AI Safety Experts: English & Marathi

What you'll do

Marathi shares the Devanagari script with Hindi, and that is exactly where AI models tend to go wrong. With far more Hindi than Marathi in their training data, models often drift into Hindi vocabulary, grammar or cultural assumptions while claiming to answer in Marathi. A native Marathi reader spots that immediately; an automated check usually does not. The same blind spots are where safety behaviour weakens.

Mercor's ad lists red-team tasks:

• Probe chat models and agents using jailbreaks, prompt injections, misuse cases, bias exploitation and multi-turn manipulation
• Annotate failures, classify vulnerabilities and flag systemic risks
• Follow taxonomies, benchmarks and playbooks
• Deliver reproducible reports, datasets and attack cases

The "who you are" section for Marathi is evaluation-led: strong judgment about whether an AI response is accurate, complete and appropriate, a habit of catching subtle mistakes, consistent use of guidelines, and clear explanations. Prior red-teaming, which the European versions of this listing expect, is not in the core profile here.

All work is text. Content covers sensitive areas such as bias, misinformation and harmful behaviour; topics are disclosed up front, and higher-sensitivity projects are optional with guidelines and wellness resources.

Who fits

• Native fluency in Marathi and English
• A good ear for when an answer is subtly wrong, incomplete or inappropriate, and the words to explain why
• Attention to detail and consistency with quality standards
• Clear written communication
• Comfort switching between task types and clients

Pluses: adversarial ML, cybersecurity, harassment and disinformation analysis, psychology, acting or writing. No degree or minimum years are stated.

What it pays

$16–22 per hour, Mercor's published band, the same as the other Indian-language variants and Urdu. The ad states weekly payment via Stripe or Wise. H-1B and STEM OPT candidates are not supported.

Worth knowing

Good:

• Evaluation-focused entry bar, so careful native speakers from editing, teaching or translation backgrounds can compete
• Weekly US-dollar payouts
• No residence requirement published
• Sensitive work is opt-in, with topics announced first

Less good:

• No hires count on the listing when we read it, so demand is unclear
• Lowest pay band in this family, well below the $48–62 European rate
• Red-team content can be distressing
• No weekly hours; projects can be extended, shortened or ended early
• Contractor terms; clients and models are not named

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.