Lesson 0.4.60.4.6 Delegate the work, keep the judgment

In everyday life. People who always follow GPS directions stop learning the roads. This is not a problem until the signal fails, in a part of town they have driven through a hundred times.

What you will learn. The three stages that lead from using AI to no longer judging it, what the research shows so far, and the habits that keep your judgment strong.

Why it matters. The skills in this chapter need practice. Early research shows what happens when people let AI do the thinking.

The three stages

How do people stop judging?

Addy Osmani is an engineer on Anthropic's Claude Code team.¹ In November 2026 he wrote about how people slowly stop judging the work of AI agents. In the past, people did their judging in the time between asking a question and getting the answer. Agents reduce that time to seconds. So a well-written answer is approved before anyone has thought about it. He described three stages. He calls the first one cognitive offloading. This book calls it delegation.²

The three words are easy to confuse. To delegate is to give someone a task. To defer is to accept someone else's view in place of your own. To surrender is to give up. A short way to remember them:

• Delegation: I check.
• Deferral: I could check, but I do not.
• Surrender: I cannot check.

Figure 0.4.8. From delegation to surrender. Delegation is the goal. Surrender is the failure.

The name for the last stage, cognitive surrender, comes from two researchers at the Wharton School, Steven Shaw and Gideon Nave. They define it as accepting an AI's judgment as your own. Osmani used their term for agents.³

Which stage was Dave in? He was about to forward the memo without checking it. But Dave is a controller. If Ana had asked him, he could have named the checks: the vendor terms, the September invoices, the cost of borrowing. He could check, and he was not going to. That is deferral. If it becomes a habit, the third stage follows.

What does the research show so far?

Two studies from 2026 measured what happens in the short term.

Shaw and Nave ran three experiments with about 1,400 people in total. Each person solved logic puzzles, and some had a chatbot to help. The researchers had secretly set the chatbot to give correct answers on some puzzles and confident wrong answers on others. These are the results of the first experiment, with 440 people:³

Figure 0.4.9. People with a wrong chatbot did worse than people who worked alone. Source: Shaw and Nave, 2026.

With a wrong chatbot, people did worse than people who worked alone. When people opened the chat and its advice was wrong, they followed it about 80% of the time. And people with the chatbot felt more confident, even when it was wrong.

One result is encouraging. In the third experiment, some people got a small reward for correct answers and saw at once whether they were right. They rejected wrong advice about twice as often: 42% of the time, against 20%. Feedback helps.

Anthropic tested 52 engineers, most of them early in their careers. All were learning a coding tool that was new to them. One group used AI, and the other group wrote the code themselves. Afterward, everyone took a quiz on the tool. The AI group scored 50% on average. The other group scored 67%. Inside the AI group, people who used AI to ask questions and get explanations scored much higher than people who let AI do the whole task.⁴

Be careful about what these studies show. The first measured decisions made while a chatbot was present. The second measured learning immediately after one task, with 52 people. Neither study followed people for years. So they do not prove that judgment becomes permanently weaker. They show how the risk begins: people accept wrong answers, and people who let AI do the whole task learn less. The long-term loss of skill is a risk that Osmani and the researchers warn about. It is not yet a measured fact.

The second study matters most for learners. A person who has not built a skill yet cannot use that skill when the AI is wrong.

Keeping your judgment

What should you give to AI, and what should you keep?

Give AI the small choices, where a mistake costs little, such as choosing between three similar hotels. Keep the thinking yourself where a mistake is expensive, such as a payment policy.² This is the rule from 0.4.5, used on your own attention. You can give AI the work, and much of the checking. You cannot give it the responsibility. In Osmani's words, "a human has to take responsibility."²

How should you read an AI's report?

First look at what the AI actually did and produced: the file, the number, the test result, or the log, which is the record of what happened. Read its summary after that. Summaries often sound more successful than the evidence. Also decide what "checked" means before the work starts, not after you see the result.²

Which habits keep the skill strong?

• Think first, then ask. Write your own short answer before you ask the AI. You learn from the difference between your answer and the AI's answer.
• Estimate before you calculate. Say "about 40 invoices" before you count 43. If your estimate and the result are far apart, one of them is wrong.
• Explain it back. If you cannot explain something in your own words, you have not learned it yet.
• Notice your own reaction. Before you accept or reject an answer, ask: What influenced me? What am I assuming? What else could be true? Thinking about your own thinking is called metacognition. The shortest form is one question: am I thinking with the AI, or is the AI thinking for me?
• Talk it through with people. Discuss a decision with a colleague before you make it. A person who knows your situation notices things that an AI cannot.
• Choose which skills to give up. List the skills you no longer use. Decide deliberately which ones the AI can keep doing for you.

For teams. People practice what is measured. If a team counts only speed and volume, the checkpoint is the first thing people skip. So count the checking as part of the work: what was checked, and what it caught.

Using AI less is not the answer. Use it fully, and keep the judgment.

Key words

• Delegation: the AI does the work, and you check the conclusion.
• Deferral: you accept the AI's output without doubting it.
• Surrender: the AI's answer becomes yours, and you can no longer say what you would check.
• Cognitive surrender: the researchers' name for the third stage.
• Metacognition: thinking about your own thinking.