Researchers test whether AI can speed up inclusive-design evaluations for farming tools in poor countries

A research team compared AI language models against human experts evaluating whether agricultural software tools work for people with low literacy and slow internet. The AI matched human judgment on some questions but not others, suggesting language models could eventually save months of evaluation time — but only if you trust the specific model on the specific tool.

Agricultural development organizations currently hire teams to audit whether farming software actually works for smallholders in Africa and South Asia — a process that takes months and costs real money. If language models can approximate that work, it collapses the evaluation timeline and cost, which means more tools get audited before deployment, or the same budget audits more tools. The catch is straightforward: the models are inconsistent. They get some dimensions of inclusiveness right and others wrong, which means you cannot simply swap a human team for a prompt. This is the actual hard problem in using AI for specialized judgment work — it is not whether the AI is smart enough, it is whether its failures are predictable enough to notice before they matter.