# What does the book's 'human supervision' rule mean for AI agents?

Date: 2025-07-27 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/books-human-supervision-rule-mean-ai-agents/

**Answer:** The book's rule is that AI tools which take actions are useful but fallible, so a person checks what they do and keeps the judgment calls. With ChatGPT agent, that rule becomes a set of concrete stopping points: you write the limits into the task, approve each consequential step, handle sign-ins yourself, and review the result before anything leaves your hands.

*A Beginner's Guide to ChatGPT* was written when ChatGPT was just starting to reach outside the chat window through plugins, small add-ons that let it look up live information or act on other services. In the chapter on plugins, the book makes a point that has only grown more relevant: these tools are powerful but not infallible. A vague request or an unexpected complication can produce a poor result, so a person needs to check both what the AI says and what it does. Ethical questions, privacy and decisions that need nuance stay with people.

The book also makes a practical observation that is easy to forget: sometimes the automated route is slower. Its example is booking a hotel through a plugin, which can take several rounds of prompts when a direct search on a booking site would be quicker.

On July 17, 2025, OpenAI released **ChatGPT agent**, which lets ChatGPT carry out multi-step tasks on its own virtual computer: browsing, filling in forms, running code, building spreadsheets and slides. An **agent**, in this sense, is an AI that decides and takes the next step toward a goal rather than only answering. That is exactly the situation the book's rule was written for, and OpenAI's own safeguards map onto it closely.

## How the rule maps to ChatGPT agent

| The book's principle | What ChatGPT agent provides | What you do |
|---|---|---|
| AI can misread an ambiguous request | It may ask you for missing details; you can interrupt and correct it mid-task | Write a precise task with limits, and answer its questions |
| A person verifies actions, not only answers | It asks permission before actions with real-world consequences, such as a purchase | Read each confirmation before you approve it |
| Some tasks need active oversight | "Watch mode" is required for critical tasks like sending email | Stay at the screen while it runs |
| Judgment and privacy stay with people | You sign in yourself through "takeover mode," and OpenAI says it does not collect what you type there | Never put passwords or card numbers in the chat |
| Know when doing it yourself is better | It refuses high-risk tasks such as bank transfers | Handle money and sensitive accounts directly |

OpenAI is open about the main new risk: **prompt injection**, where a web page contains hidden instructions meant to hijack the agent, for example to leak data from a connected account. Its advice includes turning off connectors (links to apps such as Gmail or GitHub) when a task does not need them. TechCrunch also reported that OpenAI switched off ChatGPT's memory feature for the agent to reduce that risk.

## A worked approval checklist

Before starting a task, decide where the agent must stop and wait for you. Then say so in the prompt, using the book's method of role, context, task, format and a final check:

```prompt
You are helping me plan a two-night stay for a family of four in Quebec City in mid-August.
Context: budget about 250 dollars a night, need two beds, walking distance to the old town.
Task: research options and fill a comparison table (name, price per night, beds, distance, cancellation policy, link).
Stop and ask me before you: sign in anywhere, add anything to a cart, enter any personal details, or accept any terms.
Do not book or pay for anything. I will book myself.
Format: the table, then your top two picks with one line each.
```

When the agent asks for approval during a run, use this short check every time:

1. **What exactly will happen?** Read the action, the site and the amount or recipient.
2. **Is it what I asked for?** Compare it with your original limits.
3. **Did a web page suggest this step?** If the idea came from a page rather than from you, stop the task.
4. **Can it be undone?** If not, slow down or do it yourself.
5. **What will I check afterward?** Decide now how you will verify the result.

> **Try it:** Give ChatGPT agent a research-only task with the stop rules above, and watch the on-screen narration. Note each moment it pauses for you, and whether any step surprised you.

## Which agent actions need your approval every time?

Some actions are routine enough to let the agent run: searching, reading pages, summarizing and drafting a table inside its own workspace. You review the result at the end, as the book recommends for any AI answer, by checking names, numbers, prices and links against the original sites.

Other actions should never go through without you, however many times you have done them before:

- **Spending money.** Any purchase, deposit, booking or subscription. Check the item, the total and the site.
- **Sending in your name.** Email, messages, posts, form submissions or invitations. Read the final text and the recipient.
- **Signing in.** Take over the browser and type credentials yourself; never give a password or card number to the agent in the chat.
- **Sharing personal information.** Addresses, phone numbers, ID or health details entered into any form.
- **Changing or deleting.** Account settings, files, calendar entries or anything that cannot easily be reversed.
- **Accepting terms.** Agreements, permissions for connected apps and cookie choices.

Money transfers belong in a category of their own: ChatGPT agent is designed to refuse them, and you should keep it that way by doing your banking directly.

## Sources

- A Beginner's Guide to ChatGPT: https://www.amazon.com/dp/B0BZRRPXZW
- Introducing ChatGPT agent: bridging research and action: https://openai.com/index/introducing-chatgpt-agent/
- OpenAI launches a general purpose agent in ChatGPT: https://techcrunch.com/2025/07/17/openai-launches-a-general-purpose-agent-in-chatgpt/