Why AI Should Support Thinking, Not Replace It

Most arguments about AI and research collapse into two camps. One says the tools are a shortcut to work nobody should be doing by hand anyway. The other says they are a threat to scholarship itself. Both are arguing about the wrong thing, because both treat "using AI" as a single act with a single answer.

It is not one act. There are two, and they are not remotely equivalent.

The first is offloading retrieval. Finding the paper, pulling the reference into the right format, remembering which of four similar studies used the larger sample. This is work, it is necessary, and almost none of it is thinking. A researcher who automates it has lost nothing. They have bought back the hours that the actual difficulty deserves.

The second is offloading judgement. Deciding whether this evidence supports that claim. Noticing that two sources you trust disagree, and working out which one is wrong and why. Choosing what to conclude when the data will not quite carry the conclusion you hoped for. Judgement is not the overhead around research. It is the research.

The reason this distinction matters more with language models than with any previous tool is that they blur it. A search engine returns results and leaves you to think. A model returns a conclusion, phrased with the same confidence whether it is right or invented. Fluency is not accuracy, and nothing in the output signals which one you are looking at. That is a genuinely new problem, and it is not solved by using the tool less. It is solved by knowing which of the two acts you are performing.

The test that actually works

There is a practical version of this, and it is older than any of the technology. Could you defend it?

Not "can you produce it". Defend it, in a viva, to a supervisor who has read the same literature and is unconvinced. If a machine wrote the sentence and you could not reconstruct why it is true, you do not have a finding. You have a plausible sentence. The distinction is invisible in the document and decisive under questioning.

This is the honest reason to keep judgement with the human, and it is worth being clear that it is not a romantic one. It is not that human thought is sacred. It is that you are the one who has to answer for the claim. Accountability cannot be delegated to a system that does not experience consequences, so the reasoning had better sit with whoever does.

What this looks like in practice

Working this way is less about restraint than about sequence.

Ask the question yourself. A question you did not form is a question you cannot judge the answer to, and the framing carries most of the assumptions.

Let a tool find things, then read them. The retrieval was automated; the appraisal was not. If a source was never opened, it was never evidence.

Look for the disagreement rather than the summary. A tidy synthesis of a contested literature is not a summary, it is a decision someone made on your behalf, and the interesting part is exactly what got smoothed away.

Say what you do not know. The single most useful thing a researcher can write is that the evidence is mixed. It is also the sentence a system optimised for helpfulness is least likely to produce unprompted.

The concession

The strongest objection to all of this deserves stating plainly: the boundary is not clean. Summarising a paper is retrieval right up until the summary decides what mattered, and then it is judgement wearing retrieval's clothes. Extraction is mechanical until the categories you extract into embed a theory. Anyone who claims a bright line is overselling it.

But an imperfect distinction is not a useless one. "Am I outsourcing the finding, or the deciding?" is a question you can actually ask yourself mid-task, and it is right far more often than it is wrong. Most bad uses of these tools are not subtle edge cases. They are someone accepting a conclusion they could not have reached and could not defend.

That is the line Bizarus is built on. Not because automation is suspect, but because the part that makes research worth trusting is the part a machine cannot be accountable for.