Spend enough time in an online subculture and you discover that words don't always mean what you think they mean.

Ordinary expressions become signals of belonging. Abbreviations take on new meanings. A few characters can carry an entire worldview.

In incel communities, this language is known as a cryptolect: a shifting vocabulary that expresses shared beliefs, reinforces group identities and marks the boundaries between insiders and outsiders.

Understanding it takes more than compiling a list of offensive words.

The meaning often depends on context, and the same term can signal hostility in one conversation and be part of a perfectly legitimate discussion in another.

This creates a difficult problem for automated content moderation - When the meaning is hidden in plain sight, how do you teach a machine to see it?

General-purpose tools can miss coded language while flagging benign uses of the same terms. And as subcultural vocabulary evolves, systems trained to recognise surface-level toxicity can struggle to keep up.

I investigated this problem using a corpus of 1.8 million posts from r/Braincels. I developed a taxonomy of more than 400 subcultural terms and benchmarked a general-purpose moderation tool against a domain-adapted language model.

In this talk, I'll explore what happens when you bring digital anthropology into the engineering of AI-autonomous content moderation systems.

We'll look at how context changes meaning, why detecting subcultural language is a difficult classification problem, and what domain adaptation can reveal that general-purpose moderation tools miss.

Because your model can be fluent in a language and still have no idea what anyone is talking about.