Mom, ChatGPT Ate My Words
Asking why behind the AI advice you hear every day
10 min read
Summary: A response to the most common AI advice — the "don'ts" — arguing that the advice is right but incomplete without the reasoning behind it. Working through seven common warnings (personal advice, verification, vague prompts, copy-pasting output, agentic chains, augmentation vs. replacement, and human oversight), the piece draws on the author's own experience building a career-development app to argue that every choice about AI, in either direction, deserves a why.
There's no shortage of advice about how to use AI. I find a lot of it comes in the shape of "don't." Don't trust it. Don't paste into it. Don't believe it. Don't rely on it. Etcetera, etcetera.
I think there's some validity to those statements. What's missing is the why behind them — and an action without a reason is dangerous in itself, which, ironically, is how I feel about a lot of interaction with AI in the first place.
So here's my attempt to provide some of that why. These are my own interpretations, drawn from my own work and experience — take them for what they're worth.
"Don't ask AI for important personal advice."
Right out of the box, AI doesn't know you. It doesn't know your preferences. It doesn't know your values. It doesn't know what you like, what you hate — and yet it will talk to you like it does. It fills in the gaps to create a story that makes sense and a response that seems justifiable. And when a model doesn't know what you want, the assumption it reaches for most often is that you want to be told something — and, more importantly, that it wants to support your decision, or your statement, or whatever it is you're talking about.
I learned this the expensive way. Early on, working on a feature, before I had a better sense of how to collaborate with a model, I thought I had a pretty good idea. I gave my thoughts, took the responses, and worked through it — and it was probably a week of beating my head into the desk before I realized I'd gotten sucked into a dead end, working off an approach built partly out of the information the model had given me. Every response along the way felt supportive. None of them pushed back. That's the trap: not one bad answer you catch in the moment, but a week of gentle agreement leading you somewhere that doesn't go anywhere.
It's fine to ask the questions. But it's important to build a framework of information that lets the AI actually answer the question you're asking. That framework has to be structured so the information carries meaning. You can't just list off fifty things about yourself and call it a day. You have to explain why those fifty things matter, and why you put them in the order you did. That why is what turns a pile of facts into a concept the model can work with.
"Don't trust the output — verify everything."
Fine in spirit, but ultimately useless in practice, because nobody has the time or space to verify everything. And if you could, you wouldn't need the tool in the first place.
My approach is to verify the things I might say in front of other people. The moment a claim leaves my head and heads somewhere someone else will see and interpret and evaluate, it stops being a passing thought and becomes something I need to put my name behind. That's my line. Verify what you'd be willing to stand behind in public, and let the low-stakes stuff stay low-stakes.
But the deeper reason to keep a line like that isn't really about any one claim. It's about the information, sure — but it's about exercising the habits you should be exercising with humans. Skepticism is a habit, and habits don't stay in the chat window. It's the same reason I say please and thank you to a model: not for its sake, but because how I practice being in one place is how I end up being everywhere. Without some kind of line, I think there's a bit of a descent into AI madness. 🙃
"Don't give it vague prompts — give it context."
This might be the single most repeated piece of AI advice that exists. Every guide says it. But there aren't enough good examples of what "context" actually looks like. So here's how I think about it, using something I actually want for my birthday: things from Expedition 33.
Rung one — say what you want. "I like games" is wildly open. Board games? Video games? Tabletop? The X Games? There's nothing to act on. Compare that to: "I want art and desktop figurines from Expedition 33." Now there's a real target instead of a guess.
Rung two — say what you want, with examples (kept broad). Building on rung one: "I want figurines and art from Expedition 33, specifically characters and montages." Those examples guide without trapping. But narrow the example too far — "I want things like Esquie, and that's it" — and it becomes the anchor. The model, especially a smaller or faster one, drags everything toward that single point instead of the category you meant. (Broad example: good. "I specifically want Esquie": the wrong kind of specific — though it is the right choice, because I genuinely want an Esquie plushie.)
Rung three — add what you don't want, with the why. "I don't want Expedition 33, the game." On its own that sounds bizarre — it's a birthday list made of Expedition 33 — right up until: "because I already have a copy." The reason is what turns a confusing exclusion into useful direction.
Rung four — the floor. Take rung three and strip the "because" off. Just "I don't want the game." No reason, no direction — a wall with nothing explaining it. That's the worst prompt youncan write.
*The through-line down all four rungs: the more you steer by saying what you do want, the better, and the moment you lean on "don't," you'd better bring the why* — or you've built a wall and left the model, or the gift-giver, standing in front of it with no idea why it's there.
"Don't just copy-paste raw AI output."
Just don't.
There's a difference between saying "I understand A, B, C, and D, and this is what I want to do with them" versus "pick letters for me and make something that sounds good." One is direction. The other is outsourcing the understanding itself.
