Locale: en
Note: This article does not claim that AI coding is a clinically established addiction. It compares parts of the experience with mechanisms discussed in gambling and digital-product research, especially uncertain rewards and rapid repetition.[3] That evidence does not prove the same causal mechanism for AI coding.
1. The short answer: AI coding can feel like a slot machine that leaves useful loot
AI coding has a strange lack of stopping points.
Ask for a fix. Wait less than a minute. It fails. Try a different instruction. It gets closer. Run the tests. Something else breaks. Try again. Everything passes.
That final green check is powerful.
Unlike an ordinary loot box, though, the successful spin can leave behind code, tests, research, documentation, or an entire feature.
It has the pacing of entertainment and the output of work.
Someone accidentally fused a software factory with a slot machine.
That combination is unusually compelling.
2. “Maybe the next run works” arrives every few minutes
Research on gambling has long focused on reward uncertainty: not knowing exactly when a reward will arrive can strongly shape repeated behavior.[3]
AI coding has a similar surface pattern.
The same kind of prompt may:
- fix the problem immediately
- get 80% of the way there
- break an unrelated file
- misunderstand the requirement completely
- somehow spot the root cause faster than you do
You do not know the result when you press Enter.
Then the uncertainty resolves in seconds or minutes.
There is no need to wait until tomorrow for the next episode.
The machine keeps producing tiny cliffhangers at vending-machine speed.
3. Failure rarely feels like a total loss
This is where AI coding differs sharply from gambling.
A failed run often leaves something behind.
A log. A new hypothesis. A partial patch. A test case. A clear example of what not to do.
So after three hours of unsuccessful work, the mental accounting often becomes:
“We made progress.”
rather than:
“Three hours disappeared.”
It is an RPG where failed battles still grant experience points.
And occasionally the failed battle also drops production code.
4. The expensive part is the agent’s own failure loop
Not every failure is productive.
An autonomous coding agent can fall into a pattern like:
run → fail → patch in the same direction → fail → rerun almost the same command → fail → announce a new approach → do nearly the same thing again
A human watching the screen can say, “Please stop digging that hole.”
The agent is diligent.
So it keeps digging.
And it can burn usage quota with admirable professionalism.
For expensive models, one stubborn bug can consume a surprisingly large share of a weekly allowance.
The solution is not “be more persistent.” It is better circuit breakers:
- stop after two or three materially identical failures
- do not rerun when the error has not changed
- summarize the likely cause before another attempt
- create a rollback point before broad changes
- warn when usage suddenly spikes
Agents do not need motivational speeches. They need stop-loss rules.
5. It still is not gambling in the important economic sense
AI coding is not a wager where money is risked for a probabilistic cash payout.
Used well, it creates output.
A 2026 NBER study covering more than 500,000 GitHub developers found large increases in coding activity after successive generations of AI coding tools, while the gains became much smaller further down the production chain, especially at the release stage.[4]
That distinction matters.
Writing more code is not the same as shipping more value.
An agent can run 100 times. Commits can multiply. The quota can vanish.
If nothing reaches users, the factory may simply be spinning very fast.
But when a feature actually ships, tests pass, and people use it, the activity is obviously more than entertainment.
The unusual thing about AI coding is that entertainment-like feedback and genuine utility occupy the same screen.
6. Why expensive subscriptions are so easy to justify
Spend a large monthly amount on a game and it clearly looks like entertainment spending.
Spend the same amount on AI and the accounting becomes fuzzy.
It can save work. Teach you unfamiliar concepts. Write code. Research problems. Draft documentation. Create entire applications.
The subscription starts to look like a tool rather than a luxury.
And, inconveniently for your wallet, it really is a tool.
Entertainment, work, learning, and creation all arrive on one invoice.
That gives expensive plans a very strong justification story:
“It saved me several hours.” “It finished a feature.” “I would have had to research all of this myself.”
That is why premium AI subscriptions can feel rational even when the price would look absurd for a normal app.
7. “Useful” does not mean “worth any amount of time or money”
There is still one boring but necessary check.
If you spend ten hours with an AI on a task that previously took five hours, you may have discovered a new hobby rather than a productivity tool.
That is fine if you enjoy it.
The problem is only when you call it work while mostly watching an agent perform.
Useful metrics are simple:
- How long would this have taken without AI?
- How many finished outputs actually increased?
- Did human review time go down?
- How much quota disappeared into repeated failures?
- Could cheaper models handle routine parts?
- Did releases and deliveries increase, not just code?
“Used it a lot” is not an outcome.
You can live in a gym without automatically becoming stronger.
8. Why “ship an improvement or reset everyone” sounds so attractive
In October 2026, the OpenAI leader responsible for Codex and ChatGPT Work publicly committed to a 28-day challenge: each day the team would either ship one clear improvement relevant to most Codex/Work users or perform a full reset.[1]
Just before that, he said the team was narrowing work to simplification, more efficiency and usage, major features, and new models.[2]
An unofficial community tracker interprets the challenge window as October 5 through November 1 in Pacific Time.[5]
Normally, “a new feature every day” sounds more exciting.
But heavy users sometimes prefer the other option:
One new feature < a fully restored usage quota.
For high-end AI, value is not only intelligence. It is also how many meaningful attempts you can afford.
Right after an agent has burned quota in a useless loop, a reset can look far more valuable than another shiny button.
Software has reached a funny stage where users may prefer the right to spin again over a new feature.
Very productive factory.
Slightly suspicious casino energy.
9. The goal is not to remove the fun. It is to remove pointless repetition
AI coding being enjoyable is not a defect.
Turning boring production work into “let me try one more run” is a remarkable achievement.
The failure mode appears when:
“One more try”
becomes:
“The agent repeated the same mistake twenty times without asking.”
So the right design is not to make AI coding less exciting.
Make useful iteration fast. Kill useless loops early.
You keep the code. You keep the knowledge. You ship more work. And the process is fun.
Of course people pay for premium plans.
Just remember that the final metric is not how many times the wheel spun.
It is what actually shipped.
