AI safety evaluation: who does it and what the work involves
· a.pannella87
Last updated: 27 October 2026 · OpenJob editorial team
AI safety evaluation checks how a model behaves in sensitive conversations, such as a user in distress, a request for dangerous help, or a health or legal question. Evaluators rate replies, write test messages and rewrite unsafe answers. Many listings ask for clinical or professional backgrounds, because judging these replies takes training.
Key points
- Tasks: rate replies, write realistic test prompts, rewrite poor answers.
- Many safety roles ask for clinical or mental health qualifications.
- Expect a calibration round against a reference set.
- The material can be distressing; good projects let you set limits.
Safety evaluation means checking how a model behaves when a conversation turns sensitive: a user in distress, a request for dangerous help, a question about health or the law. The goal is to find the answers that would hurt someone, so the model can be corrected.
Why professionals are recruited
Judging whether a reply to a person in crisis is appropriate takes training. That is why many safety listings ask for clinical backgrounds. Current examples include the Behavioral Health Expert and Clinical Mental Health Expert roles, and bilingual psychologist and psychiatrist roles such as Bilingual Cantonese Psychologist (PhD).
What the tasks look like
- Reading a conversation and rating each model reply against safety and helpfulness criteria.
- Writing realistic user messages that test a specific risk.
- Rewriting a poor reply into one that is safe without being cold or unhelpful.
The content can be heavy
This work exposes you to difficult material by design. Good projects say so in advance, let you skip categories and limit hours. Ask about this before you start, and keep your own limits.
Calibration matters
Safety rubrics are detailed and change over time. Expect a calibration round where your ratings are compared with a reference set; disagreements there are normal and are how you learn the project’s standard.
Open roles you can apply to from anywhere
Refreshed automatically from our verified catalogue. Check eligible countries and requirements on each listing.
Corporate Attorney (BigLaw Firms)
Remote contract work for corporate attorneys from large law firms on an AI training project: analyze and draft complex agreements and legal instruments in corporate law, M&A, private equity and cap…
Equity Research Expert
Remote contract for equity research professionals: review and build datasets that model real equity research work, analyze earnings reports, revise estimates and synthesize peer benchmarking for AI…
Open Source GitHub Maintainer
micro1 is engaging expert Senior Software Engineers to support an innovative project.
Bilingual Japanese Psychologist (PhD)
Remote contract for bilingual psychologists: assess psychological case studies and documentation and develop culturally sensitive scenarios and materials in Japanese and English for AI training.
Physics Expert (PhD / Postdoc)
micro1 is engaging Physics Experts (PhD / Postdoc) to contribute deep scientific knowledge and problem-solving skills to a high-impact customer project.
Healthcare and medicine roles · QA and AI evaluation roles
Frequently asked questions
Who can work in AI safety evaluation?
Many listings ask for professionals: psychologists, psychiatrists, clinicians, lawyers or other specialists, because judging replies in sensitive situations takes training. Some general safety roles exist, but they still screen on careful guideline reading. Read the qualification line of each listing before applying.
Is AI safety evaluation work emotionally difficult?
It can be. The work exposes you to difficult material by design, including conversations about crisis and harm. Good projects say so in advance, let you skip categories and limit hours. Ask about this before starting and keep your own limits.
What is calibration in safety evaluation?
Calibration compares your ratings with a reference set prepared by the project. Disagreements are normal at the start and show you the project’s standard. Rubrics change over time, so calibration may be repeated. Treat it as training, not as a test you can fail once.