Amazon Ads Brand Safety & Suitability protects advertisers from exposure to unsafe, unsuitable, or policy-violating content across web, mobile app, CTV, and audio advertising inventory. Our mission is to ensure that every ad impression delivered through Amazon's demand-side platform appears adjacent to content that meets advertiser trust expectations while giving brands granular controls to define suitability on their own terms. We operate at the intersection of advertiser trust, publisher quality, and supply integrity.
AI is fundamentally changing the content landscape. Content is now generated at unprecedented scale — faster, cheaper, and increasingly sophisticated. Low-quality, deceptive, AI-generated, and synthetic content evolves in real time, constantly adapting to evade detection. The volume and velocity of new content entering the advertising system has outpaced traditional classification approaches.
We are looking for an Applied Scientist to work on the next generation of AI-powered Brand Safety and Content Classification systems designed to protect advertisers and elevate supply quality at internet scale. This is not a traditional classification problem. You will build systems that make millisecond-level decisions across billions of content signals while continuously adapting to emerging content risks driven by generative AI. You will own the science strategy for LLM-powered classification and semantic understanding, real-time multimodal content evaluation, adversarial ML and adaptive model resilience, proactive risk intelligence and content risk hunting, AI-generated and synthetic content detection, and large-scale abusive content system identification and disruption.
You will define how modern AI separates high-quality advertising inventory from unsafe, unsuitable, and policy-violating content — across web, mobile app, CTV, and audio surfaces.
What Makes This Role Unique
Generative AI has dramatically lowered the cost of producing deceptive, policy-evasive content, and the adversary evolves daily. Your detection systems must reason contextually, adapt rapidly, and generalize beyond previously seen content risk patterns. Static models fail here; you will build living systems that learn and respond in real time. You will do this at internet scale, developing low-latency ML and LLM-powered systems evaluating content safety, brand suitability, misinformation risk, and emerging content risk vectors across massive real-time traffic streams, making billions of decisions per day with single-digit millisecond latency constraints. This role sits at the intersection of frontier AI research and large-scale production engineering, combining deep science, system-wide impact, and business-critical outcomes. The models your team ships directly influence billions of dollars in advertising spend and the trust of the world's largest brands in Amazon DSP.
The Science Problems Are Genuinely Hard
You will tackle challenges including detecting sophisticated AI-generated and synthetic content, understanding nuanced contextual brand risk, identifying coordinated MFA space before they scale, balancing precision, recall, latency, explainability, and fairness, designing adaptive models resilient to adversarial evolution, and leveraging LLMs for semantic understanding in real-time, latency-constrained environments.
Why This Matters
Few roles offer the opportunity to work at the intersection of frontier AI, internet-scale production systems, adversarial environments, and business-critical impact — while tackling open-ended scientific challenges with real-world societal relevance. As AI reshapes the internet, the systems your team builds will define what trustworthy, high-quality digital systems look like for the next decade.
Key job responsibilities
- Own the science strategy for AI-powered brand safety classification across web, mobile app, CTV, and audio surfaces.
- Build LLM-powered content classification systems making billions of decisions/day at single-digit millisecond latency
- Develop multimodal evaluation pipelines reasoning across text, images, audio, and video in real time
- Design adaptive ML systems resilient to adversarial evolution-- continuously learning rather than relying on static models
- Build proactive risk intelligence systems that surface emerging content risk vectors through automated hunting
- Develop semantic understanding for nuanced contextual brand risk.
- Balance precision, recall, latency, explainability, and fairness at internet scale
- Define measurement frameworks and drive continuous improvement
- Translate research into production — own the path from prototype to deployed model
- Publish at peer-reviewed venues; contribute to the scientific community in adversarial ML, NLP, and content safety
:quality(80))
:quality(80))
:quality(80))