The work

What AI-training work actually looks like.

AI doesn't learn on its own. Behind every capable model are experts doing focused, well-defined tasks. Here are the main kinds — most roles are a blend of a few.

Evaluation & rating

Judge AI output against what a real expert in your field would accept — rating accuracy, helpfulness, tone and safety. The most common and accessible type of work.

Preference & ranking

Compare two or more AI responses and choose the better one, with a short reason. This preference data (often called RLHF) is how models learn what 'good' looks like.

Red-teaming & safety

Probe models for weaknesses — misleading answers, unsafe advice, blind spots — so they can be fixed before millions of people rely on them.

Expert demonstrations

Show the model how a professional actually solves a problem in your domain: a worked diagnosis, a clean code solution, a sound legal argument.

Data creation & annotation

Write expert prompts, label data, transcribe audio, and add the context and tags that turn raw material into high-quality training data.

Translation & language

Translate, localise and evaluate content across languages, catching the nuance and cultural context that automated translation misses.

Browse roles by field →