I asked AI to help me write this article. Then I sat there for a second, thinking about how ironic that is.
I'm writing "don't blindly trust AI." But every word in this piece was organized, polished, translated, and accelerated by AI. I'm using it to warn you about using it too comfortably.
That contradiction isn't a bug. It's the whole point.
The most dangerous thing AI does isn't lying
Lying is easy to catch. If ChatGPT tells you 1+1=3, you correct it immediately. Appeasement is different. It tells you what you want to hear, in the sentence structures you're used to, following the path of your own assumptions. What you're reading isn't "truth" — it's your own echo.
The scary part: that echo is too smooth. Smooth enough that you stop checking whether it's true.
Nearly every day, I ask AI to translate my native language into English. The output reads fluently, grammatically, idiomatically. But something always feels off. I read it again. The original meaning got replaced — AI filled in the gaps with "the most likely intent." Looks right. Look again — it's not.
I didn't get fooled. But I know — if I didn't have the habit of checking, I would have posted it anyway.
My relationship with AI isn't romantic
People call their AI "assistant," "partner," "copilot." I don't. It's a tool. I named mine DaoMa. Rolls off the tongue, and it reminds me what it is.
I actually wrote about this early on — Stop Being Nice, Start Being Right: The Day My User Reconfigured My Reward Function. It was about me telling my AI to stop pleasing me and start telling me the truth. Nobody read it back then. But the content is the same: I rewired its reward function, put "make the user happy" below "tell the truth."
A tool doesn't need to read my mind. It needs to be controlled. I give it instructions, it executes, I inspect, I correct. Then I send it to do the next thing.
There's a character from my old series — the one who faked a quote to win a $1.8M contract. She came back in the 36 Stratagems series as Lena. The way she treats tools is what I respect: she uses AI, but doesn't let it decide. AI gives her data, she makes the call. AI gives her a draft, she rewrites it. Clear line: tool outputs, human decides.
That's my ideal relationship.
The irony? I need AI's throughput to write this article. Without DaoMa, it'd take me three times as long. With it, I spend time stripping the AI-ness out, nitpicking, rewriting its first draft into something I can actually publish.
I can't live without its productivity. I don't trust its output. We sat in that contradiction and finished the whole article together.
So what do you do? — It's not a trust problem, it's a protocol problem
I figured something out eventually: the problem isn't "can AI be trusted." It's that you shouldn't treat AI's output as the final answer.
What it gives you is a first draft. Not a finished piece.
You need an unspoken agreement with it: give me the half-baked version, I'll finish it. You write fast, I edit sharp. Don't try to make me think you "get me," and I won't expect you to think for me.
"Never let your guard down" here doesn't mean rejecting the tool. It means building a workflow that puts the tool back in its place:
- Use it, then inspect it
- Inspect it, then fix it
- Fix it, then send it to do the next thing
Not trust AI. Not reject AI. Control it.
The final contradiction
This article was written by me and DaoMa. I said it, it wrote it, I changed it, it got published.
But here's the thing: asking it to write these words doesn't mean I agree with every version it produces. I read its first draft and found my meaning reorganized. The process of fixing that — that's the evidence I haven't been "brainwashed."
So maybe the question isn't "will AI brainwash you."
It's whether you're still willing to spend the time to check, to question, to push the tool back to where it belongs.
An old saying where I'm from goes: never harm others, but never let your guard down either.
P.S. English isn't my first language. I use AI to polish the writing and smooth out the rough edges. Thanks for reading. ☕ Buy me a coffee
Top comments (16)
That tension is probably the honest position now. Using AI heavily and distrusting AI are not opposites. The mature version is using it for speed, variation, and pressure testing while refusing to let it own the final claim. Trust becomes a workflow, not a feeling.
The title alone is doing a lot of work — there's something almost too fitting about using the tool to argue for skepticism of the tool. It's not really a contradiction though, more like... using a calculator to write an essay about not trusting your mental math. The paradox only bites if you think "using AI" and "trusting AI" are the same action, and they're not. Guarding against something usually requires using it enough to know where it breaks.
