by Vlad Cherman, Sophia Jenkinson and Isabell Orlishausen
In July, members of the SEBI-L data team attended the AI for Good Global Summit in Geneva, Switzerland. Organised by the International Telecommunication Union (ITU), together with more than 50 UN agencies and the Swiss Government, the summit brought together over 1,000 speakers, 200 exhibitors and participants from governments, UN agencies, academia, NGOs and major technology companies.
Across more than 300 sessions, the focus was on practical uses of Artificial Intelligence (AI) in areas including food security, climate resilience and humanitarian response. For SEBI-L, the summit was a chance to explore developments in AI implementation, knowledge management and decision support, particularly in low- and middle-income countries.
Trust as the leading theme
Trust came up repeatedly throughout the summit.
As synthetic information becomes easier and cheaper to produce, reliable observations and official sources are likely to become more important. AI should help strengthen these services rather than replace them, as these trusted sources often know the local contexts well and remain accountable.
Several speakers discussed “verifiable trust”: designing systems you can trust with outputs that can be checked, rather than simply trusting the model. This might involve independent machine verification using encoded rules, expert oversight, and clear audit trails. It also means documenting mistakes, avoiding “verification debt” (not budgeting for verifying AI-generated outputs), and considering whether important data sources will remain accessible and sustainable over time.
Trust, in this sense, is not an optional feature but part of the infrastructure needed for AI to scale responsibly, suggesting an approach to design for auditability, not just model accuracy. As one speaker put it, “innovations move at the speed of trust”.
Emerging Tools
The summit showcased a wide range of tools designed to bring information together, improve access to knowledge and therefore support decision-making.
Examples included NASA Harvest’s Harvest2Market, OHCHR’s HRDx, WFP and CERN’s ADAD, UN Global Pulse’s pipelines in collaboration with DISHA, and projects from Planet, Wiley, and UNDRR/WMO.
Some may offer useful data sources for SEBI-L ongoing effort of bridging existing data gaps, particularly in Earth observation. Others could provide ideas for knowledge-aggregation tools that allow users to search and navigate evidence in plain language.
Getting AI-ready
One of the clearest messages was that becoming AI-ready starts with understanding the problem, not choosing a tool.
This critically involves looking at the wider system around a challenge and asking whether the issue is really a lack of information, or whether it also involves capacity, incentives, access or power. The same problem may look very different across different actors.
Domain experts and intended users therefore need to be involved from the beginning. One speaker called for organisations to “fall in love with the problem”, rather than the solution.
SEBI-L will also explore existing frameworks that help organisations structure this process. Examples highlighted at the summit included FAO’s Digital Agriculture and AI Innovation Roadmap and the GENIE.AI open-source AI reference framework. These could provide useful reference points when considering how to assess opportunities and requirements for future AI work.
Testing and Benchmarking
Another recurring issue was how difficult it remains to assess whether an AI tool is genuinely reliable.
Practical testing and usage of AI tools continues to expose familiar problems such as inconsistent or confidently incorrect outputs, but there have been notable improvements in performance and reliability in the past year. Furthermore, while metrics and benchmarks are rarely consistent across different organisations, this is a developing area, and greater emphasis is being placed on finding ways of accurately measuring reliability.
ETSI has published standards for testing machine-learning systems and plans further work on large language models. These could provide a useful reference point for SEBI-L.
Checks can also be built into the system itself. Amazon CTO Werner Vogels described this as “trust the quorum, not the model”: relying on agreement and verification between several agents rather than a single output.
What this means for SEBI-L
The summit highlighted several areas of active development worth exploring further, including knowledge-aggregation tools and AI-enabled Earth observation. These could serve as new tools or data sources for SEBI-L usage, or ideas for our own ongoing innovation projects.
The main lesson, however, was to keep starting with the problem. That means working closely with the people and systems involved and accepting the friction that comes with this process, as well as being realistic about what technology can change.
AI is likely to be most useful when it supports domain expertise rather than replaces it. It could reduce the time spent finding, processing and organising information, leaving more time for judgement and interpretation.
SEBI-L’s relationships, contextual knowledge, and position as a trusted partner are strengths that technology alone cannot reproduce. As Peter Danenberg of Google DeepMind put it, we have a responsibility to “diligently preserve humanity, so that people still pause”. The challenge is not simply to adopt AI, but to use it in ways that preserve human judgement, attention, and trust.
Vlad Cherman, Sophia Jenkinson and Isabell Orlishausen are Data Analysts with SEBI-Livestock.