In this Humanitarian AI Today episode, Michael Tjalve, Independent Humanitarian AI Advisor and co-author of the SAFE AI framework, speaks with guest host Jennifer Wilde, an AI strategy and social impact advisor, about AI governance, the SAFE AI initiative and the evolving security and operational landscapes. Against the backdrop of critical, sector-wide dialogues bringing together executive leadership and technical teams from organizations of all sizes, they explore the new safety and security topics of concern posed by advances in AI, compute distributed across agents, and the emergence of sophisticated models like Mithos that can expose and be used to exploit critical IT security vulnerabilities. Michael outlines his advisory work helping humanitarian organizations understand how they can confidently use AI in a way that is both effective and responsible, emphasizing the increasingly present role that artificial intelligence has in humanitarian action today. Jennifer and Michael discuss the SAFE AI framework next, the Standards and Assurance Framework for Ethical AI, which stands as the first operational, end-to-end AI governance framework tailored specifically for humanitarian actors. Developed through a vital collaboration between the CDAC Network, The Alan Turing Institute, and the Humanitarian AI Advisory, with critical funding from the UK Government’s Foreign, Commonwealth and Development Office (FCDO) and deep consultation with East African researchers and communities, the toolkit bridges the gap between high-level policy and grassroots implementation. They walk through how both large and small organizations can start utilizing the toolkit today, what is essential to know to begin, and exactly where to start. Pointing out how organizational governance is lagging behind organic, "shadow IT" AI experimentation, Michael and Jennifer discuss in detail why structured guardrails are vital for the sector. They address how crucial it is for team members and staff across entire organizations, whether at global headquarters or country offices, to have access to appropriate training and a centralized focus on AI policy and strategy, highlighting how the SAFE AI framework provides actionable guidance to achieve this. Touching on rapid technological leaps, Michael notes how the emergence of agentic AI introduces an additional layer of complexity for responsible governance. With semi-autonomous processes executing tasks on an organization's behalf, the potential for compounding errors increases while transparency declines, a distance that challenges traditional accountability models and complicates oversight. Together, Michael and Jennifer advocate for starting any AI adoption journey with a clear focus on the specific human need or use case rather than deploying the technology for its own sake. Drawing on the core tenets of the framework, co-authored by Michael alongside Suzy Madigan, Helen McElhinney, Sarah Spencer, and Anjali Mazumder, they emphasize that true accountability requires an operational commitment to "communities in the loop". This means ensuring that frontline staff can confidently explain AI-driven outcomes to affected populations, and that AI systems never silently or inequitably exclude the very people humanitarians are bound to protect. Ultimately, they agree that while the rapid expansion of AI is exciting, organizations must maintain a healthy balance of curiosity and caution, allowing their concerns to drive continuous learning and the proactive questioning necessary to confidently take responsible steps forward. Because there is absolutely no margin to get it wrong in a humanitarian crisis, this collective effort reminds listeners that stress-testing AI governance under extreme, high-stakes conditions does not just safeguard vulnerable populations; it creates a vital proof of concept with profound lessons for how responsible AI should be governed anywhere in the world.
