# RAIDP03-BP01 Address data that may be unsafe or inappropriate for your use case
To perpetuate dataset safety throughout the AI system lifecycle, establish definitions of safe and unsafe content for your use case. Create specific criteria for content exclusion across training, evaluation, and auxiliary datasets, considering both direct harms and contextual inappropriateness. Implement automated and human review filtering processes, with protective measures for reviewers. Document safety definitions and filtering decisions and regularly audit datasets to verify effective removal of unsafe content while maintaining necessary testing scenarios.
**Level of risk exposed if this best practice is not established:** High
## Implementation considerations
1. Define what unsafe content looks like for your specific use case by creating objective definitions that align with your release criteria.
1. Consider implementing filters and other mechanisms to filter out potentially unsafe or inappropriate content. There may be scenarios where human review is appropriate and helpful in identifying problematic content that models might miss or misclassify. Depending on your use case, seek legal guidance about whether and how to build in processes to filter training data for illegal content such as known child sexual abuse material (CSAM) or adopt additional measures to mitigate risks related to CSAM and exploitative content.
1. Implement protection systems for dataset labelers. For example, set content warnings, exposure limits, and support Resources. Create rotation schedules and anonymous reporting channels for reviewer wellbeing.
1. Measure filtering effectiveness regularly. For example, track removal rates of unsafe content while verifying preservation of necessary test scenarios.
1. Document safety decisions you make to create an audit trail of what content gets filtered out and why, so you can explain your choices and improve your process over time.
## Resources
**Related documents:**
+ [Flag harmful content using Amazon Comprehend toxicity detection](https://aws.amazon.com/blogs/machine-learning/flag-harmful-content-using-amazon-comprehend-toxicity-detection/)
+ [Trust and safety](https://docs.aws.amazon.com/comprehend/latest/dg/trust-safety.html)
+ [Automate media content filtering with AWS](https://aws.amazon.com/blogs/media/automate-media-content-filtering-with-aws/)
+ [Data-Centric Safety and Ethical Measures for Data and AI Governance](https://arxiv.org/pdf/2506.10217)
+ [AEGIS2.0: A Diverse AI Safety Dataset and Risks Taxonomy for Alignment of LLM Guardrails](https://openreview.net/pdf?id=0MvGCv35wi)
+ [BEAVERTAILS: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset](https://papers.nips.cc/paper_files/paper/2023/file/4dbb61cb68671edc4ca3712d70083b9f-Paper-Datasets_and_Benchmarks.pdf)
+ [CISA AI Data Security Guidelines - Best Practices for Securing Data Used to Train & Operate AI Systems](https://media.defense.gov/2025/May/22/2003720601/-1/-1/0/CSI_AI_DATA_SECURITY.PDF)
+ [Training curriculum on AI and data protection Fundamentals of Secure AI Systems with Personal Data](https://www.edpb.europa.eu/system/files/2025-06/spe-training-on-ai-and-data-protection-technical_en.pdf)
+ [AI Privacy Risks & Mitigations - Large Language Models (LLMs)](https://www.edpb.europa.eu/system/files/2025-04/ai-privacy-risks-and-mitigations-in-llms.pdf)
+ [Thorn Generative AI Child Safety Commitments](https://www.thorn.org/blog/generative-ai-principles/)
+ [ISO/IEC 42001:2023](https://www.iso.org/standard/42001)A.7.2 Data for development and enhancement of AI system
+ [ISO/IEC 42001:2023](https://www.iso.org/standard/42001) A.7.4 Quality of data for AI systems