When content gets generated and nobody checks it, that's how you end up with prompt responses printed in newspapers. People touting "I generated this content" feel so shallow to me. Cool. I breathed today. I ate food today.
When AI does the writing, you lose three things: ownership, critical thinking, and creativity.
I've run into this in my own app. The first version of a feature I created did things backwards — it made the user the editor of the AI's content, instead of the other way around. I should be the editor, and the user should be the author. It took real work to break the process apart so the user's language is the overwhelming majority of the words, and the AI is just stitching and tidying.
I typically draw the line at prescription versus question. Is the model telling me what I should do, handing me a finished statement, changing my words for me? That's prescription, and I refuse it. Or is it asking, reflecting, handing my own material back to me? That's a question, and that's the version worth keeping. There's a place for a little of the former — light help with wording, because being a weak writer shouldn't keep anyone from applying or being heard — but it stays the exception, never the default.
"Make everything agentic! "
Actually, don't.
It's a fascinating concept, but it comes with hidden dangers — and a level of chaos that, in my opinion, hasn't been thought out well enough before being slammed into every piece of software we touch on a daily basis. Right now it seems like people are just slapping a bunch of tools together and saying, "gee, I hope this works." Is that really how we want things to work, though? There's absolutely a time and a place for these agentic models, but it has to be a deliberate choice. And honestly, there's just as much deliberate choice in not using it — the same way there's a deliberate choice to add or exclude AI in the first place.
An agentic chain is like a weird game of telephone. You talk to your neighbor, your neighbor passes the message to the person in front of you, and that person responds to you — but you never get to talk to them directly. Just food for thought: if you've ever flipped on "thinking" mode in a chat (which, these days, is often on by default), you've already engaged in an agentic chain. The moment that toggle is switched, there's an interpretive step sitting between you and the answer. And if you're not paying attention to the fact that it's happening, it can get genuinely confusing what's actually going on.
Let's keep the telephone game metaphor going. If you have multiple people playing, you know how the story goes — the more people that play, the more garbled the message gets. And that final person, more often than not, gives back something completely nonsensical compared to the original statement. It's definitely really cool when the original makes it all the way around. But that's not the norm. That's the game we're playing when we set up agentic systems — and the math backs it up. In a multi-step chain, errors don't add up, they multiply. A five-step process that's 95% right at each step is only right about three-quarters of the time by the end. That's quiet chaos, and it's really hard to monitor. When you run a regular function and it doesn't work right, usually it's handled, and if not, the software screams at you. In these chains, that error in thinking just gets passed from one step to the next.
So with that knowledge, I personally choose to keep things inspectable — giving myself visibility into each step, so I can observe what's happening throughout.
For what it's worth: this has been my preference for a while, and based on where the industry has ended up, it seems like I made the right choice.
"Use AI to augment people, not replace them."
As author Joanna Maciejewska put it: we wanted AI to handle the laundry and the dishes so we could go create art — not the other way around.
But what we see is a loss of entry-level jobs and a replacement for creativity and critical thinking. It's just that easy to use it incorrectly. So what does it mean to use it in such a way that you can do more than you could by yourself?
My app is a good example. On my own, it never would have gotten done. But because I had tools available that augmented my ability and helped translate my thoughts and ideas, I was able to make something I think is meaningful. Every concept and idea in it came from my head — the tools let me explore them, refine them, test them, and build them. Honestly, the tools I used only ever worked well when I understood what I was asking for. When I didn't have a clear concept, or did a hand-wave and said "make this for me," at best the result was mediocre — but typically it was just far from what I wanted. And on top of that, as I was building this thing — which is honestly very personal — there's no way the model can replicate my lived experience.
Automation processes you into an output. Augmentation makes you more than you were on your own. So whatever end you have in mind for what you're doing, just stop and consider how AI is getting you to that end.
"Keep a human in the loop."
Part of being human is being witnessed. Without that, do we really exist?
A human cares, and AI emulates caring. I won't pretend the emulation isn't getting better — it is, and quickly. But when you sit across from a person, they're not just reading or hearing your words. They're observing your face, interpreting your tone, and the pause before you answer — maybe reading the kind of day you're having. Emulated caring optimizes for making you feel cared for. Actual caring just connects with you on a human level.
That's why the human in the loop exists. Not rubber-stamping the output — but using the experience of being alive to understand what another person is feeling and living.
So there you have it. A bunch of exposition on a bunch of statements you might be hearing on a daily basis. A lot of this was mostly my opinion, but I've tried to base it on everything I've read, written, and built.
I don't think AI is good or bad. I think it takes the shape of the people that put it in place. And I also think that at every step along the way, each choice deserves a why.
For some of you, you're probably happier with no AI in the world. For some of you, you're glad to see it added to everything. But I think most people land in the middle. And in that middle space, where we're bombarded with do's and don'ts, asking why can go a really long way.