Senior Product Manager, AI Safety Evaluation & Governance - TikTok Safety Product

The Feed Safety-Model & Data Intelligence team within TikTok Platform Responsibility ensures that AI models meet the highest bar before they make content safety decisions affecting billions of users. Our work spans three layers: - Standards & Governance — We define and iterate the safety standards that AI systems must follow, translating complex policy intent into structured, machine-interpretable frameworks. This requires deep governance thinking: navigating trade-offs between safety, fairness, user experience, and enforcement consistency. - AI/ML Solution Design — We partner closely with algorithm teams to improve model accuracy and stability across safety scenarios, tackling challenges unique to this domain — adversarial content, imbalanced distributions, and deep contextual understanding. We evaluate, select, and help shape the right AI approaches (LLMs, prompting strategies, agentic workflows, etc.) for each problem. - Rigorous Evaluation — We design statistically grounded evaluation frameworks, build high-quality ground truth datasets, and ensure our assessments are valid, reproducible, and actionable — so the platform can confidently ship AI-powered safety systems at scale. Our work sits at the intersection of AI/ML product development, trust & safety policy, and data-driven quality assurance — ensuring that AI systems can be reliably deployed for high-precision content review and risk governance at scale.

Responsibilities - Define and own the evaluation methodology for AI-powered safety models — establish frameworks for measuring continuous recall capability across risk severity tiers; define launch criteria, regression thresholds, and ongoing monitoring requirements so that no model ships or degrades without clear, evidence-based quality signals. - Design and build diverse positive-example evaluation datasets — develop principled labeling taxonomies and sampling strategies that maximize coverage of real-world content diversity (across languages, formats, content types, and adversarial patterns), leveraging LLMs as tools to surface gaps and expand coverage systematically. Define the methodology and coordinate labeling teams for execution. - Architect severity-stratified and ranking-aware evaluation benchmarks — create tiered datasets aligned with risk severity levels (e.g., critical / high / medium / low) to rigorously assess model performance at each tier, enabling differentiated quality gates, calibrated decision thresholds, and informing recommendation strategies on how to rank and distribute content by risk level. - Conduct hands-on data analysis and model evaluation — analyze model outputs, compute statistical metrics (precision, recall, F1, confidence intervals, regression analysis), identify failure patterns, and generate actionable insights for algorithm partners. - Develop and iterate prompts and LLM-based evaluation pipelines — independently author, tune, and optimize prompts for LLM-as-judge and LLM-assisted labeling workflows; diagnose prompt failure modes and drive continuous improvement. - Drive cross-functional alignment — collaborate with Algorithm, Policy, Recommendation, Labeling, and Data Science teams to translate evaluation findings into model improvement roadmaps, policy refinements, and operational calibration. - Scale and systematize evaluation operations — identify process gaps, build reusable tooling and frameworks, mentor junior team members, and ensure the evaluation system evolves alongside model and policy complexity.