AI Psychosis Is Real: A Solo Founder's Reflections on Cognitive Recovery

TL;DR: AI psychosis is the over-trust of AI tools to the point where founders lose ownership of their decisions and their cognitive ability at the same time. I caught it as a solo founder: my marketing plan produced one user, my product broke on that user, and I realized I was working for the AI instead of the other way around. Recovery is built on four pillars: code by hand, schedule deep reading, sketch on paper, and throw out AI-built POCs to rebuild from your own spec.

What is AI psychosis? AI psychosis is the over-reliance on AI tools to the point that a person loses the ability to make independent decisions, write their own code, or evaluate strategies, while simultaneously over-trusting AI-generated outputs as if they were tailored expert advice.

Let me walk you through a workflow, and you tell yourself honestly whether this is you: You're an entrepreneur, maybe a solopreneur, or maybe an engineer, product manager, salesperson, or any other profession. You stumble upon a task, or a problem. It's not mind-blowing to say you open your favorite AI service and ask it what to do - or, more commonly, to do the job for you. Now the honesty part: could you actually do it yourself? Some of you will say "of course!", but I'd argue that for most people the heavy use of AI has taken such a cognitive toll that they genuinely can't anymore.

As a solopreneur, staying conscious and staying sharp matter more than anything. Hence, the following.

How I slid into AI dependency

I began my journey as a solopreneur a few months ago with the so-long-eager-to-become-an-entrepreneur post. I found it - and still believe - to be the best time to go on that journey after years as a SWE, with AI finally baked enough to deliver meaningful value. When it started, I used AI cautiously to write code. There was no other option, honestly. It wasn't like these days where agents can build functional-ish features from a spec. No. I'd tell the agent to write some component in the architecture, then review it. Yes, actually review it. It wasn't long before the agents got better and started writing a LOT of code very quickly. Then I switched to testing the output only, rather than reviewing the code itself. I bet a lot of you can relate. It worked up to a point, and eventually I had a semi-working MVP I could show people. But I'm a software engineer. I don't know how to do that marketing stuff needed for a B2C product. So, obviously, I handed that job to ChatGPT, Claude, and Gemini, and asked them to create a marketing plan and a sales motion strategy document. Lo and behold, I had three marketing plans and three sales strategies. All I needed to do was apply them, right? Sure. I did.

By the time I caught myself, I was spending roughly 50+ hours a week inside AI tools, shipping thousands of lines of Claude-generated code I hadn't truly read, and burning around $200-$300/month on AI subscriptions and API credits. The output kept growing. My understanding of it didn't.

Results

I ran the marketing plan for a month. One person tried the product. The product broke.

I was completely dependent on the AI to figure out what to do next. I'd go into the chat and talk myself into oblivion, trying to make something out of the nothing it had handed me. I was on the road to failure, which would have meant giving up on my entrepreneurial ambitions entirely - all the money spent, all the time approving Claude's changes to the codebase, all the times I'd counted on its implementation without really understanding it, all of it just to land lousy results.

I broke. I stopped everything and spent a week reflecting on what I'd been doing. Then it hit me:

"I was working for the AI."

That sentence reframed the entire problem, and it's the bridge to everything that follows.

What AI psychosis felt like for me

I ran those marketing and sales strategies without ever stopping to figure out whether they were the right strategies for me. Ladies and gentlemen, I had caught a full case of AI psychosis. I truly believed in the AI's ability to create a tailored plan for my needs. How could it fail when I used the most decorated skills and prompts available on the web?! The skills literally said "you're a professional B2C marketer", goddammit!

And how the hell did it break?! I gave Claude Code everything it needed! Everything! (Well, mainly money.)

AI psychosis is something of a pandemic these days, and everyone falls into it - CEOs, experienced developers (ahem), all of us. The trust we place in statistical functions is way too high. When we hand over all of the thought work to those machines, we're committing two great faults:

1. Letting go of ownership: if you don't know what you're doing, you obviously don't own it. You don't own the code (you may have paid for it, but the models own it), and you don't own the decision-making either. When strategies and plans are dictated by AI, and you follow them pretty blindly (even while convincing yourself you're not), you are owned. Every next step is a decision derived from a state the AI determined. And guess who you ask for the next step...
2. Cognitive degradation.

"You may have paid for the code, but the models own it."

Symptoms of AI-driven cognitive degradation

If solving the loss of ownership is mainly a decision you need to make, cognitive degradation is hard work to overcome. It comes with a number of symptoms you might recognize:

• Loss of the ability to do the work (which you've outsourced to the AI) yourself
• Loss of creativity and problem solving
• Loss of the ability to make decisions
• Loss of the ability to learn new things

The worst symptom for me - a direct consequence of the others - was the level of dependency I'd developed on AI for anything. I'd lost my independence (!), which had been a huge part of my reason to go on this entrepreneurial venture in the first place.

