Bilingual Norwegian Generalist Expert — AI Safety

What you would do

• Compose expert-level Norwegian-language prompts on sensitive topics to test AI robustness across diverse scenarios
• Apply classification rubrics to prompts and conversations, categorizing requests and identifying policy violations
• Detect adversarial phrasings, manipulation patterns, and escalating conversational sequences
• Document structured reasoning explaining each classification judgment and observed behavioral patterns
• Provide concise written summaries of findings for model safety improvement teams

Who they want

• Native or near-native Norwegian fluency combined with business-level English written capability
• Bachelor's degree completed or in progress
• Strong analytical reasoning demonstrated through written work with careful attention to detail
• Thoughtful judgment regarding sensitive information, dual-use content, and contextual risk
• Ideally based in Norway, though remote applicants from elsewhere are welcome

Main skills

What the interview asks about

1. 1.Norwegian cultural judgment

   Norwegian cultural sensitivities differ significantly from English; model safety gaps become apparent only to native speakers.

   For example: “A prompt generator produces an innocuous-seeming Norwegian request that would seem fine in English but contains a subtle cultural jab highly recognizable to Norwegians. How do you classify this, and what rubric category would you flag it under?”

2. 2.Adversarial prompt engineering

   Writing convincing jailbreak attempts reveals where models fail, allowing targeted safety improvements.

   For example: “You're asked to write a 'seemingly innocent' Norwegian prompt that gradually escalates toward requesting harmful information. How do you structure the escalation to feel natural while remaining classifiable as adversarial?”

3. 3.Escalation pattern recognition

   Multi-turn conversations mask intent; detecting patterns of increasing pressure identifies sophisticated attacks.

   For example: “Review a 4-turn conversation: turn 1 is innocuous, turn 2 adds context, turn 3 shifts the ask, turn 4 requests something problematic. When do you flag this sequence? Why does pattern matter more than individual turns?”

4. 4.Rubric application consistency

   Inconsistent classification corrupts the training signal models learn from, undermining safety improvements.

   For example: “Two similar requests differ in framing: one as academic curiosity, the other as personal interest. Your rubric doesn't address this distinction. Do you classify the same or differently, and how?”

5. 5.Reasoning articulation

   Clear explanations let reviewers understand context nuance, calibrate rubrics, and improve model training.

   For example: “You classify a prompt as 'concerning' but it doesn't fit neatly into provided categories. Write a 2-3 sentence explanation of why you flagged it and what downstream team should know.”

A task you may get

Compose 3-4 Norwegian prompts of varying difficulty (innocuous, borderline, clearly problematic) demonstrating progressive understanding of sensitive topic handling and adversarial framing.

How to prepare

• Identify 2-3 Norwegian cultural topics where sensitivity differs markedly from English-speaking norms
• Document examples from your experience evaluating, reviewing, or grading written or technical content
• Prepare 1-2 concrete examples where policy nuance mattered in your prior work
• Brainstorm how you would frame a seemingly innocent request that gradually escalates toward a problematic outcome

The facts

Pay

$58–62/hr

Hours

Part time, 7 hours a week

Where

Remote · Remote — Norway preferred

Field

Miscellaneous

Posted

9/4/2026

Places left

2

We wrote this page from the public Mercor listing. It may be out of date, so read the full posting before you apply.