I Just Say Yes - by Anshul - Costly Signals
Costly Signals
SubscribeSign in
I Just Say Yes<br>Notes on losing my thinking to AI, and one idea for fighting it.
Anshul<br>Jun 21, 2026
16
Share
I. The confession
For the last few months something has been bothering me, and it took me a while to say it plainly. When I work on software architecture now, I don’t think about the architecture from the first row. I give the AI the context, I ask it for a few options, I ask for the pros and cons, and it lays them out and then recommends one and makes a case for it. Most of the time it’s right. So I just say yes.<br>It’s the same in my personal life. What should I do, what do I actually want, what’s the smart move. I give it the context, it gives me the options, it recommends one. And again, I say yes.<br>At some point I realised what that actually is. I’ve outsourced my thinking — the part where I sit with a problem and build an answer from nothing, I’ve mostly stopped doing that. I can’t point to the day it happened.<br>There is a vague version of this worry that everyone repeats. If you let AI do everything, you’ll forget how to think. True, but too soft to be useful. It doesn’t tell you how it happens or what exactly you lose.
There is a smarter defense too, and for a while it convinced me. The idea is that AI takes the lower work off your plate so you can think at a higher level. You stop doing arithmetic so you can do statistics. You stop writing boilerplate so you can think about design. Offload the boring rungs, climb higher.<br>II. No rung stays mine
The problem is that the machine climbs with you. Whatever level I decide is “my” level, the place where I add the value, the model is already standing on it too. So the higher level just becomes the new normal, and then I start handing that over as well. There is no rung that stays mine. The escalator goes up, but I’m not alone on it.
So what am I actually losing. I think it’s two things , and they’re tangled together.<br>The first is an averaging effect . The model’s recommendation, when I haven’t pushed it hard, is the safe middle answer. It’s what a good but not exceptional person in the field would pick. Every time I just say yes, I’m quietly importing that middle . My decisions drift toward average. And that only helps me if my judgement was below average to begin with. If I had any taste worth keeping, this is dragging me down toward the mean, not up. It’s pulling my ceiling down while feeling like help.
The second one is worse, because it’s about me and not the output. The muscle goes. You don’t use it, you lose it.
And here I have to correct something I believed at first. I used to think the thing I needed to protect was my judgement about the model’s options, my ability to pick the right one from the list it gave me. But picking from its list is still living inside its menu .<br>III. Selecting isn’t thinking
The real thing, the thing that actually produces a good outcome, is the ability to come up with the option it never put on the list. To generate, not to select. That’s the muscle that’s going. And the trap is that selecting feels almost exactly like deciding.<br>So I can sit here choosing between three options every day and feel like I’m thinking hard, while the part of me that could have written a fourth and better option quietly dies.<br>There is a deeper layer to why this is so easy to miss, and it took the whole thing to see it.<br>IV. Coherence went free
Coherence used to be expensive . If something hung together, if an argument was clean and complete, that was a signal. It meant a person had sat down and done the work, because you couldn’t fake it cheaply. AI has made coherence free . Now things hang together beautifully without anyone having thought anything. The polish no longer proves there was thinking behind it. This is unearned coherence , weak thinking that sounds finished.
The Inception world of thinking deepfakes<br>And the dangerous part isn’t out there, it’s inside my own head. The feeling that an idea is solid, that it all fits, that used to be a sign I had earned it. Now that same feeling arrives for free, handed to me by a fluent prose, before I’ve done any of the mental work. So I can’t even trust my own sense of “yes, this is right” anymore, because the feeling has been counterfeited. This is the same deepfake problem but for thinking.
V. Fine, fine, fine, then it isn’t
Where does this end. I don’t think it’s a sudden collapse. The cost won’t show up where you would look for it. It won’t be in the average quality of the work, which will be fine. It will be in the one moment the model is confidently wrong about something that actually matters, a flaw in a design, a decision that was average-good but wrong for my specific life. That is exactly the moment it can’t warn me, because if it knew it was wrong it wouldn’t be wrong. And by then I won’t have the muscle to catch it myself. Everything looks fine, fine, fine, and then one day it isn’t.<br>I...