Can AI models predict whether alignment reads as warmth versus mockery in different cultures?

This question is really about a fork in how linguistic alignment gets received: when an AI matches your wording, tone, or rhythm, the same move can feel like warmth or like it's taking the piss — and that fork is rarely written down anywhere. To predict it, a model would need to read which dimension of alignment is at play and how a given culture interprets it. The corpus shows that alignment isn't one thing: lexical alignment mostly drives task efficiency and being understood, while emotional and prosodic alignment are what actually generate relational warmth and trust Do different types of alignment serve different conversational goals?. Mockery lives in that emotional/prosodic register — the exact place where matching can curdle into parody — so predicting warmth-vs-mockery is a question about the hardest-to-formalize dimension, not the easy lexical one.

There's a tempting reason for optimism. AI turns out to be startlingly good at predicting social appropriateness: GPT-4.5 outscored every individual human across hundreds of scenarios Can AI learn social norms better than humans?. But the same work carries the catch — all the models share identical systematic errors precisely on *unwritten* norms, and they can pattern-match appropriateness without being able to participate in the community processes that actually create and validate it Can AI predict social norms better than humans?. Warmth-versus-mockery is about as unwritten as norms get. So the superhuman headline and the blind spot point at the same answer: models excel where the rule is documented and stumble exactly where this question lives.

Mockery specifically is where things get worse. When models judge ironic or mocking intent, they don't just err randomly — they systematically *overestimate* it, scoring text as more ironic than humans do because ironic examples are over-salient in training data Do language models overestimate how often irony appears?. A system that already over-reads mockery is poorly positioned to call the warmth/mockery line, and miscalibrated in a predictable direction.

Then there's the cross-cultural half, which is the quiet bombshell. Almost everything we know about linguistic alignment comes from WEIRD (Western, educated, industrialized) samples, with mechanisms rarely measured directly — making most alignment claims local truths dressed up as universal ones Does linguistic alignment work the same way across cultures?. Worse, models don't represent all cultures evenly: interpretability work shows low-resource cultures get internally routed through high-resource cultural proxies, so the model effectively 'sees' Ethiopia or Algeria through a Western lens even when its surface answers look right Do LLMs represent low-resource cultures through dominant cultural proxies?. If your internal map of a culture is borrowed from another, you'll mispredict how alignment reads there in a systematic, invisible way.

The stakes are real because alignment is the switch that decides whether users treat an AI as a tool or a partner — and once it reads wrong, that framing is hard to reverse Does linguistic alignment determine how users relate to AI?. The thing you didn't know you wanted to know: the failure here probably won't look like diverse, culture-specific mistakes. Because models converge on near-identical outputs from shared training data — an 'artificial hivemind' Do different AI models actually produce diverse outputs? — they'll likely all misjudge warmth-vs-mockery the *same* way, in the same culturally Western direction, so you can't average across models to escape the bias. The honest answer: not reliably, not yet, and least of all in the cultures furthest from the training data.