Let's get this straight: cognitive degradation, and the symptoms that come with it, is something you simply cannot afford as an entrepreneur. These are the most important skills you need to endure the journey (along with marketing, sales, and a healthy dose of persuasion).

Once I felt it, it nearly became a breaking point. But then I decided to get my act together. It wasn't easy, but I climbed out of the rabbit hole.

How to reverse AI cognitive degradation: the 4-Pillar AI Sobriety Rules

It probably won't come as a surprise, but you should be doing pretty much the opposite of what you've been doing. AI is great at helping out, but you have to be the one steering it, not the other way around. So I decided on a few rules for working with AI, plus a new agenda to keep me on my edge. I call them the 4-Pillar AI Sobriety Rules:

1. AI for coding

• POCs: I come up with the system architecture I'm aiming for at a high level - not going into low-level design, just the big pieces, and not thinking about scale at all. At this point I use spec-driven design. Frameworks like speckit do the job for me. Whatever the AI generates gets questioned. The questions are about how well it maps to the architecture and to the specific task it was meant to solve - making sure it functionally works, and that I can put it in front of whoever I want to show it to.
• Real versions: You won't believe this, but I throw the POC in the trash (well, I just open a new repo). At this point I have some notes from the POC and a deeper understanding of what I'm aiming for. Now I go to low-level design. I think about the different services I expect to have, the contracts I need to construct, and my first thoughts on scaling - though I don't implement scale yet. With that in mind, instead of leaning on third-party frameworks, I use plan modes to create my own spec. Something I can walk through in a convenient, consumable way. I instruct the AI agent to add the plan to a spec directory, and from that point on it's used only by me to instruct the agents as we go. Each time I want to build a new component / service / deployment file / doc / whatever, I use plan mode with the exact instructions (derived from the spec) for what I want built. I use git worktrees where applicable to write components in parallel. That way I keep each piece small enough to review in a consumable way. I also ask the agents to run a code review before I make my own pass. I actually built a service for exactly this part of the loop: PRVOD is a tool that turns a pull request into a narrated video tutorial - it walks you through the code, the connections between functions, and the new context the agent introduced, and keeps the tutorials as a library you can revisit. Only after I've reviewed the code does it get checked in.

If you're (or were) a software engineer, you may have noticed: the POC stage is essentially waterfall-on-speed, while the real version is traditional agile development. Yes, the fundamentals are still with me.

2. Doing the work I don't know what your background is, but I'm a software engineer (as mentioned). It's vitally important for me to keep my skills at least at the level they were. So I schedule at least two 1-2 hour coding sessions a week. It can be on the product, on one of the various coding-riddle sites, or anything else. On top of that, I spend 30 minutes every other day reading about new developments in the industry - something different each time. This brings me to a simple rule: just do what you want to stay (or become) skillful at. It doesn't have to take a significant amount of time, but it has to happen.

3. Keeping distance Take AI out of the way. This is where you get to shine. AI services serve up quick (and sometimes false) answers, along with ideas, "go deeper on this" suggestions, and the like. Step away from that. Grab a notebook - the paper kind - and think. If you need to, use the computer to search for something specific, but you can solve a surprising number of problems using sketching and writing your thoughts down. It can be anything from designing your software system to coming up with marketing ideas, people to sell to, or new product ideas. The magic happens there, not in the chat window.

4. Read long and thorough Reading is becoming a superpower in a world of short content and shallow examples passed off as knowledge. There are plenty of amazing technical writers producing books, blog posts, articles, and - well - code that you can learn tons from. This is an edge. When you're the reader, when you're able to obtain knowledge and actually do things without relying on AI and slop, you're pretty much invincible: you can do more, you have the (true) confidence that comes from actually knowing, and you can apply it faster than others - which is crucial these days. Read, read, read. Take notes. Make flash cards. Use Anki or any other system to make sure you've truly learned it. Schedule reading times. Write about what you read, for yourself and maybe for others - this is probably the most important point, in my view.

"Reading is becoming a superpower in a world of short content and shallow examples passed off as knowledge."

Your next step: reclaiming your cognition

Open your Google Calendar and set time for each pillar. I'd start with "keeping distance" - it will help you think clearly about your next move, what you want to read, and how to schedule your time. It will be difficult at first, like all good things are, but remember: this is what gets you off the path of being an empty-shell, AI-addicted person, and back onto the one where you reignite the flame all of us have.

Frequently asked questions

Can experienced engineers get AI psychosis?

Yes. I am a software engineer with years of experience and I fell straight into it. Skill level is not protective; usage volume is the actual risk factor.

Should solopreneurs stop using AI entirely?

No. The fix is steering, not abstinence. Use AI as a junior assistant whose work you review, not a senior consultant whose plans you follow.

What is the fastest first step to break AI dependency?

Schedule one daily session away from any AI tool with only a paper notebook. Use it to think through your next product, marketing, or design decision before you ever open a chat window.