What this book is about

Imagine asking an AI tool to draft a letter about a billing problem. It produces something useful. You check it, make a few changes, and save time. Next time, you ask it to do more.

That small decision is the starting point of this book. As work becomes easier to hand over, we need to understand what else may move with it. Drafting a reply is one assignment. Sending it to a customer is another. Giving software permission to change a price or reject a request changes the arrangement again.

We already rely on expertise beyond our own. We hire an electrician, trust a teacher, and ask a doctor for advice. We also expect limits, someone to answer questions, and a way to respond when something goes wrong. AI brings those expectations into work that can move much faster and reach many more people.

I use delegation to describe handing over work while retaining control over its purpose and limits. Abdication begins when nobody remains meaningfully answerable. A person may still click an approval button while lacking the information, time, skill, or authority to judge the result.

Accountability asks who must answer. Several people or organizations may share that duty according to their roles. Blame is a later judgment about fault. We need a working system for checking and correcting decisions before we know whether anyone deserves blame.

Consider a school using AI to help students learn. A quick answer may help with today’s assignment. The school still needs to know whether students can explain the work and use the skill independently. In a business using AI to handle complaints, customers need a way to challenge a wrong decision and reach someone who can change it.

Better AI performance can support more delegation. People still have to decide which goals are worth pursuing, which actions are permitted, and which consequences are acceptable. Those choices affect lives beyond the person operating the software.

No technical background is needed to follow these questions. The chapters explore work, health, education, family life, creativity, and public decisions. The practical task is to establish who can act, who can intervene, and who must answer for the result.

When Technology Finally Caught Up With Me

My relationship with technology began in 1968, when I used my first calculator. By 1975, I was programming in Fortran at the University of Pennsylvania, using computers to explore political questions.

I was fascinated by what machines could accomplish. But for most of my life, there was a fundamental limitation. To use technology effectively, I had to learn to think the way the machine worked.

Artificial intelligence has reversed that relationship.

Today, I can begin with an idea, a question, or a problem and pursue it without first mastering the technical machinery required to solve it.

I can research unfamiliar subjects, develop businesses, write books, analyze investments, challenge assumptions, and explore possibilities that once would have required teams of specialists.

But the most important change isn't what artificial intelligence knows. It's what artificial intelligence allows me to do with what I already know.

Seven decades of curiosity, experience, successes, mistakes, and accumulated judgment have suddenly become more powerful.

For most of my life, I had to learn how to think like a computer. At seventy, the computer finally learned how to work with the way I think.

That is what makes this moment so extraordinary.

Artificial intelligence has not replaced my intelligence. It has expanded my ability to use it.

And perhaps that is the greatest promise of this technology. Not that machines will eventually surpass human intelligence, but that human beings will finally have the tools to make fuller use of their own.

The responsibility for what we do with that power, however, remains ours.

From answers to actions

When I first began working with ChatGPT, I thought of it as a glorified Google. I could ask a question and receive an explanation that helped me understand an issue. That was my impression of how I used it, rather than a technical description of how it worked. I read the answer and decided what to do next.

Then my assignments changed. I asked AI to help prepare documents, work on the Living Edition website, and set up notices for readers. Eventually, I authorized a continuing task: notify confirmed subscribers when an approved book update had actually been published. I still decided which changes belonged in the book and when to release them. The system had permission to carry out a defined part of the work.

That experience made the change concrete for me. An explanation arrives for me to consider. An action can alter something outside the conversation. A message can go to another person. A document can be changed. Material can become public under my name. The benefit is considerable: work I would otherwise have to arrange or perform can move forward. The permission behind that work deserves equal attention.

An AI agent is a system that uses AI and tools to take steps toward a goal. Depending on its design and access, it may retrieve information, edit a file, or send a message. A continuing assignment may be started by a schedule or an event. The word “agent” does not tell us how much authority a system has, how reliably it works, or whether a person checks each action.

Even an answer can cause harm if someone relies on it uncritically. Tools add another route for mistakes to have consequences: the system may carry an error into action before anyone reads a reply. Understanding what it can reach and change matters as much as judging what it says.

One task with four different permissions

Consider a reader newsletter. Each instruction below changes the assignment. These are choices about authority, not stages that every reader or every AI system must progress through.

Explain. “Explain how an email newsletter works.” The assignment is to provide information. I assess the explanation before relying on it.

Prepare. “Draft an announcement about this book update.” The assignment is to prepare wording for review. Drafting the message does not authorize sending it.

Act. “Send this approved announcement to confirmed subscribers.” The instruction authorizes a particular action for a defined audience. The content and recipient boundary both matter.

Continue within agreed limits. “When an approved update is published, notify confirmed subscribers, at most once a day.” This grants continuing permission under stated conditions. It requires a way to check what happened, handle failures, and withdraw the permission.

In that final assignment, authority begins before the next email is sent. I have chosen the conditions under which sending may proceed. Successful delivery on one occasion does not establish that every future notice will be correct. Continuing permission creates a continuing need for oversight.

Three questions before handing over more work

What have I authorized this system to do? Name the information it may use, the things it may change, and the people it may contact.

What requires my approval before it happens? Make the point of approval explicit. Approval of a draft and approval to release it are separate decisions.

How will I know it worked, and how can I stop it? Look for evidence of the result and a practical way to pause future actions. A claim that work is complete needs checking; some completed actions cannot be fully undone.

These questions are a starting point for the fuller Accountability Test later in the book. When the consequences exceed your expertise, competent help is part of the arrangement. The person operating the tool may share accountability with the organization and the people who designed and authorized its use.

As the technology changes, return to the assignment: is the AI explaining, preparing, acting, or continuing under standing instructions? More capability does not automatically mean more permission.

The Argument in Brief

This book makes one argument in many settings. The chapters that follow refer to these principles by number; Part Six develops the Doctrine and Accountability Test in full.

The governing rule. Delegate the work. Never delegate the accountability.

The central distinction. Delegation is not the danger. Abdication is.

Delegation assigns work while the delegator keeps the goal, boundary, success standard, and consequences. Abdication leaves no identifiable party meaningfully answerable. Accountability asks who must answer. It may be distributed, but it must remain traceable. Blame is a later judgment about fault.

The eight principles

Principle 1. Accountability does not transfer. Work and discretion can move. Accountability can be shared, but it cannot disappear. “The system decided” does not close the chain.

Principle 2. Capability is not permission. Competence can justify more work, but not authority. As machines perform more execution, the remaining human decisions matter more.

Principle 3. Every delegation has a boundary. State what the delegate may do, must escalate, and may never do.

Principle 4. You did not delegate what you cannot stop. The owner needs enough evidence, time, authority, and competence to judge and intervene. Practical control is the test.

Principle 5. Speed and architecture move accountability upstream. They do not erase it. Accountability moves to the earlier choices about scope, design, safeguards, and stopping rules.

Principle 6. Consequences must reach the owner. Meaningful review and consequences must reach the accountable role.

Principle 7. Some decisions must carry a human signature. Human involvement is not human accountability. The accountable human must have authority over the decision itself. Irreversible, high-stakes decisions require a named, competent human with case-specific authority and real power to disagree or override. Decisions whose legitimacy depends on human judgment require independent judgment, not mere ratification.

Principle 8. Accountability decays. It must be renewed. Habit, overreliance, deskilling, and lost intervention capacity can turn delegation into abdication. Boundaries, review, override, and human competence must be renewed.

The Doctrine is not a defense of human judgment against machine competence. It protects the accountability loop that allows competence to be corrected, governed, and legitimately used. The full Doctrine and Accountability Test appear in Part Six.