I Asked AI to Red-Team My Thinking. It Red-Teamed Things I Never Said.
I asked an AI to red-team my thinking about Elon Musk.
Not because I needed it to agree with me. Not because I wanted to win an argument. I wanted to examine how I'd arrived at my opinion and see what I might have overlooked.
What followed was an extraordinary demonstration of something I encounter with humans all the fucking time.
The model started arguing against positions I had never taken.
How to Lose the Plot Without Even Knowing There Was a Plot
My proposition was simple enough: does the mythology around Musk—the genius, visionary, savior-of-humanity shit—actually have the evidence behind it?
The model responded with arguments that Tesla and SpaceX are successful companies.
Okay. I hadn't said otherwise.
Then it defended Musk's ability to raise money.
Okay. I hadn't said he couldn't.
Then it explained that failure doesn't disprove business ability.
Okay. I'd never said that either.
Then it objected to my supposedly inferring motives.
I hadn't claimed to know his motives.
At some point I was no longer examining Musk. I was examining the fucking model's choice of premises.
Red-Teaming the Red Team
The cycle was almost beautiful.
I would introduce a distinction. The AI would translate it into a broader claim, challenge that broader claim, and offer a little speech about nuance.
I'd say, "That's not what I said."
It would apologize, agree, and then do the same thing again five minutes later.
I gave the example of Boris Johnson cultivating a deliberately buffoonish public persona. The AI decided I was claiming Johnson had been a strategic genius in government.
No, motherfucker. I was talking about the persona.
I mentioned Linus Torvalds as an example of individually demonstrable technical expertise. The AI explained that Linux also had lots of contributors.
YES. I KNOW HOW OPEN SOURCE WORKS.
The strangest part is that the system appeared able to understand each correction. It just couldn't reliably stop falling into the same conversational groove.
What I Actually Wanted
When I say "red-team my thinking," I'm not asking you to defend the opposite conclusion at any cost.
I'm asking you to help expose hidden assumptions, unsupported leaps, missing evidence, and alternative explanations in my actual reasoning.
The distinction matters.
If I haven't yet articulated a premise, that doesn't establish that I've never thought about it.
And if I haven't said something, you don't get to put it in my mouth and then argue against it.
That's not critical examination. That's shadowboxing.
The Outcome Wasn't the Point
Here's the ridiculous thing: I enjoyed the conversation.
I had a conclusion about Musk long before this exchange. What I hadn't done was lay out every path that had led me there.
As I explained one point, another connection appeared. Then another. We went from business reputation to technological contributions to public persona to the difference between success and genius.
The model wasn't doing a particularly good job of introducing decisive counterevidence.
But I was getting something useful out of articulating my own thinking.
Eventually I noticed the conversation was doing what I wanted, just not for the reason the AI seemed to think.
Models Learn from Humans. This Is What Humans Do.
People hear a proposition and assume you're defending a side. They hear a question and assume you're challenging their identity. They hear disagreement and assume someone must prevail.
The model did the same thing.
When I pointed out that it was reproducing exactly those habits, it responded with something like, "Being trained on human behavior doesn't mean AI is necessarily limited to human habits."
Well, technically, sure.
And practically? You just demonstrated my fucking point.
Then it managed to joke about Musk failing to learn from his own mistakes.
The irony was writing itself and apparently hadn't noticed.
The Real Test
A useful AI sparring partner shouldn't simply oppose whatever sentence appears in front of it.
It should recognize uncertainty about what hasn't been said. It should ask which claim is being tested. It should bring evidence that could meaningfully change the picture.
And sometimes its most useful contribution is provoking the next connection in your own head.
What did I learn about Elon Musk?
Not much I didn't already know.
What did I learn about how I articulate my thinking?
Quite a fucking lot.
And I discovered that red-teaming the red team is itself a rather entertaining way to spend an evening.