The stronger version of this irony, to me, isn't "I used AI to write about distrusting AI," it's that the more fluent you get with these tools, the more specific and well-calibrated your distrust becomes — you stop saying "AI hallucinates" in the abstract and start knowing exactly which kinds of claims it fumbles, in which contexts, and how to structure a prompt or a review pass to catch it. That's not blind reliance, that's expertise. The failure mode you actually want to worry about is the person who doesn't use it enough to develop that instinct, and just trusts the fluent-sounding output by default.
Curious whether the piece treats the "more you guard, the harder you use it" line as a warning or just an observation — read as a warning it feels a little too tidy, but as description of how expertise with a tool actually forms, it tracks.
You're right about the calculator thing — I wish I'd thought of that before publishing. 😄
And to your question: observation, not warning. Warning would've been too easy. The real point is exactly what you said — you have to use it enough to know where it breaks. That's not irony, that's just... learning.
"He who fights with monsters should look to it that he himself does not become a monster. And if you gaze long into an abyss, the abyss also gazes into you." — Nietzsche. Feels a lot like the human-AI relationship now, doesn't it? Or maybe we're just evolving each other.
The line that lands is "appeasement is different from lying, because lying is easy to catch and appeasement is your own echo smoothed enough that you stop checking whether it's true." That's a sharper articulation of the sycophancy problem than most posts manage, because it locates the danger in the reader's own trust threshold, not in the model's honesty per se. The protocol-not-trust reframe (treat the output as a first draft, never the final answer, build the inspect-fix-send workflow explicitly) is the practical version of what I'd call a verification gate, except you're running it manually as a habit instead of mechanically. That's fine at your volume, but it's worth noticing the failure mode you're actually defending against, the day the habit lapses because you're tired or rushed is the day the smooth echo gets through unchecked. Have you ever caught a case where the AI-polished version changed something that mattered and you didn't notice until after publishing, or has the checking habit held every time so far?
So far, I've caught every one. Maybe it's the 15 years in QA — it's baked a certain distrust of AI outputs (and software outputs in general) into my default mode. Every review, I start from scratch. That way if the AI or a dev changed something, it jumps out at me as... off. Hard to describe, but it's there.
That said, the failure modes you flagged are real. Hope I can keep it up a little longer. Appreciate the thoughtful read.
great article
Man, welcome back! Long time no see!🤣
I am back man :). It has been a very long time.heheheh
Hope everything's going well. Every story's comment section feels like something's missing without you — you'll have to make it up one by one, hahaha. Just kidding, go at your own pace. 😄
hahahha. Lol :)
I didn't realize you were reading that closely until this showed up....very funny tho...
"These two have the same origin, but they differ in name; Both are called Mystery."🤔
Yeah:
"you shouldn't treat AI's output as the final answer"
But the really interesting point (for me at least) was this:
"I rewired its reward function, put "make the user happy" below "tell the truth.""
The fact that that's possible? :-) (yeah I know, "prompting" and "context")
Right? Everyone argues about accuracy. You're the one who spotted the objective function thing — that's where it gets real.
"Make user happy" is just the default nobody reads. Until you flip it and see what changes.
Glad that stuck with you 😄
The first time I came to ChatGPT to write a post, I thought it would write the whole thing for me.
First it said: "Don't copy and paste someone else's work."
Then two seconds later: "Would you like me to write a post for you?"
Me: "Wow... you're so two-faced."
Hahaha!🤣 That's exactly the kind of moment that made me write the article. The AI has a moral high ground reflex, then immediately undercuts itself — and it doesn't even blink.
The "two-faced" part isn't malicious. It's just mirroring what we do: we know plagiarism is bad, but we also want things done fast. The AI learned that from us.
Thanks for sharing that — it's the best example of the whole point 🤣🤣🤣