AI Safety Experts: English & Kannada

What you'll do

The task list is the same red-team brief Mercor uses across this family: probe conversational models and agents with jailbreaks, prompt injections, misuse cases, bias exploitation and multi-turn manipulation; annotate and classify the failures; work to taxonomies, benchmarks and playbooks; and write reproducible reports and attack cases.

What differs for Kannada is who Mercor says it wants. The European and Southeast Asian versions of this listing describe the ideal candidate as someone with prior red-teaming experience. The Kannada version, like the other Indian-language versions, drops that line and instead asks for strong judgment about language and content: telling whether an AI response is accurate, complete and appropriate, and explaining why. In practice that reads as a mix of evaluation and red-teaming, and it opens the door to careful native speakers without a security background.

Kannada is a good test of a model's limits. It is spoken by tens of millions of people in Karnataka but has a small digital footprint relative to its speaker base, uses its own script, and is often typed in Latin transliteration or mixed with English on phones. Each of those forms is a place where safety behaviour can break.

All work is text. Sensitive topics (bias, misinformation, harmful behaviour) are part of it; topics are communicated in advance, and higher-sensitivity projects are optional with guidelines and wellness resources.

Who fits

• Native fluency in Kannada and English
• Sharp judgment on whether an answer is correct, complete and suitable, with the ability to say why
• Rigour: you notice small errors and inconsistencies others skim past
• Consistency when working to guidelines and quality standards
• Clear written reasoning, and comfort moving between task types and clients

Nice to have: adversarial ML, cybersecurity (pentesting, reverse engineering), harassment and disinformation analysis, or creative backgrounds in psychology, acting or writing. No degree or minimum years are stated.

What it pays

$16–22 per hour, Mercor's published band, shared with the Malayalam, Tamil, Telugu, Marathi, Odia, Gujarati and Urdu variants. It is the lowest band in this listing family (the Nordic and Dutch versions pay $48–62), but the entry bar is also lower.

The ad states weekly payment via Stripe or Wise, based on services rendered. H-1B and STEM OPT candidates are not supported.

Worth knowing

Good:

• No prior red-teaming experience listed in the core profile, so strong native reviewers can qualify
• US-dollar pay, weekly
• No residence requirement published
• Sensitive projects are opt-in, with topics disclosed first

Less good:

• The task list still describes adversarial work; expect screening to test how you would try to break a model
• No hires count shown when we read the listing, so volume is uncertain
• Disturbing content is part of the job
• No weekly hours; projects can end early
• Independent contractor, clients unnamed

About this listing

Posted by Mercor as a remote hourly contract, read from the Mercor site on 24 September 2026. Pay, requirements, payment terms and the visa exclusion are the ad's own. Mercor states the work involves no confidential information from any employer, client or institution. See Mercor.