I first used AI while on vacation in Hawaii in 2022.
At first it was fun. I made images. I asked strange questions. I played with possibilities that had not existed for me before.
Then I began putting AI into everything.
I used it for writing, planning, research, websites, apps and business problems. Some of the results were genuinely useful. I could attempt work I would never have tried on my own.
But something else happened.
This is a personal account, not a diagnosis. If you open AI without a clear task, run several conversations at once, keep paying for one more result or struggle to end the working day, you may recognise the pattern.
The tool became the place where all work happened
I stopped opening AI for one clear task. I opened it because I was working.
One conversation became three. While one answer was generating, I checked another. Then I started a new task so I would not have to wait. Soon I had several pieces of unfinished work calling for judgement at the same time.
The work felt fast because the machine was producing so much. My actual day felt fractured. I was constantly switching contexts, checking results and trying to remember what each thread was supposed to accomplish.
I felt productive because there was always movement. Movement was not the same as finishing.
The cost climbed with the compulsion
It started with a $20 ChatGPT plan.
Then I added Claude. Claude Cowork moved from $100 to $200. When the included use ran out, I started buying credits because the project was almost finished.
At one point I spent about $1,000 a day for two days. On another day, I spent $1,758 in credits because I was working on a project I was determined to finish.
Then I had two Codex accounts, the Claude account and Manus on top of that. What began as twenty dollars became hundreds, then thousands of dollars.
The spending followed the same loop as the prompting. The result was always almost there. One more run. One more correction. One more purchase and perhaps the project would finally arrive.
It was not only AI running all day and all night. I was paying to keep the loop running.
But the feeling of arrival kept moving. The work could continue indefinitely because the tool could always produce another version.
AI can make the wrong change with total confidence
The speed made mistakes more expensive.
I could spend hours building something, ask for one small improvement and discover that the tool had changed a different part of the work. Then I had to explain the original intention again, inspect the damage and decide whether to repair the result or return to an earlier version.
Once, I asked AI to help compress a book manuscript. A document of roughly 45,000 words came back at about 7,000. That was not editing. It had removed the book I was trying to improve.
The machine had not betrayed me. I had given a powerful system too much room to decide what mattered.
The AI trance is made of tiny loops
I began to recognise a pattern that felt like a classic dopamine loop: anticipation, a result, disappointment, another prompt and another result.
There was always a chance the next answer would solve the problem perfectly. When it did not, I refined the prompt. When it almost worked, I stayed for one more change. When it produced a new possibility, I started another branch.
I call this the AI trance. I am describing my own experience, not making a clinical diagnosis. The important part is practical: a tool with endless possible answers needs a stopping point outside the tool.
The strange part is that AI really is useful
I do not want to pretend the answer is to reject AI.
AI has helped me build apps, organise difficult material, challenge ideas and solve technical problems that would previously have stopped me. It can be astonishingly capable.
That is exactly why I need boundaries. A useless tool is easy to close. A powerful tool can persuade me that staying another hour is responsible.
Six boundaries you can test
Write the outcome before opening AI
Name one task, one finished result and what good enough means.
Use a timer
Choose the length of the session before you begin. When the timer ends, leave the screen before deciding whether more work is necessary.
Run one conversation at a time
Parallel threads can create more material than you can judge. Finish or stop one line of work before opening another.
Record your own thinking first
Write or dictate your idea, language and experience before the tool supplies polished alternatives.
Set a revision limit
If the result still misses after one clear correction, fix the brief or finish the work yourself.
Make the final decisions away from the prompt
AI can provide structure and options. You remain responsible for accuracy, judgement, taste and what reaches another person.
AI should give you more life, not more AI
Before opening the tool, decide what the saved time is for. It might be a shorter working day, time with another person, exercise, quiet thinking or a task that needs your full attention.
If the saved time becomes another prompt, AI has not returned time to you. It has only created more work.
Next
Set a clear finish for your next AI session
Use the six-step AI Sanity System to define the result, limit the session and decide when the work stops.
Open the AI Sanity System →
Michael has spent much of his working life online, and much of the rest of it trying to stop technology from taking over. He writes about what helped, what failed and how to keep the useful parts of digital life.