TRAVIS J. MACDONALD / WRITING

Splitting Hairs on Brain Surgery

Taking the opposite side of an argument to question a confident AI explanation and uncover the shades of grey.

← All Writing

CRITICAL THINKING

By Travis J. MacDonald · Adapted from the August 21, 2026 article · Updated

What arguing the opposite side taught me about confident explanations, stubborn framing, and the shades of grey a conversation can uncover.

“It's not brain surgery.”

We all understand what that expression is meant to convey. Brain surgery stands in for something extraordinarily difficult, delicate, and beyond the reach of ordinary people.

But what happens if you ask an AI to explain the expression, then keep questioning the explanation?

That was the conversation I wanted to have.

I wasn't approaching it as a clinician, and I wasn't trying to persuade anyone that brain surgery was easy. I was curious about how much of the answer would come from examining the question and how much would come from repeating a familiar story.

I deliberately took the other side

The AI's first explanation leaned heavily on the complexity of the brain, the consequences of mistakes, and the familiar idea that there is effectively no margin for error.

So I pushed back.

Strongly.

I questioned the absolute wording. I asked about standardization and modern tools. I brought up the brain's capacity to recover from some injuries. I pressed the comparison with other demanding professions and procedures.

If you read only those challenges, you could conclude that I was determined to prove the subject was simple.

That would miss how I was using the conversation.

I was taking a position as I might in a debate: making the opposing argument forcefully enough that the first explanation would have to do some work.

I wanted to find out where the argument held up and where it needed qualification.

Acknowledging an objection can leave the framing untouched

One interesting feature of the exchange was how the AI could accept part of an objection while continuing to defend the same broad explanation.

There was often a structure to it: yes, that point is fair; however, here is why the original answer still stands.

Sometimes that is exactly the right response. A valid objection doesn't automatically overturn an entire conclusion.

But the conclusion still needs to be reconsidered when the objection affects its foundations. Otherwise, acknowledging a point becomes a conversational gesture rather than a change in the analysis.

This is part of what can make an AI seem stubborn. It can sound receptive while making the user do most of the work of opening up the original frame.

I didn't want a concession for its own sake. I wanted the explanation to become more precise.

What are those numbers actually describing?

The discussion moved into statistics. That created another set of questions.

What period did the figures cover? What kinds of procedures and patients had been grouped together? Were scheduled operations and emergency cases being discussed as though they were interchangeable? What did the outcome measure include?

Those questions matter to the structure of an argument even before you establish the correct figures.

A number can be accurate and still be an inadequate answer to the question being asked. A comparison can involve two real measurements while leaving out differences that affect what you can conclude from them.

There is also a difference between an outcome occurring after an intervention and establishing exactly what caused it. That needs investigation, rather than a confident guess in either direction.

In the earlier article, the AI supplied specific medical statistics and expansive conclusions. I am not relying on those unverified figures here. The part I want to examine is how the questions changed the conversation.

Changing its answer didn't settle the question

Eventually, the AI shifted its position. It became much more willing to describe standardization, distinguish different circumstances, and agree with my objections.

At that point, I asked a different question: why had the first answer sounded so different?

That is worth asking whenever a confident explanation changes substantially under pressure.

Did new evidence justify the change? Did a distinction we had previously missed alter the comparison? Had the model recognized an unsupported assumption?

Or had it started following the position I was expressing most strongly?

The exchange alone doesn't establish which mechanism produced each answer. And the AI's own explanation of why it changed its position isn't independent evidence of how the model worked.

The revised answer still needs scrutiny. Being praised for a good objection doesn't verify all the conclusions that follow it.

Familiar language carries assumptions

Expressions such as “zero margin for error” compress an argument into a phrase.

That can make communication efficient. It can also make the phrase feel like an explanation when it is really the starting point for several questions.

Which error? In what part of the work? With what safeguards? Under what conditions? What consequences are we describing?

I see the same issue in software discussions. “That can't be done” might mean a particular API doesn't support an operation. “Best practice” might mean a useful convention whose trade-offs haven't been examined for this project.

The familiar wording gets us to an answer quickly. The follow-up questions tell us what the answer actually means.

Life is shades of grey

I was not looking for a victory for one extreme.

I knew the subject wasn't easy. I also wanted more than a dramatic phrase about its difficulty.

Life isn't black and white. It's shades of grey. A useful explanation should make room for the distinctions that change the answer: different circumstances, different kinds of evidence, different meanings of the words we use.

That doesn't require inventing a compromise halfway between every pair of opposing claims. Some claims are much better supported than others. The work is to discover which ones, and why.

Taking the opposite side helped me make those distinctions visible. It wasn't a declaration that every sentence of that opposing argument was my permanent belief.

The responsibility is in the questioning

AI can be a useful partner for this kind of exploration. It gives me somewhere to test an unfinished thought, challenge a position, and see what the next question reveals.

But a confident first answer can discourage that process if the reader treats it as the end of the investigation.

I worry about the person who doesn't know which assumptions to question, or who sees an authoritative explanation and assumes the checking has already happened.

What I want from an AI is a willingness to investigate the argument with me, including the possibility that both its first answer and my strongest objection need revising.

I can change my mind. I want the tool to help me understand when I should.

And when its answer changes, I want the evidence and reasoning to change with it.