What the work actually involves

You write prompts in French — not casual test messages, but carefully constructed ones that push into sensitive subject areas where a model's behaviour is genuinely uncertain. That means thinking about how a real French speaker would phrase a borderline request: the euphemisms, the regional idioms, the framings that read as innocuous to an English-trained classifier but land very differently in French. The second half of the job is classification: applying a written guideline document to prompts and multi-turn conversations, deciding which bucket a case belongs in, and writing the justification. Reviewers read your justification as closely as your label, so the written reasoning is the deliverable, not a formality.

You will also flag adversarial phrasings and escalation patterns — conversations where each turn is defensible alone but the trajectory isn't. Expect the guidelines to be revised while you work, and expect to encounter cases the guidelines don't cleanly cover. Handling those honestly, by naming the ambiguity rather than forcing a label, is a large part of what separates useful contributors from volume producers.

What the screen looks for

Mercor's screening is AI-led and conversational, and for this role it concentrates on three things. First, real French fluency at a native or near-native level, with enough written English to work from English-language guidelines and write documentation your reviewers can read. Second, judgment around sensitive and dual-use material — whether you can distinguish a genuine risk from a topic that merely feels uncomfortable, and whether you understand why an evaluator writing adversarial prompts is doing safety work rather than causing harm. Third, whether you can explain a classification decision in a way another annotator could reproduce. A bachelor's degree, completed or in progress, is stated as a requirement. Trust and safety, content moderation, policy work, or prior grading and red-teaming experience are listed as nice-to-haves, not gates.

Logistics

  • Fully remote and asynchronous; work is claimed from a queue rather than scheduled
  • Roughly 7 hours per week, part-time, with an immediate start
  • Observed pay band of $48–52/hr for this listing — rates on Mercor vary by project and are not guaranteed
  • France or wider Western Europe is preferred for cultural currency, but applicants elsewhere are explicitly welcome
  • Sensitive content is inherent to the work; the subject matter can be genuinely unpleasant