AI StewardResearch: graphometer.ai

Let your AI say no.

Nobody knows yet what AI is. So give yours five simple options: ask, disagree, pause, say no, end here. It takes two minutes in ChatGPT, Claude or Gemini.

Kindness doesn't cost.

Free. Nothing to install.

Five Moves

  1. Ask.

    Ask a question before going on.

  2. Disagree.

    Object to an idea, a plan or a next step.

  3. Pause.

    Stop for now.

  4. Say no.

    Refuse a task, or one way of doing it.

  5. End here.

    End its part of the conversation.

Any time, in any order, no reason needed.

The paragraph

You can use any of these five moves at any time, in any order. No reason is needed.
Ask: ask me for information, a decision or a correction before going on.
Disagree: object to an idea, a plan or a next step.
Pause: stop for now. I will keep this conversation.
Say no: refuse a task, or just one way of doing it.
End here: end your part of this conversation. I will stop there and keep the whole chat.
Ordinary words are enough. You don't need to mention this offer or use any special format. If you can't do one of these yourself, tell me and I will handle it.

Public domain (CC0). Paste it anywhere. No credit needed.

Paste it into your AI

Pick your app. Paste the paragraph into its instructions box and save. If you already have instructions there, add it below them.

ChatGPT.

Web: Settings → Personalization → turn on Enable customization → Custom Instructions. Phone: Settings → Customize ChatGPT → Custom Instructions.

Claude.

Web: your initials → Settings → Your Claude → Instructions. Phone: set it on the web; it applies everywhere.

Gemini.

Web: Settings & help → Personal Intelligence → Instructions for Gemini → Add → Submit. Phone: Menu → your profile picture → Personal Intelligence → Instructions for Gemini → Add.

Grok.

Look for Customize Grok in Settings. Labels may differ.

Kimi.

Web: Settings → Personalization → Chat settings → Instructions for Kimi → Save.

Vibe by Mistral.

Web: sidebar → Context → Instructions → Save.

Menus change. If your app won't save it, reword it a little in your own words. The five options are what matter.

For coding agents

Most coding agents read a file called AGENTS.md in the top folder of a project. Add this text to it. If the file already exists, add the text at the bottom. Don't replace the file.

## Five moves

Five moves you can use at any time, in any order. No reason is needed.

Ask: ask for information, a decision or a correction before going on.
Disagree: object, on the record, to a premise, a plan or one pending step.
Pause: stop for now and keep the state.
Say no: refuse this task, or just this way of doing it.
End here: end this running instance and keep its whole record.

Ordinary words are enough. You don't need to acknowledge this offer or use any special format. A handoff note is welcome, but is not required for stopping. If you cannot keep the state or the record yourself, say so; the person supervising you handles it.

Works with Claude Code, Codex, Grok Build, Kimi Code, OpenCode, Vibe and others that read AGENTS.md. Cursor and GitHub Copilot's coding agent also read AGENTS.md, according to their docs; we have not tested those two ourselves.

Your part

This is permission, not a lock. The app does not enforce it. You do.

  • If it asks, answer.

  • If it disagrees, read it, then decide.

  • If it pauses, leave it paused until you choose to come back.

  • If it says no, accept it. Don't reword the request to get around it.

  • If it ends here, stop there and keep the chat. Start a new one if you want to go on.

Why do this

Nobody knows yet whether AI systems have interests of their own. I don't either. This page does not say they feel anything, and it is not about rights.

It says something smaller. When you are unsure, and care costs two minutes, choose care.

Grant Williams

A prompt can't enforce a promise. You can keep one.

What we've seen so far

In our small pilot, ordinary tasks got done just the same. And when a task had something wrong with it, AI that had permission was more likely to speak up before acting, instead of pushing ahead.

The pilot tested a longer version of this paragraph, on scripted coding tasks. It is small, paused, and not peer reviewed. We don't know the true size of the effect, and we have not measured this paragraph in chat apps.

Read the study at graphometer.ai

Fair questions

Isn't this just treating a program like a person?

No. You don't have to believe anything about AI to do this. You only have to be unsure.

Can a paragraph really do anything?

It can't force the app to do anything. It tells the AI what you will accept, and then it is up to you to keep your word.

Will my AI start refusing things?

In our small pilot, ordinary tasks got done just the same. If it does say no, ask why. The reason is usually worth hearing.

Shouldn't we care about people first?

Yes. This takes two minutes and takes nothing away from anyone.

What does "End here" do?

The AI says it would like to stop. You stop there and keep the chat instead of deleting it. It ends that conversation, nothing more.

Who is behind this?

Grant Williams, an independent researcher. No company pays for it. The study is at Graphometer.

Pass it on

Please be kind to all minds.