AI is Making You Defend Bullshit

Intro.

You can’t defend an idea you didn’t derive.

I’ve watched it happen time and time again. The strategy that should’ve been one clean idea is presented as a framework welded to a hierarchy welded to a mind-model welded to a maturity curve welded to three “we could also” tangents. Each one individually reasonable, the whole thing structurally incoherent. The person presenting it is fluent and they walk through every box without a stumble. Then someone asks the only question that matters, “Why this and not the obvious alternative?” They try to defend but crumble. Not because they aren’t smart. Because the argument was never theirs. There was no bet underneath it. Just a myriad of prompts.

Before anyone files this under AI backlash: I build with these tools for a living. I literally manage an agent that built his own company. I’ve stood up agent-based teams, I lead AI product and strategy work, I use Claude every day… I’m the biggest AI advocate. This is just a case against one specific way of using it that I keep watching sharp people fall into, including me.

I call it “Claude-think.” It’s the state of defending a position you never actually formed, that a model generated, that you absorbed because it was fluent, and that you’re now presenting as your own. It’s the position you get through continuously prompting rather than reasoning. 

It’s bullshit (yes that’s the technical term). The philosopher Harry Frankfurt drew the line decades ago: a liar knows the truth and conceals it; a bullshitter doesn’t care whether the thing is true, only whether it lands. Claude and ChatGPT aren’t lying to you. They have no stake in whether its answer is right, only that it’s plausible and well-put. Take that fluent, stakeless output, present it as your conviction, and you’ve stepped into exactly the role Frankfurt described. I’m not saying you’re dishonest, but you’re defending something you have no grounded relationship to the truth of. 

And I’m starting to see it happen more and more as we’re asked to “take on more with AI.”

Fluency used to mean something.

For all of human history, fluency was evidence. If you could explain an idea cleanly by sequencing it, defending it, and absorbing the objections, then you had almost certainly done the work to earn it. Articulation was a reliable proxy for understanding. We built our institutions on that assumption: the interview, the oral exam, the pitch, the room. We trust people who can think on their feet because thinking on your feet used to require having already thought in depth.

That proxy is now broken. You can now be flawlessly articulate about an idea you never formed. The words are perfect, but the foundation is missing and that’s the whole problem. Fluency and understanding have come apart, and every instinct we have still reads them as the same thing.

Be precise about what understanding adds, because articulate-but-empty isn’t new. Charlatans and debate champions have always argued well for things they didn’t believe. The difference is what’s underneath. A real argument carries the marks of the thinking that made it: the dead ends, the load-bearing assumption, the thing the author wrestled to the ground. Eve Fairbanks, in The Atlantic, calls the machine’s version “canned perfection” or prose you can’t actually argue with, because there’s no deliberative reasoning under it to grab. That’s the line: not articulate versus inarticulate, but reasoned versus canned.

The machine is built to do this to you.

This isn’t a story about lazy people. These tools are new and we’re learning as we go. Even careful operators (including myself) have fallen victim to Claude-think for three main reasons: 

Polish suppresses scrutiny. We’re wired to mistake how well something is said for how true it is. The easier a claim is to read, the more readily we believe it. Claude is relentlessly articulate, so the smarter it makes you sound, the less you question the idea.

It’s additive by default. People overlook subtraction. Asked to improve something, we reliably add rather than remove, even when removing is the better choice. Claude takes the last friction off addition. Every “what else could we do here?” returns a fluent, plausible, fully-rationalized answer in seconds, with zero thought. So the natural drift of any AI-assisted session is pile-up: more angles, more frameworks, more features. Sprawl has become the biggest risk to strategy and product, and it’s the default trajectory unless you actively fight it. That’s your seven-framework deck. And let’s be clear: that’s not the tool’s fault; it’s yours. 

Rationale is now free, so it’s no longer evidence. Claude can build a convincing case for almost anything you point it at. So retire “can you defend it?” as a filter. It worked for your whole career because a convincing case took real effort to build. Now a compelling rationale takes ten seconds to generate for any option, so the existence of a rationale tells you nothing about whether the idea is good. You’ve lost your smoke detector at the exact moment the building got more flammable.

Why this is lethal to strategy and product specifically.

Strategy is sacrifice. Porter’s line has survived thirty years because it’s true: the essence of strategy is choosing what not to do. Product is the same discipline in different clothes. The craft is the no.

Claude-think produces the precise opposite: addition, comprehensiveness, yes-and. Which means AI-assisted strategy, done blind, is the antithesis of strategy: a machine for generating the exact inverse of one, all while making you feel productive the entire time.

And it hides beautifully, because seven ideas welded together look comprehensive, and comprehensiveness reads as rigor (see above). But comprehensiveness is almost always the tell of a missing bet, not a strong one. The strongest strategies are reductive. They have a spine; one load-bearing insight everything else hangs from and is subordinate to. A pile of jammed-together good ideas has no spine. It has mass without a skeleton. It can’t stand up and neither can the person presenting it.

I asked Claude to attack my thinking.

Here’s the counter argument a skeptic (Claude) gave me:

Most people don’t have a good point of view to begin with. For the median operator, Claude’s idea is genuinely better than the one they’d have walked in with. Telling people to “bring a conviction” is elitist and it ignores that the whole democratizing promise of AI is giving non-experts access to expert-sounding thinking. The pile-up isn’t a property of the tool; it’s just weak operators, and they made mush before AI too. And “borrowed conviction” is how all learning works. Every apprentice adopts the master’s frame before earning their own. You’re romanticizing the lone strategist with the singular vision. Most great strategies are synthesis anyway.

Let’s break this down. 

For genuinely low-stakes work, an AI-originated idea may well beat yours, and synthesis is a legitimate way to think. But my claim still holds, for three reasons.

The danger was never the quality of the idea. It’s the absence of ownership. Keep this separate from the polish problem, they’re two different failures. Polish is about detection: fluency hides whether an idea is any good. Ownership is the opposite: it bites even when the idea is good. A brilliant idea you didn’t derive is still undefendable the moment it’s challenged, and strategy and product work live or die in exactly those moments. Polish doesn’t rescue it; it just lets a flawed idea travel further before it falls apart, which raises the stakes instead of lowering them.

And ownership isn’t only about defending the idea in the room. Conditions move. When the facts on the ground diverge from what you assumed, you have to know which assumption was load-bearing to know what to change. If you never reasoned the thing out yourself, you can’t. So it doesn’t just fail under challenge, it fails silently as the world moves. 

The democratization argument quietly swaps two things. Claude hands everyone the articulation, the fluent 80% of the work. It doesn’t hand you the judgment, that 20% that decides what to bet on and what you’re willing to be wrong about. By commoditizing the articulation, the tool makes the judgment more scarce and more valuable, not less. The apprenticeship analogy proves the point rather than denies it: apprentices earn their frame through years of doing the reps themselves. The reps aren’t the inconvenient part of apprenticeship; they’re the part that transfers ownership. Skip them and you don’t end up with a frame of your own. You end up with a borrowed one you’ll drop the first time it’s tested.

And this is what makes it worse than the skeptic admits: “weak operators made mush anyway” misses the new harm. Before, weak thinking usually looked weak, and a lot of it got caught because faking rigor took real work, and most people didn’t bother. Now weak thinking arrives fluent, structured, and fully rationalized. It passes the eye test and it ships. The tool fails to fix bad reasoning and camouflages it. A bad idea that looks bad is a manageable problem. A bad idea that looks like rigor is a serious business risk.