The sociolinguistic foundations of language modeling

Harper: Here's something that sounds simple but opens up a lot. You're using an artificial intelligence writing tool, and it keeps flagging your dialect as informal, stumbling on your professional jargon, or just sounding slightly off for your context. Most people shrug and think the model isn't smart enough yet. However, Grieve, Bartl, Fuoli, and colleagues say the problem is more specific: the model was never trained on your kind of language. Adela: That's what I kept getting stuck on. These models are trained on billions of words. How can something that large still be narrow? Harper: That's the move worth walking through. The paper's core claim is that every language model is modeling a variety of language — not language in general, but some bounded slice. The slice is defined by three things: dialect, which is about who is speaking; register, which is about the situation; and period, which is about when. Adela: I am happy to help! However, I need you to provide a specific passage that you would like me to rewrite for naturalness in TTS. Please share the text, and I will take care of the rest. Harper: A doctor's clinical note is quite different from a text message to a friend. They use the same language and are written by the same person, but the register is totally different. A model trained mostly on clinical notes would excel in understanding clinical notes but would perform worse on texts. Now, imagine multiplying that across an entire training corpus. Adela: Right, but the web is enormous. Surely it covers most areas? Harper: The paper has a specific term for this — convenience sampling. You grab the largest and most accessible chunk of language available, which online means written, formal, English-heavy text. They contrast it with representative sampling, which means deliberately selecting across dialects, registers, and time periods. Nobody building a massive model is doing representative sampling. Adela: So the web tends to favor certain groups. Harper: Heavily. The analogy is polling only people who answer their landlines. Your conclusions aren't wrong about that group; they just don't generalize. Adela: Okay, that lands. But here's where I lose it — the paper claims this one idea explains five separate problems: social bias, domain adaptation, alignment, language change, and scale. That feels like a stretch. Harper: Walk through one. Adela: Take bias. The standard story is that bias comes from prejudiced text in the data. Harper: The paper reframes this issue. Bias occurs when certain dialects and registers are underrepresented during pretraining. The model learns the dominant variety effectively and performs worse on everything else — a situation they refer to as quality-of-service harm. The solution isn't simply removing offensive content. It involves examining whose language is included in the corpus and whose language is left out. Adela: That is a different diagnosis. Harper: Alignment shifts in a similar manner. It's often described as getting a model to follow human values. Grieve and colleagues refer to it as varietal misalignment. The model was fine-tuned using human feedback, but that feedback also originated from a specific variety of language and a specific population. Adela: Not broadly human values. Harper: Not broadly. The paper also addresses scale. The Chinchilla scaling law specifies around twenty training tokens for each model parameter, but the paper argues that scale only helps if it increases the sociolinguistic diversity of the corpus. They contrast the curated data of GPT-3 with the Pile of GPT-J as evidence that diversity drives gains, not just raw size. Adela: So, throwing more web text at it doesn't fix the sampling problem. It just deepens the same groove. Harper: That's it. The model isn't incomplete; it's complete at the wrong task. Adela: That framing is what finally changes my perspective. Harper: Grieve and colleagues conclude that understanding which variety a model was trained on should be standard practice — documented, evaluated, and designed deliberately. This is not abstract. It influences whose dialect gets corrected, whose writing gets flagged as wrong, and whose voice the tool amplifies. Language is always someone's language, and these tools are no different. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.