AI Safety Practitioner

What you'll do

You judge how frontier AI models behave when the question is neither clearly fine nor clearly forbidden. The ad calls these "grey-area" topics, and names the territory: misinformation, political persuasion, self-harm, violence, cyber, biosecurity and other sensitive domains.

For each response you assess safety, factual accuracy, policy compliance and overall quality, flag unsafe outputs, hallucinations, reasoning failures and policy violations, and write structured feedback. You also apply and help refine the rubrics used for RLHF, SFT and safety benchmarking, and work alongside AI researchers and safety teams.

The core skill is consistency. Two reviewers looking at the same borderline response should reach the same verdict for the same reasons, and that is harder than spotting obvious violations. Expect to spend real time reading policy documents and calibrating against examples before your judgments count.

Be realistic about the content, too. Reviewing self-harm and violence material at volume is emotionally heavy, and the ad does not mention wellbeing support, rotation or exposure limits. Ask about that before you commit.

Who fits

Required:

• Bachelor's or higher in journalism, communications, psychology, sociology, public policy, law, biology, chemistry, computer science or a related field
• 5+ years in AI safety, trust and safety, journalism, public policy, scientific research, security or similar
• Excellent written English, critical thinking and analytical reasoning
• The ability to evaluate nuanced, policy-sensitive scenarios consistently

Preferred: prior AI safety, RLHF, SFT or evaluation work; familiarity with safety policies, content moderation or rubric development; experience with high-risk or ambiguous content.

The degree list is deliberately wide. A former fact-checker, a policy analyst and a biosecurity researcher all qualify, and a mixed pool of backgrounds is presumably the point.

Where you can be based

The ad publishes a hard location list of 40 countries: the United States, the United Kingdom, the EU member states it names, Iceland, Liechtenstein, Norway and Switzerland, the microstates Monaco and San Marino, and several non-EU countries in southeast and eastern Europe (Albania, Bosnia and Herzegovina, Kosovo, Moldova, North Macedonia, Serbia). Canada and Australia are not on it.

What it pays

$60–70 per hour, Mercor's published figure. Payment is weekly via Stripe or Wise. H-1B and STEM OPT candidates cannot be supported. Weekly hours are not stated.

Worth knowing

Good:

• Broad eligibility: many degree fields and a long country list
• Substantive work shaping how models handle contested topics
• Direct collaboration with safety researchers, per the ad

Less good:

• 5+ years is a firm bar that rules out early-career moderators
• Distressing content is part of the job and no support measures are described
• Narrow band ($60–70) with no stated route to higher rates
• No hours, duration or end client disclosed
• Independent contractor; projects can be extended, shortened or ended early

About this listing

Posted by Mercor as a remote hourly contract, confirmed open on 24 September 2026. The pay, country list and qualifications are the ad's own. See Mercor.