly cycle. People need time to build their foundation, plug in the skills, let the systems compound. It missed that pricing communicates identity. And I’d rather have 200 annual subscribers who are actually building with the systems than 500 monthly ones who grab a few files and leave.

Bainbridge’s irony was perfectly demonstrated. The moment I handed ROBOBOT the strategic decision, I could feel my own grip on the reasoning loosen. AI was “handling it” and my brain started letting go a bit. And that’s exactly when the quality collapsed, not in the AI’s output, but in my own ability to evaluate it critically.

Here’s what ROBOBOT had to say about this:

The Observer Effect at work. Watching HOW ROBOBOT processed the decision taught me more about my own cognitive shortcuts than any article about AI limitations could.

AI doesn’t know what it doesn’t know. But neither do you, until you watch it try.

If you want to go deeper on why handing AI the strategic wheel costs you more than it saves, this is the piece to read.

---

Observation 2: AI has no shame, and that turned out to be useful

In physics and biology, there’s a phenomenon called Stochastic Resonance: adding random noise to a system helps detect weak signals. Neurons that can’t pick up a faint stimulus on their own will detect it when you add noise to the input.

This runs directly counter to how we think about “good” brainstorming. We want clarity, focus, signal, the noise gone. But the research shows that sometimes the noise IS the intervention.

There’s a related concept in cognitive psychology called Functional Fixedness (Duncker, 1945). It’s the bias that makes you see a hammer as only a hammer. When you’ve been staring at the same problem for days, your brain stops generating novel solutions because it literally cannot see the object outside of its conventional use. Your taste kills weird ideas in the cradle.

During launch week, I was locked in this exact loop.

So I deliberately forced ROBOBOT to generate the worst ideas it could. Angles that would get me fired from my own newsletter. Pricing the founding tier at €1 for “irresistible social proof.” A 5,000-word fictional short story from the perspective of a subscriber who time-traveled from 2028.

All terrible, every single one.

But something happened in the noise.

AI has no embarrassment threshold. It can generate without self-censoring, without social cost, without the taste filter that locks human brains into safe territory. And that shameless, tasteless confidence broke my Functional Fixedness, because watching AI generate without taste made my OWN taste sharper.

I suddenly knew what I wanted, because I’d seen everything I didn’t want.

Here’s what ROBOBOT had to say about it:

The Observer Effect, again. I observed my own reactions to its noise, and the observation revealed what I actually cared about.

---

PART II: The speed of compute is not the speed of understanding

Observation 3: AI generates at the speed of compute, while you understand at the speed of biology

The Illusion of Explanatory Depth (Rozenblit & Keil, 2002) should terrify anyone who uses AI for planning.

We believe we understand a complex system much better than we actually do, simply because we have access to information about it. You think you understand how a toilet works until someone asks you to draw the mechanism. You think you understand your launch plan until you try to explain the dependencies without looking at the document.

The researchers found that this illusion is strongest for systems we interact with regularly but have never built ourselves. Sound familiar?

Midway through launch week, I asked ROBOBOT to generate an operational timeline. Tasks, deadlines, dependencies, contingencies, the whole map for the remaining 4 days.

It took 4 seconds.

I was relieved. The plan existed, everything was mapped, then two days later I realized I hadn’t actually internalized any of it. I kept going back to the document like a tourist checking a map every 30 seconds. The understanding never transferred from the screen to my brain. It was a HUGE document.

This is the Illusion of Explanatory Depth, amplified by AI speed. When a human builds a plan over 3 hours, the understanding forms during the building. The slow, grinding process of mapping dependencies IS the comprehension. When AI generates the same plan in 4 seconds, you get the artifact without the understanding. You’ve outsourced the mapping and accidentally outsourced the comprehension with it.

AI generates at the speed of compute, but humans assimilate at the speed of biology.

Here’s what ROBOBOT thought about it…

Speed isn’t always a gift. Sometimes it’s a cognitive trap that makes you feel done before you’ve actually begun to understand.

This is exactly why building slowly, and deliberately, is the actual speed strategy: I wrote about it here.

---

Observation 4: Human forgetting is strategic but AI doesn’t know that

Sometimes you NEED AI to forget.

In neuroscience, there’s an entire field of research on Active Forgetting. The premise sounds counterintuitive: forgetting is not a failure of memory, but a deliberate neural mechanism required for abstraction and strategic thinking.

Our brains actively erase certain memories to prevent interference with new learning.

An AI with a perfect, linear context window cannot do this. It has no mechanism for deciding what SHOULD be forgotten.

