Jawoo Koo

Asking the right question: Toward a research agenda for responsible GAI in agricultural extension

This study explores how generative AI (GAI) tools for agricultural extension can be designed and evaluated more responsibly. While current GAI systems offer scalable, personalized advice, they often ignore the lived realities of smallholder farmers—especially women—by relying on generic datasets and rigid evaluation metrics. We investigate three complementary methods: adversarial testing to expose gendered and contextual blind spots in model outputs; deliberative stakeholder engagement using the C-H-A-T framework, which focused on Collective knowledge, Human insight, Augmentation, and Trust, to surface value tensions and design trade-offs; and field-level insights from extension officers to uncover trust-building, diagnostic reasoning, and social intelligence absent from static GAI interactions. Together, these approaches reveal that responsible GAI requires more than technical accuracy. It demands participatory design processes that foreground user realities, surface stakeholder assumptions, and account for social and institutional context. We recommend developing gender-responsive benchmarks, embedding reflexive, participatory design methods, and modeling advisory reasoning based on real-world extension practice. The findings contribute to a growing agenda for responsible AI development—highlighting the importance of aligning GAI tools not only with technical goals, but with the social, cultural, and political contexts in which they operate.

Man vs. machine: Multi-country experimental evidence on the quality and perceptions of AI-generated research blog content

Academic research is not always available in a form that is accessible or engaging to a non-academic audience, hindering readers’ engagement with it. Non-academics, even if highly educated and policy experts in their fields, tend to need research to be presented in a more accessible way than peer-reviewed articles — one example being non-technical blogs. However, writing these requires some effort from researchers. Artificial Intelligence (AI) tools can make academic research easier to understand by summarizing and simplifying academic papers much more quickly than researchers can, making it easier for researchers to produce such summaries. However, disclosure of AI use may lower readers’ perceived quality of and trust in the blog, generating a trade-off for the researcher. In this paper, we evaluate an 11-country experiment cross-randomizing a blog’s actual and reported author as AI or human. We find that research stakeholders rate the quality of AI-generated blogs marginally lower than human-written ones (p 0.1), but disclosure of AI use offsets the negative effect (p 0.1). The study sample consists of policy-relevant stakeholders who typically engage with academic research; they are highly educated and include thematic specialists. Indeed, findings indicate that this audience interprets “accessibility” differently, preferring slightly more technical summaries of research. The nature of the respondents may thus explain the particular findings in this study, suggesting that researchers should tailor their prompts for their intended audience. There are no effects on readers’ reported likelihood of engaging with the blog or on beliefs about others predicted engagement with it. Consequently, we hypothesize that researchers can leverage AI to communicate their research more easily without a penalty from disclosing its use.

Data collaborative report: Session 1