Custom LLMs and collaborative datasets: does the aid sector need AI?

As the international aid sector begins to consider the role of artificial intelligence in its work, Oxfam International has revealed that it will be developing a custom large-language model (LLM) for the INGO sector in 2026 and 2027. 

“It is going to be developed, supported, and delivered out of the Global South,” said Kenny Kamal, chief information officer at Oxfam International, who spoke from the UK. 

“If you look at all of the LLMs today, they’re all delivered by Western entities.”

While the ambition is to host and run this new LLM out of Kenya, Kamal also acknowledged that Oxfam will likely have to contend with infrastructure constraints. 

This means that collaboration between aid organizations will be key to this project becoming a reality, he added. 

“Oxfam is very collaborative with the wider INGO space globally,” Kamal said. “So if we build something, we’re not going to just build it for us.”

The humanitarian aid sector has already started collaborating on data and information sharing. For instance, HumSet is a dataset that combines multilingual information about different humanitarian responses between 2018 and 2021.  

“HumSet, in contrast to the current resources which mostly originated from social media, is created with humanitarian experts through an annotation process on official documents and news from the most recognized humanitarian agencies,” explains the team behind the dataset. 

The dataset is made up of information from 46 aid organizations around the world. 

Ensuring the aid sector is AI – and risk– ready

Alongside conceptualizing an LLM for the NGO sector, Kamal chairs Oxfam’s global AI working group. 

His team has deployed AI training to everybody across the confederation of national aid-delivering agencies that make up Oxfam. 

According to Kamal, one “uncomfortable truth” the team had to reckon with was that despite their wealth of data, it all remained disconnected, making it challenging to use it to train artificial intelligence tools.  

Oxfam International is not the only major aid organization considering where artificial intelligence can best slot into their work. Others, such as the Norwegian Refugee Council, are using AI to improve how information is shared internally, including “messy” unstructured data and images.  

At Mercy Corps, there is a renewed focus on collecting geospatial and satellite image data, said Alicia Morrison, director of data science. 

They also considered using machine learning to forecast the price of basic groceries, particularly in contexts with inflationary pressures and where the organization was giving out cash assistance, Morrison added. 

“It’s not that difficult in a lot of locations [but] it is more complicated in other places, particularly once you get into really volatile places,” she said. “You have to really think hard about whether or not it’s the right thing to do.”

Roy Hanna, director of data and analytics at Save The Children International, added that AI had recently become a part of his team’s remit. They are particularly interested in what they can learn from existing qualitative data about their past programs and results. But building machine learning algorithms from this information has its limitations, he said. 

“It’s really hard to say just based on an absolute evaluation metric how successful the program was, because context varies so much,” Hanna said. “This is another piece where you really need the human in the loop to make those assessments.

“And we don’t want an AI tool to design a program – we want the AI tool to help find the information, bring it together and help the person who is designing the program.”

Ideally, choices about data and technology are made by in-country experts who are delivering specific programs on the ground, Hanna added.

Aid cannot be automated 

Save the Children has developed a risk framework for using generative AI within the organization and for frontline work. Practitioners are encouraged to consider the risk of outputs based on flawed data and how they might harm the children they work with. 

Other risks include cybersecurity, data privacy and potential non-compliance with legal and regulatory obligations. 

Environmental impact and labour exploitation—negative externalities of the AI industry that the organization cannot control—are also factored into the risk framework.  

Risk remains front-and-centre for each organization that Future of Good spoke with. Most still expressed discomfort with deploying AI tools that can be directly accessed by the communities they serve, focusing instead on how AI can support internal operations and frontline staff. 

“Once you get into participant-facing technologies, there is quite a high risk,” said Morrison. 

Kamal felt it was important to acknowledge the limits of AI’s “practical help” in program delivery.

“The fundamentals of the work that we do does not change.”

For Zineb Bhaby, who shifted from leading data solutions to AI solutions at the Norwegian Refugee Council earlier this year, the sector has yet to balance risk and opportunity. 

“Participant-facing AI is on hold because we don’t have the capacity to do it properly,” she said. 

“But I really think the humanitarian sector needs to pool resources together so we can fully leverage the value of AI. We cannot always be too risk-averse.”

Technology vendors become AI funders

AI programs are proving to be an attractive granting opportunity for technology-forward philanthropic organizations, such as the McGovern Foundation, and technology vendors themselves. 

In a recent report, Project Evident – a U.S.-based organization helping the social and education sectors prioritize data and evidence – found that there is still a significant gap in grant funding for AI projects among philanthropic funders. 

“Grantmakers should consider funding that allows flexibility and innovation so that the social and education sectors can experiment with approaches,” they write in an executive summary. 

“Most importantly, funders should increase their capacity and confidence in assessing AI implementation requests along both technical and ethical criteria.”

Both Morrison and Hanna found that large technology companies were willing to fund AI development pro bono or as a gift-in-kind. In other words, they could provide some access to their products or software developer time. 

The cost of products, such as Microsoft Copilot licenses, remained prohibitively high despite non-profit rates, Hanna and Bhaby said. Collaborating with smaller start-ups also felt risky “because they may disappear in a year,” Hanna said. 

Kamal insisted that Oxfam International should “always maintain a core level of expertise within [its] teams.” 

“You can’t outsource everything to a vendor. You will undoubtedly pay more,” he said. “But that knowledge is not then in your hands. If you want to change vendors or platforms, you’re relying on the vendors to do that.”

Tell us this made you smarter | Contact us | Report error