Tokens AnonymousFrançais

Tokens Anonymous. Hellomy name is

and it’s beensince my last prompt.

The fellowship for people who have said “one more prompt and I’ll go to bed”… at 3 a.m.

Take the testRead the 12 steps

Tonight’s shares

Member #001 goes first. The numbers are real: they come from their own usage log, without a single word of their conversations.

Heard at other meetings

It’s 2:47 AM on a Tuesday. I’m not debugging a production outage. There’s no deadline. I’m just watching Claude Code refactor a module — and I can’t stop.

Quentin Rousseau, CTO at Rootly (1)

So addicted to Claude Code, I stayed up 19 hours yesterday and didn’t sleep til 5AM.

Garry Tan, president of Y Combinator (1)

Agentic coding is addictive. You will hear it more and more often.

Steve Yegge (1)

The 12 steps

Adapted from the original program, with all the respect it deserves. Read aloud at the start of every meeting.

  1. 1

    We admitted we were powerless over the Send button, and that our nights had become unmanageable.

  2. 2

    Came to believe that a Power greater than ourselves could restore us to sanity. We checked: it was not a bigger model.

  3. 3

    Made a decision to turn our will and our lives over to an alarm clock, not to --dangerously-skip-permissions.

  4. 4

    Made a searching and fearless inventory of our open sessions: four terminals, two background agents, and one that has been running since Tuesday.

  5. 5

    Admitted to ourselves and to another human being (not a chatbot) the exact nature of our prompts.

  6. 6

    Were entirely ready to give up “last prompt, then I’m done”.

  7. 7

    Humbly asked our brain to tolerate boredom for more than four seconds.

  8. 8

    Made a list of everyone we had bored with “have you tried the new model?”

  9. 9

    Made direct amends to such people, except when it meant writing the message ourselves.

  10. 10

    Continued to take inventory, and when we were wrong, promptly admitted it, instead of regenerating until the AI agreed with us.

  11. 11

    Sought to improve our conscious contact with the real world, the one without a context window.

  12. 12

    Having had a spiritual awakening (around 8 a.m., for once), we tried to carry this message to other prompters, writing it ourselves.

Sound familiar?

Tick what sounds like you. Nothing is sent anywhere: it all stays in your browser.

Sound familiar?

0 of 9

Recreational use

Or denial. Either way, your chair is still here.

This is not a diagnosis. The 9 questions follow the structure of the 9 criteria for gaming disorder, which research scales on generative AI also build on4.

The chips

At meetings, you get a chip for every milestone. These ones you have to earn.

  1. 24 hourswithout a prompt

  2. 7 dayswithout an agent running overnight

  3. 30 dayswithout --dangerously-skip-permissions

  4. 1 yearwithout asking the AI if you’re addicted

For real

A parody site with serious sources. Here’s what holds up, and what doesn’t yet.

What we know

The slot machine
An agent sometimes nails it and sometimes fails, with no warning. That’s variable-ratio reinforcement, the same mechanism that makes slot machines so hard to walk away from. Well-known developers describe exactly this1.
The ADHD brain
In young adults with ADHD, it’s reward processing, not attention deficits, that predicts internet addiction. The study followed 682 people over time2.
The known profiles
One study describes three ways to get hooked on a chatbot: escapist roleplay, the pseudo-social companion, and the epistemic rabbit hole3. One is missing: the builder, who starts a new project every night.
The order of magnitude
A scale tested on 1,500 people in China and the US puts problematic generative AI use at 5 to 10%4.
AI helps, too
People with ADHD who talk about their chatbots online are mostly positive about them6. That’s exactly the problem: nobody gets hooked on something that doesn’t work.

What we don’t know

  • “AI addiction” is not a recognized diagnosis. Some researchers question the concept entirely5.
  • No study directly measures the link between ADHD and generative AI. We’re extrapolating from internet and gaming research.
  • A token counter doesn’t measure suffering. 35 billion tokens can also be a very good month of work.

If it’s not funny anymore

Sleeping badly, pulling away from people, unable to stop even when you want to? Talk to someone. A human.