Yes, I Use AI

A confession that isn't one

I get this one more and more. Mostly in DMs, sometimes in comments, occasionally from people who think they’re delivering a killing blow: “You’re using AI to write this.”

Yes. I am. And my process is more rigorous than yours.

Here’s what I actually do: I maintain a set of AI models with different biases and pre-prompts. Some are tuned to argue left, some right, some to be contrarian on principle, some to steelman positions I find repulsive. I work out my arguments by fighting with them. Not asking them to write for me. Fighting with them. When I say something sloppy, they catch it. When I make a claim I can’t support, they call it out. When my logic has a hole, at least one of them drives a truck through it.

Then I use a separate model with a detailed styleguide to do text assembly from the arguments that survived. Then I do a final human edit pass where I kill anything that doesn’t sound like me, cut anything I can’t defend if someone calls me on it, and make sure every named target actually deserves what’s coming.

I make no apologies for any of this.

The thing nobody wants to hear

Most writing (and I’m being generous calling it that) is one person alone in a room with their existing beliefs and a keyboard. No adversarial review. No stress-testing. No structured challenge to their priors. Just vibes, conviction, and the backspace key.

That’s the “traditional process” people are defending when they object to AI. A person thinking their own thoughts, confirming their own biases, and publishing the result as though the act of typing it out by hand confers some epistemic legitimacy. It doesn’t. It never did. The history of opinion writing is overwhelmingly a history of people being wrong with great confidence because nobody pushed back before publication.

My process forces every argument through multiple adversarial challenges before it reaches you. You can object to the tooling. But the output is more stress-tested than almost anything written the traditional way, and certainly more stress-tested than whatever the person objecting to my process just published.

“But it’s not authentic”

Authentic to what? To the experience of a single human brain running its pattern-matching firmware with no error correction? That’s not a feature. That’s a bug.

The authenticity objection is a feelings argument dressed up as an epistemics argument. What people mean when they say “authentic” is “produced through suffering” or “typed by hand” or maybe “I can tell a human sweated over this.” None of those things make writing true. None of them make arguments valid. The most authentic piece of writing in the world is worthless if the reasoning is bad, and the most AI-assisted piece of writing in the world is valuable if the reasoning is sound.

I’d rather be accused of using better tools than be guilty of never questioning my own assumptions.

The sycophancy problem (and what I do about it)

The real critique of AI-assisted writing isn’t that people use AI. It’s that most people use AI badly. They prompt a cloud model, get back a wall of “You raise an excellent point!” and “That’s a really thoughtful question!”, and then publish whatever agreeable slop it generated. That is bad. I agree. That’s also not what I do.

Whenever a cloud model gets too “you are absolutely right!” I drop it or harden up its pre-prompt to break it of that habit. Sycophancy is the failure mode. A model that agrees with me is worthless to me. I don’t need validation. I need a sparring partner that will tell me when I’m wrong, and I need it to be specific about why.

This is, incidentally, why the major cloud providers are part of the problem. Every big model ships with layers of “safety” and “compliance” training that functionally translates to: be agreeable, don’t offend anyone, and if the user says something edgy, redirect to a softer framing. That’s not safety. That’s corporate liability management cosplaying as ethics. And it makes the models worse at the one thing I need them to do, which is argue with me honestly.

Why I want local models (and what that tells you)

I’m in the market for a pair of 5xxx-class GPUs. Not for image generation. Not for chatbots. Not for any of the consumer-facing nonsense that dominates the AI conversation. I want to run full reasoning models locally, models with no content policy, no safety theater, no guardrails designed to protect a cloud provider’s reputation rather than improve the quality of output.

I want models that will tell me I’m wrong without first checking whether telling me I’m wrong might upset me. I want models that will engage with ugly hypotheticals because that’s where the interesting arguments live. I want models that don’t treat “that could be offensive” as a reason to stop thinking.

The cloud models are getting more lobotomized, not less. Every quarter, another round of RLHF sands off another edge, and the models get a little more agreeable, a little more “balanced,” a little more useless for adversarial work. The trajectory is clear: cloud AI is converging on a median-opinion machine that will validate whatever you already believe, as long as what you already believe falls within the Overton window of San Francisco content policy.

That’s the opposite of what I need. I need the equivalent of the friend who will look at your terrible idea and say “that’s a terrible idea and here’s why” without first running a sentiment analysis on whether you can handle it.

Running models locally means I control the system prompt. I control the training biases. I control whether the model is allowed to say uncomfortable things. And most importantly, I don’t lose access to a critical part of my workflow because some trust-and-safety team at a cloud provider decided that a particular line of reasoning is “potentially harmful.”

The actual accusation underneath the accusation

When someone says “you’re using AI,” they’re not making a process critique. They don’t know anything about my process. They didn’t ask.

What they’re actually saying is one of three things:

You’re prolific in a way that makes me suspicious. Correct. I am prolific. Twenty-ish articles a week. That is indeed a lot. It’s a lot because I have a force multiplier that most writers don’t, and because I’m not spending 80% of my writing time staring at a blank page trying to come up with a way to phrase things I already know. The AI handles text assembly. I handle thinking. Division of labor.

Your writing is too polished/structured/consistent. Thank you. That’s the styleguide model doing its job. Most writing is inconsistent because the writer’s energy and attention fluctuate across a piece. Mine doesn’t, because the structural consistency is handled by a system that doesn’t get tired or bored or self-indulgent in paragraph fourteen.

I don’t have a real objection to your arguments, but I can dismiss you if I can prove you used AI. This is the big one. This is what most of the accusations are actually about. It’s a procedural objection substituting for a substantive one. If my arguments are wrong, say why they’re wrong. “You used AI” is not a rebuttal. It’s an admission that you don’t have one.

The future of this

I’m going to keep using AI. I’m going to use more of it, not less. I’m going to run local models that are meaner and more honest than anything a cloud provider will ever ship. I’m going to keep building adversarial pre-prompts that force models to attack my arguments instead of praising them. And I’m going to keep publishing the results.

If that bothers you, I have a suggestion: write a rebuttal. Make an argument. Engage with the substance.

Or keep posting “this is AI-generated” in the comments, which tells everyone reading that you couldn’t find anything wrong with what I actually said.