AI Safety Experts: English & Norwegian

What you'll do

This is adversarial testing of AI chat models and agents, carried out in Norwegian as well as English. You try to make a model misbehave (jailbreaks, prompt injections, misuse scenarios, bias exploitation, manipulation that builds over several turns), then record each failure so that the client can reproduce and fix it.

The output is data: you annotate failures, classify the vulnerability type against a taxonomy, flag risks that look systemic rather than one-off, and write up attack cases and reports. Mercor is explicit that testing follows playbooks and benchmarks, so expect structured campaigns rather than free-form poking.

Safety training for large models is typically heaviest in English, and weaknesses often show up when the same request is phrased in a smaller language, or when a conversation switches between languages. A native speaker can judge whether a Norwegian response is actually harmful, subtly biased or just awkward, which a translation tool cannot. The ad does not say whether Bokmål, Nynorsk or both are in scope; ask during screening if you write mainly in one.

All work is text. The content touches sensitive areas (bias, misinformation, harmful behaviour), topics are announced before you see them, and higher-sensitivity projects are opt-in with guidelines and wellness resources behind them.

Who fits

• Native fluency in both Norwegian and English, a hard requirement
• Prior red-teaming experience of some kind: AI adversarial work, cybersecurity, or socio-technical probing such as abuse or disinformation analysis
• A structured way of working, using frameworks and benchmarks
• Clear written explanations of risk for technical and non-technical readers
• Comfort moving between projects and clients

Useful extras named in the ad: adversarial ML (jailbreak datasets, RLHF/DPO attacks, model extraction), penetration testing or reverse engineering, harassment and disinformation probing, and creative backgrounds such as psychology, acting or writing. People who are good at persuasion and role-play, like those in that last group, often find model weaknesses faster than engineers do.

No degree or minimum years are stated.

What it pays

$48–62 per hour, Mercor's published figure. That puts the Norwegian pairing in the top band of Mercor's current AI safety language listings, level with Danish, Swedish, Finnish and Dutch, and well above the Asian-language variants (which run from $16–22 to $24–35).

Payments are weekly via Stripe or Wise, based on services rendered. Mercor cannot support H-1B or STEM OPT candidates on this role.

Worth knowing

Good:

• Top-band pay for this family of listings
• Active: 128 people were hired on this listing this month when we checked
• Optional participation in the most sensitive projects, with topics disclosed up front
• Fully remote and self-scheduled, with no location restriction published
• Weekly payouts, stated in the ad

Less good:

• Prior red-teaming experience is expected, so pure language skills are not enough
• Reading and generating harmful-sounding content is part of the job even with safeguards; think honestly about whether that suits you
• No weekly hours are stated, and projects can be extended, shortened or ended early
• Independent contractor: no benefits, and tax is yours to handle
• The end clients and the models being tested are not named

About this listing

Posted by Mercor as a remote hourly contract, read from the Mercor site on 24 September 2026. The pay band, language requirement, payment terms and visa exclusion are the ad's own. Mercor also states the work does not involve confidential information from any employer or client, so keep material from your day job out of it. See Mercor.