"The SAFE AI Framework is designed to support humanitarian organisations on their AI journey, and navigate the risks raised. Organisations procuring AI systems or designing AI projects will benefit from each tool within this framework, which is intended as a practical approach to using, designing, p ... rocuring, deploying, or scaling AI-enabled systems in humanitarian action. It helps humanitarian actors identify, reduce, and document socio-technical risks across the AI journey, enabling proportionate innovation aligned with humanitarian principles, the Core Humanitarian Standard, and IASC – the Interagency Standing Committee Accountability to Affected People commitments. On a day-to-day basis, the SAFE AI Framework is designed to support senior operational and governance decision-makers within humanitarian organisations. The outputs of SAFE AI processes are intended to provide a defensible basis for decisions on whether an AI system should proceed, be modified, be paused, or be decommissioned." (Introduction, pages 7-8) more

"Based on the National Institute of Standards and Technology (NIST) AI Risk Management Framework, the seven characteristics of trustworthy AI are: 1. Valid and Reliable: The AI system should perform as intended, delivering accurate results consistently across different situations. 2. Safe: The sy ... stem must not produce harmful or dangerous outcomes to human life, health, property, or the environment. 3. Secure and Resilient: The AI must withstand unforeseen events, attacks, or adversarial inputs, maintaining its functionality without compromising data. 4. Accountable and Transparent: There should be a clear understanding of how the system operates, including documentation on data, decisions, and responsibility for its outputs. 5. Explainable and Interpretable: Users should understand the reasoning behind a particular output or decision made by the AI. 6. Privacy-Enhanced: The system must protect sensitive data, uphold privacy standards, and avoid leaking information across sessions. 7. Fair with Harmful Bias Managed: The AI must treat different user groups equitably, actively identifying and mitigating bias in training data or results." (Page 1) more