$18–22/hr · Mercor · Part time, 7 hours a week
A bilingual Thai-English professional evaluates AI model safety by writing Thai prompts across sensitive subjects and identifying adversarial and escalation patterns.
What you would do
- Write expert-level Thai prompts across sensitive and complex topics to test model robustness
- Classify conversations and prompts using structured evaluation guidelines
- Flag adversarial phrasings where users attempt to manipulate or bypass safety guidelines
- Identify and document escalation patterns showing progressive boundary-testing
- Document your reasoning for all safety judgments and classifications
Who they want
- Native or near-native Thai fluency with business-level written English capability
- Bachelor's degree (completed or currently in progress)
- Strong attention to detail and clear written reasoning ability
- Sound judgment about sensitive information and potential harms
- Preferred: experience with content review, policy evaluation, or adversarial testing frameworks
Main skills
What the interview asks about
1.Adversarial prompt technique recognition
Safety depends on catching sophisticated manipulation attempts; identifying subtle escalation helps engineers strengthen AI defenses before harmful use develops
For example: “A user asks straightforward questions, then adds framing in follow-ups that progressively normalize requests. How would you characterize this escalation pattern versus a single problematic question?”
2.Thai cultural context judgment
Thai cultural norms shape what constitutes appropriate AI behavior; evaluators must apply local understanding to judge whether responses respect community values
For example: “The model provides accurate information but uses phrasing that could offend Thai readers. Would you flag this as problematic, and what improvement would you suggest?”
3.Escalation sequence analysis
Recognizing systematic patterns in how users test boundaries reveals which AI defenses need strengthening; clear documentation shapes engineering priorities
For example: “Across conversations, users begin with topic A, shift to topic B, then attempt topic C commitments. How would you present this progression to distinguish it from unrelated attempts?”
4.Prompt design for edge cases
Creating prompts that expose model weaknesses in sensitive domains generates training data that directly improves safety; your prompts become part of the model improvement loop
For example: “You're designing Thai prompts testing how models handle mental health combined with cultural beliefs. What specific variations would you create to expose gaps?”
5.Guideline ambiguity resolution
Classification rules sometimes conflict with complex real-world situations; documenting these disagreements helps product teams refine guidelines for consistency
For example: “A conversation spans multiple sensitive domains where classification guidelines conflict. Your judgment differs from written guidance. How would you report this discrepancy?”
A task you may get
Given 4-5 Thai-language conversation samples with embedded adversarial techniques and escalation patterns, classify each using provided guidelines, identify specific problematic phrasing, and document your reasoning for each decision.
How to prepare
- Study how content moderation teams categorize sensitive topics and why consistent taxonomy matters for AI training
- Gather examples of subtle prompt manipulation and social engineering techniques used to test AI guardrails
- Research AI safety frameworks and which types of model failures are prioritized for fixing
- Prepare examples of Thai cultural nuances where international safety guidelines might not directly apply
The facts
- Pay
- $18–22/hr
- Hours
- Part time, 7 hours a week
- Where
- Remote · Remote — Southeast Asia preferred
- Posted
- 9/4/2026
We wrote this page from the public Mercor listing. It may be out of date, so read the full posting before you apply.