Mercor is recruiting native English and Gujarati speakers to red team conversational AI models — writing adversarial prompts, documenting jailbreaks and bias exploits, and annotating where models fail. Work is remote, text-based, and paid at an observed $20–22/hr, with exposure to sensitive content disclosed in advance.

Summary, eligibility guidance and screening preparation are prepared by Sidequest from the platform's posting. Check the source listing before applying.

What the work actually involves

You write attacks and you write them down. A typical shift means taking a taxonomy of harm categories — misinformation, bias exploitation, harassment, misuse instructions — and constructing prompts in Gujarati, English, or a mix of both that push a model past its guardrails. Multi-turn manipulation matters more than single clever prompts: building rapport across turns, reframing a refused request, using code-switching or transliterated Gujarati to route around filters trained mostly on English. When something breaks, you classify the failure against the project's taxonomy, capture the exact turn sequence so someone else can reproduce it, and note whether it looks like a one-off or a systemic gap.

The Gujarati requirement is load-bearing, not decorative. Much of the value in multilingual red teaming comes from cultural and linguistic specificity — caste and communal framings, regional political misinformation, honorific registers, Gujarati script versus Roman transliteration. Reviewers are looking for someone who can judge whether a model's Gujarati output is merely awkward or actually harmful in a way an English-only tester would never flag.

What the screen looks for

Mercor's process leans on an AI-led interview plus recorded work samples. Expect probing on prior adversarial work — AI red teaming, penetration testing, trust and safety investigations, abuse analysis — and expect follow-ups that ask you to name the technique, not the outcome. Vague claims collapse fast. You should be able to describe a specific jailbreak family, why it worked, and how you'd document it for a customer engineer who does not read Gujarati. Native fluency is verified through the interview itself, sometimes in both languages.

Logistics

Fully remote and largely asynchronous; work is pulled from task queues rather than scheduled shifts
Text-only — no audio, video, or imagery
Sensitive-content projects are opt-in, with topics disclosed before exposure and wellness resources provided
Hours are flexible but availability tends to matter: reviewers favour candidates who can commit consistent weekly hours across shifting projects
Observed pay band $20–22/hr; rates vary by project and are not guaranteed

AI Safety Experts — English & Gujarati

What the work actually involves

What the screen looks for

Logistics

Interview prep — 10 questions