What the work actually involves

You write adversarial prompts in English and Urdu, run them against conversational models and agents, and record what breaks. In practice that means multi-turn manipulation attempts, jailbreak framing, prompt injection through pasted content, role-play pressure, and probes aimed at bias or misinformation — then annotating which guardrail failed and why. A large share of the value in a bilingual seat comes from attacks that only work in Urdu or in code-switched English-Urdu: transliteration, Nastaʿlīq versus Roman script, idiom and honorifics, regionally specific misinformation narratives that an English-only tester would never think to try.

The second half of the job is documentation. Every finding needs to be reproducible by someone else: the exact prompt chain, the model response, a classification against the project taxonomy, a severity judgement, and often an English gloss of Urdu-language content so reviewers who do not read Urdu can assess it. Work is delivered as datasets, attack cases, and short reports customers can act on.

What the screening looks for

Mercor's process is largely AI-led and leans on verifiable specifics. Expect probes on prior red teaming or adversarial work — AI-side, cybersecurity, or socio-technical abuse research — and on whether you work from frameworks and taxonomies rather than improvised one-off tricks. Urdu fluency is tested substantively, not asserted: register control, script handling, and the ability to explain why a particular Urdu phrasing evades a filter. Follow-up questions go one or two levels deeper than the initial answer, so vague claims tend to collapse.

Logistics

  • Fully remote and asynchronous; work is claimed in batches rather than assigned shifts.
  • Hours are flexible but uneven — availability of 15–20+ hrs/week makes you far more useful to project staffing.
  • $20–22/hr is the band observed on this listing, not a guarantee; rates vary by project and customer.
  • Content touching bias, misinformation, harassment or harmful behaviour is disclosed before exposure. Higher-sensitivity tracks are optional and come with guidelines and wellness resources.