Get paid to stress-test AI models for harmful outputs, security flaws and bias for labs and enterprises.
Why that grade Rate ranges come from job-board aggregation (theinterviewguys, infosec.qa) rather than an audited operator sample. Course-seller index 1/10.
Figures are researched estimates, not guarantees. Check local rules before you trade.
As AI safety regulation and enterprise deployment risk both increase, labs and companies need people who can systematically find where a model fails before real users do — and this is genuinely technical work that most gig workers can't do, which keeps the rate high.
Reported hourly ranges roughly $59-120/hr on general job boards, with specialized/senior AI red-teaming roles reported higher; this sits above typical generalist data-labeling pay due to the specialized skill required.
Someone with a background in security research, ML, linguistics, or domain expertise in a risk area (misinformation, bias) who can document findings rigorously.
Anyone treating this as an accessible entry-level microtask — without relevant background, applicants typically get routed into much lower-paid generalist labeling work instead.
The highest-paying roles favor candidates with a real security, ML or linguistics background over generalists, so someone without that background typically gets routed into lower-paid general data-labeling work instead of the rate advertised here.
5 steps to your first paying customer
Security research, ML, linguistics, or domain expertise in a risk area (misinformation, bias) — or a portfolio of documented findings if you're newer to the field.
Done when You can point to a specific credential, work history, or a written portfolio of at least three documented findings that backs up the background you're claiming.
Mercor, Surge, specialized red-teaming vendors, or AI labs' own contractor programs.
Done when You've submitted applications to Mercor, Surge or an AI lab's contractor program and have at least one response back — accepted, rejected or interview scheduled.
The rest of the playbook: what to charge, what you need in place before you take money, where the first customers come from, and what each step costs.
Free forever · no card · 30 secondsN/A — this is skilled contract/gig work, not a business.
Deep specialization in one risk category that's genuinely scarce
A documented track record of high-quality findings across multiple clients/labs
None as a standalone asset, but this work builds credibility toward a fractional AI-safety consulting practice.
Use real professional credentials to train and evaluate AI models via Mercor or Handshake AI.
Find and responsibly disclose security vulnerabilities for cash rewards on HackerOne/Bugcrowd.
Get paid to break AI models — extract harmful outputs, bypass safety filters, find factual or security flaws.