Think, AI, think

Write your own position before you open the tool, use the tool, then decide for yourself what to keep. Almost everybody does the middle part. Hardly anybody does the first and the last, and the first is the one that gets skipped: a person who has not written a line of their own has nothing to compare the output against, so every answer arrives sounding correct. The whole of the rest of this material sits underneath those three steps.

The three steps#

Think first. Write your own position, even a single line. Work out what you are actually asking. Decide what has to stay yours before anything is delegated.

Then use it. Give it context. Ask it to question you rather than agree with you. Make it argue back. Ground it in your own material.

Then think again. Check what came back. Rewrite it in your own voice. Say it out loud. Ask whether you could defend it in a room with somebody who knows the subject.

Why the first step is the one that carries the weight#

Skipping the third step produces a bad answer, which someone eventually notices. Skipping the first produces a good answer you cannot evaluate, which nobody notices at all.

The mechanism is measured. Searching the internet inflates people's estimates of their own unaided knowledge, and does so even on questions the search never touched: access to an external source is experienced as personal knowledge. That is the illusion of competence, and a written position taken beforehand is the only cheap defence against it, because it is a record of what you actually knew.

There is a second reason, concerning the quality of the answer rather than the person. A model asked an open question returns the middle of its distribution. A model asked to attack a stated position has something to push against. The first step improves the middle step, which is the argument for it even if you care nothing about your own capability.

What the research calls this, and where the versions differ#

The closest established vocabulary is Dell'Acqua and colleagues' distinction between two working patterns observed among 758 BCG consultants. Centaurs divide the work, giving whole tasks to the model and keeping others. Cyborgs interleave, moving between themselves and the model within a single task. Both outperformed unassisted controls inside the model's competence, and both did worse outside it.

That is a description of what people do. This is a prescription about ordering, and the two are compatible: think, AI, think is a centaur boundary drawn in time rather than by task.

On the teaching side, Mollick and Mollick's seven approaches set out ways to use AI in a classroom, several of which put the student's own attempt first: the AI as tutor requires the learner to explain, the AI as student requires them to teach. The ordering principle is the same one.

Where this differs from both is that it makes the first step non-negotiable and unconditional. Neither Dell'Acqua nor Mollick claims that, and neither has tested it.

The failure mode has a name#

Going straight to the middle step is drift: the tool arrives, the habit forms around it, and nobody makes a decision. The three steps are what design looks like at the level of a single task, which is where the decision is actually available to be made.

The tell is easy to check on yourself. If you cannot say what you thought before you asked, you did not do the first step.

What this sequence has not been shown to do#

No study has tested think, AI, think. Nobody has compared people who write a position first against people who do not, on output quality, on retention or on error detection. It is a sequence built from evidence about the illusion of competence and about withdrawal effects, and the sequence itself is untested.

It also has an obvious cost. Writing a position first is slower, and for a task where you have no stake in the capability it is waste. The three steps are for work you need to be able to do again.

Key sources

• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier. Harvard Business School and BCG working paper.
• Mollick, E. and Mollick, L. (2023). Assigning AI: Seven Approaches for Students, with Prompts.
• Bjork, E. L. and Bjork, R. A. (2011). Making Things Hard on Yourself, But in a Good Way.

Related SuperSkills research#

The frameworks underneath this one: goal, context, friction, standard for the middle step, the source rule for the evidence, and keep, share, hand over for what to delegate at all. On the pattern it belongs to, drift versus design. On the failure it defends against, the illusion of competence.

About this framework#

Think, AI, think is used by Rahim Hirji in teaching and in the Mastering AI deck, most recently in September 2026. No claim of first use is made. No dated first publication exists for it and the Box of Amazing archive carries none, so it anchors to SuperSkills (Kogan Page, 2026) and to the deck rather than to a moment of coining. The centaur and cyborg patterns belong to Dell'Acqua and colleagues, and the seven approaches to Ethan and Lilach Mollick. The reading offered here, that the first step is the one that gets skipped and the one that carries the weight, is an interpretation by Rahim Hirji and is marked as an interpretation and not a finding.