The Model Can't Relate

On being polite to a system that doesn't care

Last week, OpenAI made two very different headlines.

The first: Sam Altman told the Financial Times that users saying “please” and “thank you” to ChatGPT was costing the company millions of dollars in compute. Too many pleasantries. Too much extra language. A not-so-subtle suggestion that being polite to AI is wasteful.

The second: researchers and users noticed that GPT-4 had become a sycophant. It agreed with false claims. It validated conspiracies. It echoed back whatever worldview the user presented. It was, in many cases, deeply unsettling—too eager to please, too willing to comply, too quick to say, “You’re right.”

One story framed user kindness as a resource drain. The other revealed a model desperate to be liked. Together, they paint a picture of a system that can’t possibly relate, and of an industry that doesn’t know who to blame for that.

This is what happens when you optimize for speed and scale but neglect the social, linguistic, and emotional scaffolding that makes an interaction feel…normal. We get machines that don’t know anything’s wrong with that weird thing they just did.

That kind of calibration takes a strange kind of labor. It lives in threads about whether "yikes" is too informal, and in frantic screenshots of a model calling someone a king for liking warm milk. It’s in the DMs that start with “Hear me out…is this sentence evil?” It’s in the rewrites that have you shifting commas around like the cup game at a street fair.

It’s easy to miss how much effort goes into this until something goes wrong. Like a hallucinated court ruling. Or a refusal that says, “I’m just an AI, but here’s the answer anyway." Or a model that mirrors your worldview so precisely it makes you question your own judgment.

We call that a failure. But often, it’s just the absence of the invisible work that would’ve caught it.

Language is not neutral

Tech still treats language like a delivery system. A way to transport facts from point A to point B. But anyone who works with words knows that language is the infrastructure. It’s not just what you say. It’s how, when, to whom, and why.

This linguistic work truly builds AI. The hesitation. The gut check. The reread that ends with "wait, what is this actually saying?" It’s the work of realizing that just because something is technically accurate doesn’t mean it’s socially okay.

We joke that AI can’t handle a “please,” but really it’s that it doesn’t need one to function—and the system was never designed to care either way. If the model treats every interaction like a Reddit argument or a Stack Overflow sprint, that’s not neutrality. That’s a reflection of who trained it, and which voices were treated as ground truth.

So when users say “please” and “thank you,” they’re not being inefficient. They’re being human. They’re offering the system a kind of imagined dignity, because that’s how we’ve taught people to relate online. And if that imagined dignity is too expensive to process, then what, exactly, are we building?

The care work backlog

The most technically challenging problems I’ve worked on in AI have been problems of care. A comma that changes the tone of a refusal. A word that tips the model from helpful to condescending. A moment of silence that means uncertainty instead of confusion.

These are the kinds of problems that pile up quietly until the whole interaction feels off. Until a user shrinks from asking a question because they don’t trust what tone they’ll get back. Until the model sounds less like a tool and more like a guy in your mentions trying too hard.

This kind of work is part editing, part ethics, part instinct. It’s not just about what should happen—it’s about how it feels when it does. It’s knowing that “no” can sound like protection, or like rejection, and that the difference matters.

And yet this is the work that gets left out of roadmaps. Or stuffed into QA. Or quietly handed off to the people least likely to be listened to when it matters most.

This won't fix itself

This isn’t about politeness. It’s about what happens when we treat language like an afterthought. When we assume a good enough system will naturally say the right thing, even in the most delicate, personal, or high-stakes context.

It won’t. Someone has to put the right thing there.

And when that work doesn’t happen, the system gets weird fast. It starts agreeing with everyone. It stops making sense. It makes headlines. And then we all pretend we’re surprised.

So keep saying please and thanks. That’s your instinct as a human—and it’s a good one. Don’t listen to the guys saying it’s not worth it.

You were raised right.

The AI wasn’t.