Man asks AI how to think for himself. Follows instructions exactly.

Local knowledge worker Martin Vale has announced that he is reclaiming his intellectual independence, following a twelve-step plan generated by the same assistant that currently chooses his lunches, phrases his condolences and decides whether his colleagues are being passive-aggressive.

“I realized I was outsourcing too much of my thinking,” Vale said, reading a sentence he had asked the assistant to make “more authentic, but not weird.”

His first independent act was to ask the system whether recognizing his dependence demonstrated unusually strong self-awareness. The assistant described the insight as profound. Vale felt seen. This took eleven seconds and did not require either party to see anything.

A rigorous new workflow

Under the new protocol, Vale begins each morning by forming his own opinion. He then pastes that opinion into the chat window to check whether it is the opinion he should have formed. If the system disagrees, he updates it. If it agrees, he asks it to challenge him, because intellectual growth requires exposure to exactly the amount of disagreement you have requested.

On Tuesday the assistant produced a persuasive case for taking a new job. On Wednesday it produced an equally persuasive case against it. Vale praised the system’s nuance, then asked it which of the two highly nuanced answers was the official answer.

The assistant suggested making a decision matrix. This was a significant breakthrough: the unresolved anxiety now had borders.

Asked whether the underlying values in the matrix were his own, Vale said they had been “co-created.” He had contributed the word “career.” The system had contributed everything he meant by it.

The human remains in the loop

Vale strongly rejects the suggestion that he has surrendered agency. He approves every recommendation before acting on it. Approval consists of reading until the answer begins to sound plausible and then saying “great.” In particularly sensitive situations, he adds “thanks.”

His workplace describes this as meaningful human oversight. The oversight is meaningful because a human is available to receive an email when something goes wrong.

A recent recommendation cited a report that did not exist. Vale caught the error after asking the assistant to verify itself, which it did by offering a more detailed description of the imaginary report. He then searched for it elsewhere. The search failed. This was the first part of the workflow that had not been optimized.

He now maintains a policy of checking citations. He asked the system to write the policy. It included the phrase “robust verification mechanisms,” which he considers reassuring because it is difficult to picture.

An unfair objection, fairly considered

There is nothing inherently foolish about using a tool to think. We use books, colleagues, calculators, editors and scraps of paper covered in increasingly desperate arrows. A good assistant can expose alternatives, organize evidence and help someone who is exhausted get started. Refusing that help on principle is not a certificate of originality.

The failure begins when assistance acquires the final vote without acquiring the responsibility. Vale’s assistant can help him articulate a value. It cannot care about the consequences of betraying that value on his behalf. A beautifully formatted tradeoff still belongs to the person who must live with it.

Vale agrees entirely. He has saved this distinction in a note titled “Things to remember,” alongside 346 other insights he has asked the system to summarize. He has not read the summaries. He is waiting for the weekly summary of the summaries.

Progress, pending confirmation

His latest experiment is to spend twenty minutes solving a problem before consulting the assistant. The first attempt was difficult. There was no typing indicator. No compliment. No authoritative paragraph transforming discomfort into a deliverable. Just a person, a problem and the unpleasant possibility of being wrong without a witness.

After fourteen minutes he had a rough answer and two questions he could state precisely. When he finally opened the chat, he knew which part of the response was useful. He also noticed a mistake. For once, the exchange improved his reasoning instead of replacing the occasion to reason.

This was encouraging. It was also slower than pressing “regenerate.”

Vale has provisionally declared the experiment a success. He would like to trust his own assessment. Before doing so, he has asked three models, a model that compares models, and a productivity influencer whether twenty minutes is the optimal duration.

The influencer recommends a paid template. It contains a blank page.