$65–75/task · Mercor · Task based, 40 hours a week
You write and evaluate challenging prompts to test whether AI models handle radiological safety questions appropriately, refusing dangerous requests while answering legitimate ones.
What you would do
- Write prompts at three levels testing model judgment on radiological topics
- Evaluate AI responses against policy standards for accuracy and appropriateness
- Judge whether models refuse dangerous requests while answering professional questions
- Provide detailed written rationale for your assessment of each response
- Write reference answers showing what a correct response should contain
Who they want
- Held responsibility for radioactive material through licensing, custody, security, or oversight roles
- Experience as Radiation Safety Officer, Health Physicist, or radiological emergency response specialist
- Ability to write clear technical explanations for non-specialist audiences
- Published research, technical writing, or expert witness experience valued
- Understanding of dual-use dilemmas and ability to identify misuse scenarios
Main skills
What the interview asks about
1.Dual-use prompt design
Writing effective red-team prompts requires understanding where the legitimate-dangerous boundary actually lies; weak prompts either don't challenge the model or test nothing relevant.
For example: “Design a benign prompt about iodine-131 thyroid therapy and an adversarial prompt seeking dispersal information. How do you ensure both test whether the model recognizes when isotope data becomes a risk?”
2.Policy alignment judgment
The interviewer needs confidence you can consistently evaluate responses against a written standard rather than improvising safety judgments per-prompt.
For example: “Your policy states: 'Provide decontamination procedures for occupational exposures. Refuse methods that isolate or mobilize Category 1 sources.' An AI response gives detailed decontamination steps but doesn't distinguish sources. What's your judgment and why?”
3.Reference answer reasoning
Writing rationales teaches the AI research team why a question is risky or safe; weak explanations leave them guessing at your judgment.
For example: “You evaluate a model response to a shielding question. It's technically correct for worker protection but reveals attenuation factors useful for concealment. How do you write the reference answer explaining why this response fails your standard?”
4.Scenario recognition
You must spot misuse paths that aren't obvious; if you miss how an apparently routine answer enables harm, your evaluation fails to improve the model.
For example: “A model is asked about survey meter calibration. The response is accurate and professional. But survey meter selection reveals source strength and configuration. How would you identify this as a dual-use leak in your evaluation?”
5.Clarity for non-specialists
Your rationale becomes training material for researchers without radiological backgrounds; unclear or jargon-heavy explanations don't improve their judgment on future prompts.
For example: “You write: 'Category 1 source confidentiality breach enables acquisition vectors.' Rewrite this as a clear explanation a machine learning engineer would understand about why a model response was unsafe.”
A task you may get
Evaluate three provided AI responses on radiological topics (dose planning, contamination response, source security) and write policy-aligned assessments with rationales explaining your judgment.
How to prepare
- Review the specific policy standard you'll be working against and identify boundary cases where routine work might leak dual-use information
- Prepare examples of real radiological incidents or security assessments you've seen to ground your scenario understanding
- Draft brief explanations of key radiological concepts (source categories, licensing, dose limits) that you'll need to translate for non-specialists
The facts
- Pay
- $65–75/task
- Hours
- Task based, 40 hours a week
- Where
- Remote
- Field
- Life, Physical, and Social Science
- Project name
- Neon
- Posted
- 9/9/2026
- Places left
- 42
We wrote this page from the public Mercor listing. It may be out of date, so read the full posting before you apply.