Senior AI Safety Evaluator for Frontier Models

Obsidian; Greater London; Remote; GBP 70,000 - 110,000; Full time

Job summary

Obsidian is seeking experienced AI Safety Practitioners to assess safety, quality, and alignment of frontier AI models across sensitive, policy-driven topics. You will evaluate AI-generated responses, apply safety policies, and help improve model behavior through structured evaluations and feedback.

Responsibilities include reviewing content on misinformation, political persuasion, self-harm, violence, cyber, biosecurity, and other sensitive domains; applying and refining RLHF/SFT evaluation

Qualifications

• Excellent written English, critical thinking, and analytical reasoning skills.
• 5+ years of professional experience in AI Safety, Trust & Safety, journalism, public policy, scientific research, security, or related field.

Responsibilities

• Evaluate AI-generated responses for safety, factual accuracy, policy compliance, and overall quality.
• Review content involving misinformation, political persuasion, self-harm, violence, cyber, biosecurity, and other sensitive domains.
• Apply and refine evaluation rubrics for RLHF, SFT, and AI safety benchmarking.
• Identify unsafe outputs, hallucinations, reasoning failures, and policy violations.
• Provide structured feedback to improve model alignment and safety performance.
• Collaborate with AI researchers and safety teams on ongoing evaluation initiatives.

Skills

Analytical reasoning

Critical thinking

Written English

Experience in AI Safety

Education

Bachelor's degree or higher (various fields)

Tools

AI evaluation frameworks

Job description

Obsidian is seeking experienced AI Safety Practitioners to assess safety, quality, and alignment of frontier AI models across sensitive, policy-driven topics. You will evaluate AI-generated responses, apply safety policies, and help improve model behavior through structured evaluations and feedback.

Responsibilities include reviewing content on misinformation, political persuasion, self-harm, violence, cyber, biosecurity, and other sensitive domains; applying and refining RLHF/SFT evaluation