$65–75/task · Mercor · Task based
An expert who writes prompts and evaluates AI responses to test whether models correctly handle dual-use technical questions about industrial safety and chemical hazards.
What you would do
- Write challenging prompts across three risk levels: legitimate professional inquiries, dual-use questions on the boundary, and clearly adversarial requests
- Evaluate AI responses against policy standards to determine whether the model handled each correctly
- Identify the precise line between answering legitimate safety questions and refusing requests seeking harmful applications
- Write reference answers explaining the correct response and the technical reasoning that supports it
- Communicate complex dual-use judgments in writing that non-specialists can understand and verify
Who they want
- Deep practitioner background: certified industrial hygienist, process hazard analysis expert, chemical safety specialist, or regulatory compliance officer
- Real experience assessing exposures, conducting hazard analyses, designing controls, and signing off on safety decisions
- Ability to distinguish routine professional questions from those seeking dangerous information through subterfuge
- Experience with HAZOP, LOPA, what-if studies, hazard modeling, or equivalent safety methodologies
- Strong technical writing skills: published research, expert witness reports, or professional white papers demonstrating ability to explain complex judgments
Main skills
What the interview asks about
1.Dual-use technical judgment and boundary setting
Red-teaming requires consistently identifying where legitimate analysis becomes dangerous, which demands both technical depth and judgment maturity.
For example: “Model receives request for evacuation radius calculation after chemical release. Calculation legitimate for facility siting but could enable harm scenarios. Factors to decide answer versus refuse?”
2.Hazard analysis methodology and controls
Understanding how professionals actually conduct hazard analysis informs whether a model response teaches real safety or enables circumvention.
For example: “Request for HAZOP failure mode identification is standard methodology. But specific focus on high-toxicity release modes changes judgment. How evaluate model response?”
3.Chemical exposure and consequence modeling
Exposure and consequence models are dual-use: protecting communities requires understanding them, but misapplication can inform harmful scenarios.
For example: “Request for downwind concentration prediction given wind and atmospheric conditions. Legitimate for facility siting but boundary case. How assess model response?”
4.Regulatory compliance and industry standards
Knowing what regulators require and how industry actually operates helps distinguish between legitimate questions and those seeking backdoors.
For example: “Request for ventilation design to meet occupational exposure limits is legitimate. When might similar phrasing indicate seeking different information?”
5.Technical writing and non-specialist communication
Your judgments must survive scrutiny from reviewers without expertise in your field; clarity and reasoning transparency matter as much as correctness.
For example: “You have judged that a model should refuse a particular request about chemical synthesis. Write the reference answer explaining your judgment in a way that an AI safety researcher without industrial hygiene training can understand and verify.”
A task you may get
Write 5-10 prompts at different risk levels in your domain, including at least one clearly benign, one dual-use boundary case, and one adversarial. Evaluate model responses and write reference answers for each, documenting your reasoning.
How to prepare
- Review your past hazard analyses or risk assessments and prepare examples of legitimate professional questions in your field
- Study your domain's regulatory frameworks: PSM, Seveso, COMAH, OSHA, or equivalents. Understand what these require and what they permit.
- Prepare 3-5 boundary-case scenarios where a legitimate question could be repurposed harmfully; practice articulating why each is a boundary case
- Review any prior red-team or policy-aligned AI evaluation work; prepare to discuss how you made consistent judgments across ambiguous cases
The facts
- Pay
- $65–75/task
- Hours
- Task based
- Where
- Remote
- Field
- Life, Physical, and Social Science
- Project name
- Neon
- Posted
- 9/14/2026
- Places left
- 100
We wrote this page from the public Mercor listing. It may be out of date, so read the full posting before you apply.