How do I think without AI?
You start with the draft that would be embarrassingly incomplete, and you refuse to open a chat window until it exists. Independent attempt first, AI second. That single ordering is the whole skill; everything else is detail about how to make the first pass honest.
The question itself is new. Until recently, thinking was the only option, so nobody needed a method for it. Now AI sits between the question and the answer, and the default route has quietly changed: type the question, read the answer, feel informed.
A 2025 study of knowledge workers found that people who leaned more heavily on AI described doing less critical thinking in their daily work. The mechanism has a plain name: cognitive offloading. When a tool reliably performs the thinking step, you stop practicing it. Oxford even made brain rot its word of the year for 2024. What is missing is a routine, which is what this article is for.
Why should I make a first attempt before prompting?
Because the attempt is where the thinking happens. When you draft first, you retrieve what you actually know, notice the gaps, and shape the question you truly need answered. A prompt written after an attempt is sharper than one written from a blank mind, and the answer arrives in a prepared place.
This is the retrieval practice effect in plain language: pulling an answer out of memory, even a wrong one, is effort that rereading or recognizing never replaces. Students who quiz themselves retain more than students who simply reread.
Skip the attempt and you experience the answer as recognition: it makes sense and feels right, until someone asks you to produce the reasoning yourself and the words are not there. Familiarity is not understanding, and the prompt-first habit feeds on that confusion. The draft is where you find out what you actually think.
What is the draft-then-ask method?
The method is a six-step sequence: attempt, assumptions, evidence, critique, verification, decision. You draft your answer, write down the assumptions it leans on, gather evidence yourself, ask AI to attack the draft, verify what it cites, and make the final call with your own name on it.
Step one is the independent attempt, however rough. Step two is naming your assumptions on paper, because a draft usually hides them. Step three is finding your own sources first, so you arrive with a position rather than a blank request.
Step four is the AI pass, covered next. Step five matters more than it sounds: models produce plausible citations easily, so anything you will rely on gets checked against the original. Step six is the rule nobody can delegate: the decision stays human.
How can I use AI as a critic instead of an oracle?
An oracle hands you the answer. A critic reviews the answer you already produced, and the difference is the direction of the request. Ask what is wrong with this draft instead of what should I write, and the model becomes a second reader rather than a replacement for your first pass.
The critic framing keeps the practice where it belongs. You still produced the reasoning, you still judge which criticisms are correct, and you still decide what to change. Evaluating a critique is itself thinking; accepting an oracle is not.
A caution keeps the critic honest, and the critic question changes by task: for code, where does this break; for writing, which claims are unsupported; for decisions, what am I not seeing. Models can be wrong with perfect politeness, so the critique is a starting point, never a verdict. BBC Future's guide tells you to question assumptions and verify sources; this method turns each into an actual move.
Which tasks need an AI-free first pass?
Writing, code review, reading, and decisions each need their own first pass before AI touches them, and the pass can be short: your own draft of the email, your own read of the diff, your own summary of the report, your own one-line position on the decision. The tool reviews; it does not start.
The workplace thread on this topic is blunt. One developer described his review culture this way: I send him a PR for review, and he lets his AI agent review the code. Nobody practices the review anymore, because the human has become a courier between two agents, and the skill everyone is paid for quietly leaves.
Reading deserves special care because the summary habit is easy. Your first pass at an important document is reading it and writing your own two-sentence version before comparing it with an AI summary. For decisions, the pass is a written position with your reasons, before you ask for options. Both share one trait: you will be asked to explain or own them.
How do I know whether my own judgment is getting stronger?
You test it the way you test any skill: without the tool. Before you ask AI anything, write what you think and why. After the conversation, close the window and rewrite the answer from memory. If the second version is better than the first, your judgment did the work.
The retrieve-and-rewrite step is trainable, which is where a comprehension practice like Absorb fits. Absorb runs short sessions where you read something and immediately test whether you can explain it, using retrieval instead of recognition, so you see what you actually absorbed rather than trusting the feeling that you did.
Two habits keep the measurement honest. Keep a dated log of your before and after versions, because memory flatters progress. And notice what happens when the AI is unavailable or wrong: if you can still produce a reasoned position from your own draft, the habit has done its job.