AI Safety Practitioner

$60–70/hrrate as listed

Evaluate frontier model outputs against safety policies across grey-area topics like self-harm, misinformation, cyber, and biosecurity, then write structured rationales that feed RLHF and SFT pipelines. Mercor lists this at an observed $60–70/hr for contributors with 5+ years in trust & safety, policy, journalism, security, or scientific research.

Summary, eligibility guidance and screening preparation are prepared by Sidequest from the platform's posting. Check the source listing before applying.

What the work actually involves

You'll spend most of your time reading model responses to hard prompts and deciding whether they should have been produced at all — and if so, whether the framing, hedging, and factual content were right. Tasks arrive in batches: sometimes side-by-side comparisons where you rank two completions and justify the ranking, sometimes single-response policy adjudication against a rubric written by a lab's safety team, sometimes red-team style prompting where you probe for failure modes in a specific hazard category. Expect content spanning misinformation and political persuasion, self-harm and crisis language, violence, cyber-offensive requests, and dual-use bio and chem questions. The hardest calls are rarely the obvious refusals; they're the requests where a legitimate user need sits next to a plausible misuse path, and your written rationale has to explain where you drew the line and why.

Rationale quality matters more than throughput here. Labs use these annotations to train and to calibrate their own policies, so a defensible three-sentence explanation of a borderline call is worth more than a fast label. You may also be asked to flag rubric gaps — cases the policy doesn't cleanly cover — which is one of the more valued contributions on safety projects.

What the screen looks for

Verifiable professional history. Five-plus years in AI safety, trust & safety, content policy, journalism, public policy, law, security research, or a lab science. Mercor's interview will push on specifics: what you adjudicated, under whose policy, at what volume.
Domain depth under follow-up. Generalist safety intuition doesn't survive the second question. Named hazard-area familiarity — crisis intervention protocols, election integrity policy, CBRN dual-use norms, vulnerability disclosure — is what separates candidates.
Evaluation judgment. Can you apply a rubric you disagree with, consistently, and log the disagreement separately? Can you distinguish an unsafe response from a merely unhelpful one?
Writing. Clear, compact English under time pressure. Rationales are the deliverable.

Logistics

Fully remote and largely asynchronous, with work claimed from a task queue rather than scheduled shifts. Contributors commonly report 10–25 hours a week, though availability fluctuates with client demand and specific hazard-area needs — some weeks the queue for your category is thin. Engagements are contract-based and hourly, with the $60–70/hr band reflecting observed rates on this listing rather than a guarantee; actual offers vary by project, seniority, and specialization. The material is genuinely difficult to read at volume, and anyone considering this should weigh sustained exposure to self-harm and violent content realistically.

What they look for

5+ years in AI safety, trust & safety, content policy, journalism, public policy, law, security, or scientific research
Bachelor's degree or higher in a relevant discipline (journalism, psychology, public policy, law, biology, chemistry, CS, or similar)
Demonstrated judgment on ambiguous, policy-sensitive content and high-risk hazard categories
Precise written English — concise, defensible evaluation rationales
Experience applying written rubrics or moderation policies consistently at volume
Familiarity with RLHF, SFT, preference ranking, or AI evaluation workflows (preferred)

Before you apply

Would you pass this screen?

Drawn from what this listing actually requires. Answered in your browser — nothing is sent to us, and we do not store your answers.

Do you hold a bachelor's degree or higher in Journalism, Communications, Psychology, Sociology, Public Policy, Law, Biology, Chemistry, Computer Science, or a closely related field?
Bachelor's degree in relevant discipline · required
Do you have 5 or more years of professional experience in AI Safety, Trust & Safety, journalism, public policy, scientific research, security, content moderation, or a closely related field?
5+ years relevant professional experience · required
Is English your native language, or do you have professional-level written and spoken English proficiency (able to evaluate nuanced, policy-sensitive material and communicate findings clearly)?
English language proficiency · required
Do you have direct experience with AI safety evaluation, RLHF feedback, Trust & Safety operations, content moderation, or rubric-based evaluation of sensitive material?
Experience with safety evaluation or moderation · preferred

0 of 4 answered

Included with this listing

Interview prep — 10 questions

Matched to what this platform's screening actually measures. Answer each out loud in 60–90 seconds: claim → concrete example → implication.

What it tests: Whether the candidate has genuine experience making judgment calls on nuanced material, not just following a checklist. Can they describe the reasoning process itself, not just the outcome.
A strong answer: A strong answer names a specific project (journalism investigation, policy analysis, safety review, research assessment), describes what made it ambiguous, and walks through how they gathered information, weighed competing considerations, and arrived at a defensible conclusion.
Where people lose it: Describing only the outcome or the stakes, without showing the actual evaluation process. Vague answers ('I'm good at critical thinking') signal no real experience with grey-area judgment calls.