KNOWLEDGE FOR LIFE Mitigating risks and maximizing benefits of generative AI in agricultural advisory GenAI for Ag Advisory Ethics Toolkit CABI’s GAIA project team KNOWLEDGE FOR LIFE GAIA phase II: Overview GAIA phase II is developing GenAI chatbots further to deliver sophisticated, accurate and timely agricultural advisory. The focus is on: a. Improving sharing and access to reliable, robust scientific data for AI training, finetuning and development b. Examining the role of model content licences to improve such access to AI-ready content c. Addressing ethical concerns in GenAI chatbots (e.g. gender and other biases, transparency, legal compliances) KNOWLEDGE FOR LIFE GenAI for agricultural advisory use cases Chatbots on common agricultural practices and queries (e.g. pest and inputs management) Weather predictive AI systems (e.g. advising on issues such as cropping patterns) Crop and soil monitoring AI systems (e.g. soil health related advisories) Image © Manx Technology Group https://manxtechgroup.com/soil-monitoring-with-iot-smart-agriculture/ KNOWLEDGE FOR LIFE GAIA Ethics Toolkit: Background • Use of AI systems, including generative AI, can pose several risks and ethical challenges • Existing frameworks are either generic or not focused on specific aspects of risk mitigation • Extant work on AI ethics has limited practical resources for the agricultural advisory use case KNOWLEDGE FOR LIFE Classification of AI risks in agriculture Data-driven risks Non-data risks – Hallucinations and reliability – Biased outputs – Access to traditional knowledge – Repetitive or redundant tech solutions – Environmental cost – Farmer autonomy (top-down adoption) – Labour impacts (e.g. employability of traditional advisory and extension officers) KNOWLEDGE FOR LIFE GAIA Ethics Toolkit: Theory of Change Problem statement: Farmers lack timely, context-specific insights. Developers face barriers (specifically around data access) to build reliable and explainable GenAI systems for agricultural advisory. Inputs • Fairly licenced content is made available to GenAI developers • GAIA Ethics Toolkit allows developers to consider and address key ethical concerns in advance Outputs • A clearly structured checklist for the different stages of the AI lifecycle (covering design, development and deployment) • A checklist that addresses risk mitigation, legal compliance and feedback mechanisms Outcomes • Developers are mindful of the risks and downsides and aim to address these adequately when designing and developing safe and reliable GenAI models • Reliability improves trust and adoption of these systems in agricultural advisory Impact • Sustainable, equitable, and data-driven agricultural systems empowered by AI- driven advisory solutions • Solutions that improve farmer livelihoods, productivity and climate resilience Assumptions: • Developers adopt the GAIA Ethics Toolkit • Farmers and institutions engage with AI advisories • Data and digital infrastructure remain accessible KNOWLEDGE FOR LIFE GAIA Ethics Toolkit: Objectives • The Ethics Toolkit aims to create prompts (questions) for developers of GenAI systems for one or more use cases in agricultural advisory • The prompts will cover the entire lifecycle of AI design, development and deployment • The prompts will target four priority areas: i. Risk identification and mitigation ii. Community and stakeholder engagement iii. Legal and regulatory compliance iv. Feedback and complaints KNOWLEDGE FOR LIFE Stages for the GAIA Ethics Toolkit Design • Defining the use case (understanding context, problem identification and proposed GenAI intervention) • Data collection or accessing existing data assets Development • Model training (building the GenAI model) • Risk assessment and pre- deployment testing and finetuning (potentially use sandboxes, where possible) Deployment • Legal compliance • Feedback loops • Complaint mechanisms • Iterative improvements to the GenAI model KNOWLEDGE FOR LIFE IDEATION African AI observatory OECD AI policy global dashboard Problem definition: FPF and GenAI playbook chapter 2 UNESCO stakeholder engagement plan (p15) Define the problem, define the use cases and users Understand the enabling environment (policies, strategies) Data asset identification (FPF step 3) Data asset ethics assessment Validate with Ag community Ready to proceed? DESIGN Develop launch criteria (ethical scoring assessment) Defining roles and responsibilities, e.g. who manages community feedback? Who checks policy updates? Metrics for datasets and models and algorithms to mitigate bias in datasets and models Red teaming Independent model audit Evaluate model and system for safety Build the model BUILD AND TEST Responsible / ethical scoring assessment before deployment DEPLOY Phased launches GenAI ethics playbook chapter 7 Post deployment governance, e.g. AI ethics committee MEL User feedback and engagement Implement plan for redressal / grievance and complaint procedures e.g. automatic triage Automated complaint management Prioritize complaints on: severity of harm, number of reports, legal implications Links back to the role and responsibilities – who is monitoring changes in the wider environment that might impact how the tool operates FINETUNE AND RISK ASSESSMENT Feedback mechanism and plan to improve (return to the build, test and deploy) NARRATIVE / EXPLIANER (INDIA / KENYA) Downloadable checklist that covers each stage of the lifecycle ELSA scan Qs ALTAI Qs KNOWLEDGE FOR LIFE Checklists of the GAIA Ethics Toolkit Questions serving as prompts for each stage, covering one or more of the four priority areas. Example: Design phase regarding ideation: • How has the problem been identified? • Have local agricultural communities and stakeholders been engaged with to define the problem statement? • Has a landscape review been undertaken to determine if existing interventions are insufficient? If yes, is this documented? KNOWLEDGE FOR LIFE Phase Question Response (Y/N) Reasoning Design (Ideation) How has the problem been identified? TBA TBA Deployment (Legal compliance) Are there any privacy laws that affect the GenAI system? Are there any legal standards or disclosures mandated under law? TBA TBA Deployment (Feedback) Is feedback collected digitally? Can feedback be submitted in other ways? TBA TBA KNOWLEDGE FOR LIFE Outcomes of the GAIA Ethics Toolkit • Create a baseline of understanding and addressing ethical concerns and risks of GenAI system (in agricultural advisory) for different developers • Embedded AI safety by design into the development of GenAI models • Improve participation of agricultural communities in the process which should aid: – Trust and confidence – Adoption and scaling KNOWLEDGE FOR LIFE CABI as an international intergovernmental not-for-profit organization, gratefully acknowledges the generous support received from our many donors, sponsors and partners. In particular we thank our Member Countries for their vital financial and strategic contributions. More information on the GAIA project can be accessed here https://www.cabi.org/projects/generative-ai-for-agriculture-advisory/ Slide 1: GenAI for Ag Advisory Ethics Toolkit Slide 2: GAIA phase II: Overview Slide 3: GenAI for agricultural advisory use cases Slide 4: GAIA Ethics Toolkit: Background Slide 5: Classification of AI risks in agriculture Slide 6: GAIA Ethics Toolkit: Theory of Change Slide 7: GAIA Ethics Toolkit: Objectives Slide 8: Stages for the GAIA Ethics Toolkit Slide 9 Slide 10: Checklists of the GAIA Ethics Toolkit Slide 11 Slide 12: Outcomes of the GAIA Ethics Toolkit Slide 13