How to Build Trust in AI

Amazon treats RAI (Responsible AI) not as an add-on feature but as a fundamental part of product design, with science and policy teams working together to anticipate risks in advance, handle ambiguity, and build systems that adapt to change. To this end, Amazon operates a pipeline spanning pretraining, post-training, evaluation, and third-party monitoring, and has accumulated 70+ internal and external tools, 500+ research papers, and tens of thousands of hours of training.

• In pretraining, data and learning tasks embodying principles of safety, security, and fairness are added so the model understands policy.
• Harmful content isn't simply removed; instead, it's re-trained with added context so the model learns to recognize risk.
• Amazon researches modality alignment to cover not just text but also images and audio, and checks whether learning occurred using perplexity and sparse questions.

In the RLHF stage, human preferences and policy fit are converted into rewards to align the model's judgment criteria. An auxiliary-reward model built from human comparisons of responses, together with an independent LLM judge that scores responses against rubrics, are used together to reinforce responses that adhere to policy.

The evaluation team checks directional signals with lightweight benchmarks during training, and once training is complete, compares saved checkpoints against broad test sets to find the optimal point. They create model-breaking datasets designed to induce policy violations, and verify models across 8 axes: privacy and security, safety, fairness, veracity and robustness, explainability, controllability, governance, and transparency. Data is collected from red teamers, external security partners, publicly available university benchmarks, and social media, alongside research into detecting deception in long-term interactions and automated red-teaming.

The most sensitive frontier risks include research on CBRN attacks and misuse for cyberattacks. Amazon identifies risks through automated benchmarks and external expert review, and for trusted researchers, is examining limited access with heavy monitoring and LoRA-based configurability.