AI Model Policy Trainer, Mental Health (Seattle)

Written by the T&S Network from Handshake's posting.

Evaluates how conversational AI systems respond to users discussing emotional distress, mental-health conditions, unusual beliefs, and possible crises. The work is rating model responses and tool actions against a detailed mental-health safety policy, with trainers calibrating to keep judgments consistent. The posting places the role within Handshake AI, which works with frontier AI lab researchers on evaluations and model improvement.

• Self harm and suicide

What you'd do

• Review user conversations and AI responses on mental health, crisis, paranoia, hallucinations, and mania
• Apply detailed policy taxonomies to classify user requests and model behavior across single messages and longer conversations
• Separate validating a person's emotions from validating an unsupported or harmful explanation, and flag reinforced or escalated beliefs
• Judge whether the AI's tool actions, from searches to purchases and posting, were appropriate
• Write rationales tying each decision to conversation evidence and policy
• Take part in calibration exercises and give evidence-based peer feedback to other trainers
• Flag ambiguous examples, gaps, and edge cases that need policy clarification

What it asks for

• Must have prior hands-on work helping individuals or families through mental or behavioral health struggles
• Background in clinical, social service, caregiving, peer support, advocacy, or crisis roles
• Clinical licensure valued but not required
• Familiarity with conversational AI tools (preferred)
• Experience with rubric-based evaluation, case documentation, triage, or quality assurance (preferred)
• Experience working with people in crisis, psychosis, mania, or suicidal ideation (preferred)
• Knowledge of how cultural background, language, disability, and access to care shape mental-health talk (preferred)

Worth knowing

• Regular exposure to sensitive material including suicide, self-harm, abuse, violence, and psychosis
• Support resources available for exposure to distressing content
• Full-time, non-exempt schedule
• Employed through TCWGlobal, a W-2 vendor assignment supporting Handshake AI

Our best summary of the posting, and it can be wrong. The employer's posting is what counts: read it in full on Handshakenorth_east before you apply. If we got something wrong, report this listing.