Sustainability Challenges in Humanitarian AI: From Theory to Practice

Sustainability Challenges in Humanitarian AI: From Theory to Practice
Thomas Byrnes
Thomas Byrnes
Humanitarian & Digital Social Protection Expert | Director of MarketImpact

November 5, 2024
At a recent AI for Development Conference, I had the privilege of attending an eye-opening presentation by Emily Springer, PhD, about the Aunt Zuzi chatbot project in South Africa. Her presentation offered three crucial insights that resonated deeply with my own work in humanitarian AI:

"AI for good must be proven, not assumed" - This powerful statement challenged the room's comfortable assumptions about social impact AI.
The critical importance of building AI systems from the community up, rather than technology down.
The often-overlooked challenge of compute costs in humanitarian AI projects - a factor that can threaten the sustainability of otherwise well-designed initiatives.

These insights from Dr. Springer's work with GRIT's Aunt Zuzi project highlight the gap between AI's promise and the practical challenges of implementation in humanitarian contexts.

Article content
Emily Springer at AI for Development
Our NRC Ukraine Experience
At MarketImpact, we recently developed an AI-powered decision support system for the Norwegian Refugee Council in Ukraine to help field teams make rapid decisions about aid delivery in conflict zones. Having worked in humanitarian response for over 15 years, I've seen firsthand how operational costs can kill otherwise promising projects.

We made a radical decision early in the design phase: the entire system needed to run on just $100 worth of compute credits for a full year. This wasn't arbitrary - it was based on the reality that while humanitarian organizations can often secure funding for development, ongoing operational costs are a different story.

This constraint drove every aspect of our design:

We opted for Google's Gemini Flash at 15 cents per million tokens, rather than GPT-4 at $15 per million tokens
We eliminated free-form text inputs in favor of structured questionnaires
We leveraged NRC's existing Microsoft/Google enterprise agreements to avoid additional compliance costs
Every feature was evaluated not just for its utility, but for its token cost

Interestingly, these constraints improved our design. When you know you have limited compute resources, you focus ruthlessly on what's essential. Our system isn't chatty or conversational - it doesn't need to be. It does exactly what field teams need: help them make faster, more consistent decisions about aid delivery.

The Broader Implications
This crisis highlights fundamental challenges in humanitarian AI:

Funding Structures

Development grants rarely account for operational costs
Traditional humanitarian funding cycles don't align with AI operational needs
Need for new funding models that consider ongoing compute costs

Ethical Standards

Current AI ethics frameworks don't adequately address resource constraints
Need to define minimum viable standards for trauma-informed AI
Question of who bears responsibility for sustainable funding

Technical Solutions

Need for more efficient models that maintain quality
Potential for hybrid approaches combining AI with other resources
Innovation in reducing compute costs without compromising care

When the choice is between providing inadequate support and no support at all, what is the ethical path forward? Their story suggests we need fundamental changes in how we fund, develop, and deploy AI in humanitarian contexts.

Moving Forward
The humanitarian sector needs new approaches to AI development and funding that acknowledge these realities:

New funding models that include operational costs
Better understanding among donors about compute requirements
Design approaches that prioritize sustainability from day one
Realistic conversations about what features are truly essential

Dr. Springer's work with the Aunt Zuzi project is helping push these important conversations forward, highlighting the critical need to think about sustainability from the outset of AI humanitarian projects.

A Call to Action
For Humanitarian Organizations:

Start with sustainability in your AI design processes
Budget for compute costs from day one
Consider hybrid approaches that balance AI capabilities with resource constraints

For Donors:

Recognize compute costs as essential operational expenses
Develop new funding models that support ongoing AI operations
Invest in sustainable, long-term solutions rather than just pilot projects

For Technologists:

Focus on developing more efficient models for humanitarian contexts
Create solutions that balance capability with sustainability
Build with resource constraints as a primary design consideration

Join the Conversation
I encourage everyone interested in ethical AI development to:

Follow the groundbreaking work of the GRIT team and their Aunt Zuzi project. You can learn more at (1) GRIT - Gender Rights in Tech (Formerly Kwanele South Africa): Posts | LinkedIn
Connect with Emily Springer, whose research and insights into ethical AI implementation are pushing our sector to confront these critical challenges. Her work can be found at (3) Emily Springer, PhD | LinkedIn
Join the ongoing discussion about sustainable humanitarian AI. Share your experiences, challenges, and solutions on the aidgpt.org discourse server, where practitioners are working together to build more sustainable approaches.

Final Thoughts
Emily Springer's presentation reminded us that "AI for good must be proven, not assumed." But perhaps more importantly, it must be sustainable to be truly good. The future of humanitarian AI lies not in pushing technical boundaries without regard for costs, but in creating sustainable solutions that genuinely serve vulnerable populations.

As we continue this journey, let's ensure we're building AI systems that don't just work in theory, but work in practice - and keep working when our communities need them most.

Thomas Byrnes is the founder of Market Impact, a consulting firm specializing in humanitarian assistance and social protection. Connect with him on LinkedIn or join the conversation at aidgpt.org