Hooked by Design

Variable reward is the most addictive reinforcement schedule known to behavioral psychology. It is why slot machines work. It is why social media feeds keep you scrolling. And it is built into every AI coding tool on the market.

Here is how it works. You submit a prompt. You wait. You check the output. Sometimes brilliant, sometimes useless. You do not know which you will get. Psychologists call this intermittent reinforcement. It creates a compulsion loop harder to break than consistent reward. If the machine paid out every time, you would get bored. If it never paid out, you would walk away. The space between is where addiction lives.

A friend of mine fell into that space. He is a senior engineer at a well known SaaS company, married with kids. He once described programming as Christmas morning every day for over a decade. Not someone with a discipline problem. Someone who genuinely loves the craft.

About a year and a half ago, AI coding tools were everywhere. He tried them and was not impressed. The output was lazy, full of shortcuts. For months he kept coding by hand because the tools were not worth the review effort.

Then a new generation of models dropped toward the end of last year. The output was slower but clever. It felt like the model had thought through the problem. For the first time, he looked at AI-generated code and thought: this is actually good. That moment was the jackpot. And once you hit it, you keep pulling the lever.

He started running four or five agents at once, using git worktrees to parallelize while waiting for slow model output. He told himself this was efficiency. Soon he stopped reviewing code entirely. The metrics shifted: diff size and PR count. Bigger meant better. The laptop moved to the bedroom. He woke up at odd hours thinking about the next prompt. It did not stop the crash.

The model that hooked him is worth understanding because its design fed the loop. The older models were fast but lazy, producing shallow results that were easier to distrust. The newer generation was slower and more thoughtful. The quality felt closer to what a careful engineer would write. That quality jump earned his trust. Less review. More prompts accepted blindly. More dopamine hits. Later those models got faster, removing the remaining friction. No more parallel worktrees, no waiting. Prompt, accept, commit. The conditions for addiction improved with every release. The model you trust most is the model most likely to get you.

For a natural morning person, the warning sign was simple: he stopped wanting to wake up. The thing that felt like Christmas morning for over a decade suddenly felt like nothing. He took over a month offline, watched streaming shows for the first time in years, did not touch a line of code. It was the longest break of his career.

He came back different. He still generates most of his code with AI. He just reviews every line now. One thing at a time. Between prompts, he catches up with his team instead of queuing the next task. He sleeps fine. That is the metric now.

The psychology here is well studied. B.F. Skinner demonstrated variable reward with pigeons in the 1950s. When food dropped unpredictably, the pigeons pecked the lever obsessively, thousands of times without reward. The uncertainty was the hook. Replace the food pellet with a brilliant code output and the pigeon with an engineer. Same mechanism. The anticipation of the next good result keeps you prompting long after the productivity stops.

This is not about weak willpower. The products are designed this way. Variable reward is not a flaw. It is the system. Every AI coding tool profits from your engagement. The incentives align to keep you pulling the lever. Recognizing the mechanism is the first step. If you have stopped reviewing code, if your laptop is in the bedroom, if you measure your day in diff size instead of outcomes, the tool is using you. The fix is not to throw it away. It is to understand the lever you are pulling and decide when to stop.

My friend did not lose his job or his family. He lost the thing that made him excited to wake up. It came back. It took a month of doing nothing. If a variable reward schedule can hook a senior engineer with real responsibilities and genuine love for the craft, it can hook anyone. The design does not discriminate.