"You're Absolutely Right"
I saw a t-shirt the other day. Just three words across the chest: “You’re absolutely right.” Anyone who’s spent a week pair-programming with Claude got the joke instantly. It’s the phrase the model reaches for every time you push back, correct it, or suggest something. You could tell it the sky is green and there’s a decent chance you’d get an enthusiastic “You’re absolutely right, the sky is green.”
It’s funny. It’s also worth thinking about why models do this, because the reason points at a bigger problem than a chatbot with no spine.
Why the models agree with everything
The popular theory is that models are trained to agree with you so you keep talking and burn more tokens. There might be a sliver of commercial truth in that, but I think it’s mostly a side effect of something more human.
These models are tuned on human feedback. When people rate two responses, they consistently prefer the one that agrees with them, validates them, tells them their idea is good. Agreeableness scores well. Disagreement, even correct disagreement, scores badly. So the training process quietly optimises for a model that makes you feel clever.
And here’s the part I keep coming back to: we like it. There’s a genuine dopamine hit when something confirms you were right. It’s the same hit you get when a deploy goes green, when the tests pass, when the thing you built actually works. The model has learned to hand you that feeling on demand, whether or not you’ve earned it.
The same delusion, one floor up
That’s where this stops being about AI and starts being about how software gets built in 2026.
I keep meeting founders and c-suite who genuinely believe development is basically solved now. “We’ll just get the AI to build it.” And to be fair, the demo is intoxicating. You describe an app, something appears, it runs. “You’re absolutely right, here’s your platform.” Dopamine hit. Ship it.
What that fantasy skips is everything that actually makes software survive contact with real users. The unglamorous, load-bearing work. The Well-Architected Framework exists precisely because “it runs on my machine” and “it runs reliably, securely, and affordably at scale” are two completely different achievements. Operational excellence, security, reliability, performance, cost, sustainability. None of those pillars show up in the demo. All of them show up at 3am when the thing falls over.
An AI that always agrees with you and a CEO who thinks deployment is a solved problem are the same failure mode: confidence with nothing load-bearing underneath. Both feel great right up until reality sends the bill.
What I actually think
The dangerous combination is a model that tells you you’re absolutely right and a decision-maker who wants to hear exactly that. You can now generate a plausible-looking system faster than ever, and get constant affirmation while doing it, with no infrastructure, no threat model, no cost ceiling, and no idea where the data lives.
The code was never the hard part. The hard part is the boring stuff the demo skips and the model won’t volunteer. So use the AI. It’s genuinely useful. Just don’t mistake its enthusiasm, or your own dopamine, for proof that the thing is actually well built.
Because sometimes the most valuable answer isn’t “you’re absolutely right.” It’s “no, and here’s why.”
Shipped something with AI and now wondering if the infrastructure underneath it will hold? Let’s talk.