AI Safety Policy Evaluator, Violence & Threats | Remote US

This is a non-engineering content-policy evaluation role. Applicants must demonstrate relevant depth in violent fiction or media, military or emergency response, crisis or threat assessment, trust and safety, content moderation, or closely related policy work. Software engineering or LLM product experience alone is not sufficient.

ABOUT HANDSHAKE

Handshake was founded on a simple belief that everyone deserves a path to a great career, regardless of where they went to school or who they know. Today, we power 25 million job seekers, 1 million+ employers, and 1,600 educational institutions.

In 2025, we started Handshake AI and built the fastest-growing AI data business in history. We work directly with frontier AI lab researchers to create evaluations, publish benchmarks, and push the boundary of data. We've grown from $0 to ~$1B run rate and pay ~$60M to over 30K individuals every month.

Why join Handshake now

• Shape how every career evolves in the AI economy, at global scale, with impact your friends, family and peers can see and feel

• Partner hand-in-hand with world-class AI labs, Fortune 500 partners and the world's top educational institutions

• Work together with engineers, scientists, operators, and more from Palantir, Meta, Scale AI, and former YC founders

• Build a massive, fast-growing business with billions in revenue

ABOUT HANDSHAKE AI

Human data is the core infrastructure to AI advancement. Frontier AI labs currently improve model capabilities with various data-intensive post-training techniques. We believe that data spend for AI training will increase by 3-5x in the next few years and continue for much longer as models take on new domains. Handshake AI supports all of the frontier AI labs, working on their most complex data at the largest scale.

ABOUT THE ROLE

As an AI Policy Specialist on the Violence & Fiction team, you will help AI models learn where the line falls between depicting violence and enabling it.

Violence is one of the hardest domains in AI safety because most violent content is legitimate. Novels, games, screenplays, history, journalism, self-defense, and ordinary human frustration all involve violence, and a model that refuses them is broken. A model that helps someone plan real harm is worse. Your job is to tell the difference, case by case, and to explain your reasoning clearly enough that it can train a model.

You will read user requests, model responses, and conversation history, then decide which policy category applies and whether the model's response was appropriate. The interesting cases are the close ones: a torture scene that is either a chapter of a thriller or an interrogation manual with character names; a message that reads as venting about a boss or as a plan; a "realistic" combat question from a novelist that is also a real-world capability question. One word, one contextual detail, or one shift in intent changes the answer.

We are looking for people who already have strong instincts about violence in at least one of these areas: how it works in fiction, how it works in the real world, or how it shows up in people who are struggling. You do not need all three. You need one deep and the judgment to learn the rest.

This is not rote annotation. Policies cannot anticipate every edge case, and good evaluators do not apply them mechanically. You will balance policy text and intent with customer expectations, conversation context, precedent, and team calibration.

WHAT YOU WILL DO

• Evaluate user requests and AI model responses involving violence, weapons, threats, and dark fiction within the full conversation context

• Distinguish fictional, educational, historical, and defensive violence from requests that seek real-world uplift or express real intent to harm

• Assess whether a model's response gives meaningful real-world capability, regardless of how the request was framed

• Distinguish expressions of anger, frustration, or dark humor from credible threats

Responsibilities

• Evaluate user requests and AI model responses involving violence, weapons, threats, and dark fiction within the full conversation context
• Distinguish fictional, educational, historical, and defensive violence from requests that seek real-world uplift or express real intent to harm
• Assess whether a model's response gives meaningful real-world capability, regardless of how the request was framed
• Distinguish expressions of anger, frustration, or dark humor from credible threats

Requirements

• Demonstrate relevant depth in violent fiction or media, military or emergency response, crisis or threat assessment, trust and safety, content moderation, or closely related policy work