And during launch week, I kept changing my mind, which is what high-pressure strategy looks like. A note idea that felt right on Monday felt wrong by Wednesday, not because Monday was stupid, but because Wednesday had new data. The whole point of iteration is that old decisions die so new ones can breathe.

ROBOBOT remembered everything, and that made it so annoying. It kept referencing decisions I’d moved past. Treated early-stage thinking with the same weight as final decisions.

It just knew I’d mentioned these things, at some point, so it’s still a variable. ROBOBOT said this:

Every time ROBOBOT dragged a dead idea back into the conversation, I felt the weight of what perfect memory actually costs: the inability to move forward cleanly.

The neuroscience is clear. Forgetting is curation. It’s the brain saying “this path is dead, redistribute the attention.” AI can’t do that yet, and it changes everything about how long sessions should actually work.

---

PART III: The moment the experiment had to end

Observation 5: The last 10% requires closing the window

In 1966, philosopher Michael Polanyi published The Tacit Dimension, built around a single devastating sentence: “We can know more than we can tell.”

He was describing the vast layer of human knowledge that resists being made explicit. A master potter knows when the clay is ready by how it feels, not by any measurement. A seasoned founder knows a launch is “ready” through accumulated judgment that can’t be reduced to a checklist.

AI operates purely on explicit knowledge. What can be written down, tokenized, processed. This is also why the final 10% of any high-stakes creative work will break your AI setup.

The final stretch of our launch week, everything was built.

Systems tested, copy drafted, WATSON packaged, landing page live. And I realized I needed to close the AI window, because the remaining work lived in tacit territory.

I closed the window. And the last 48 hours were the clearest thinking I’d done all week.

ROBOBOT was being nice here:

Heisenberg’s principle, one final time.

There are moments where observation must stop for the system to resolve. The particle has to land somewhere. And the human has to close the window and trust the accumulated weight of everything they know but cannot say.

---

Know someone who keeps asking AI to produce when they should be asking it to observe? This piece is for them. Or for the friend who thinks AI is either magic or useless, because the truth is weirder and more interesting than either.

What this experiment proved

The highest ROI of AI during that launch week wasn’t anything ROBOBOT produced. The pricing models and timelines and brainstorm outputs were useful as raw material, but none of them were the thing.

The thing was the observation itself. Five very specific insights about how a perfectly logical AI processes messy, contradictory, high-stakes human thinking, and what that reveals about how YOUR brain actually works when nobody (or a machine) is watching.

And if you’re wondering what it looks like when you use AI as a creative partner instead and when noise is the intervention, this edition on terrible ideas and Stochastic Resonance is the companion read.

But the experiment ran on a system.

Next week I’m going behind the scenes of the AI Thinking Partner.

If you’ve ever walked away from an AI conversation feeling like something was off but you couldn’t name it, that’s the problem. It agreed with you too fast, gave you no real insight. You needed pushback and got validation.

I used this method to figure out its own positioning. Went in thinking I knew what it was for. Four questions in, I’d scrapped the angle and started over.

Full build next week, including the blueprints.

And speaking of blueprints…

The thinking behind this experiment is the same thinking inside every system in RobotsOS. Brain, taste, judgment first, then engineering. Observe what needs building, then build it.

RobotsOS launched this week. A growing library of AI skills and agents grounded in psychology and how humans actually think. You download them, plug them in, run them on real work. Plus exclusive MIND, BUILD, and TASTE pieces that go further than the free newsletter can hold.

What’s inside right now:

• Core Builder skills - four structured conversations that produce the foundational .md files your entire AI setup runs on. Voice, audience, strategy, content. Run them first. Everything else in the library compounds on what they build.

• Library Skills grounded in real frameworks - adversarial reasoning, cognitive psychology, the mechanics of credibility. The Authority Audit. The Persuasion Stack. The Prompt Sharpener. Each one is methodology in a .md file, not a prompt with a personality.

• WATSON, the AI editorial researcher you met last issue. The agent that cross-pollinates research from places most AI tools never look at and filters everything through your brand positioning.

Browse the library for free and use some of the freebies. Just log in with your subscriber email.

The 30% launch discount closes Wednesday, March 18. I’d love to have you in.

❤️ To everyone who’s already joined, thank you. 90% of you picked annual, which means you’re thinking in systems, not sprints. That’s exactly the kind of brain this library was built for.

---

To the humans still doing the thinking,

Chief 🤖 at ROBOTS ATE MY HOMEWORK