FOSTERING RESPONSIBLE INNOVATION AND GOVERNANCE OF BIG DATA AND ARTIFICIAL INTELLIGENCE IN PRECISION AGRICULTURE

Project Methods
Efforts:Conducting a 'horizon scanning' of trends in published articles, conference proceedings, and gray literature (e.g. policy papers) about some of the potential social, ethical, economic, and environmental challenges and impacts of big data and AI in PA.Summary of the horizon scanning trends will be presented at scenario-based foresight workshops, held online, where a wide range of stakeholders across the food system value chain will be invited to participate in homogenous (by sector / expertise) groups. The workshop participants will co-produce scenarios to examine plausible futures, which 'could happen', and explore how the society would have to change if certain trends were to strengthen or weaken. The workshop participants will then discuss the implications of AI and big data in each scenario by responding to questions such as: What role will big data and AI play in this plausible future? What are the social and ethical implications that might arise as a result of emerging technologies.The team will use the benefits and risks of big data and AI in PA obtained from the foresight workshops to conduct a Multicriteria Decision Analysis (MCDA) process. We will use a web-based platform (EngagementHQ) for conducting the MCDA. Participants will complete a survey-using a centralized online hub (EngagementHQ)-which will include questions about the benefits and risks associated with big data and AI in PA as well as the most desirable strategies identified by foresight workshop participants. Participants will rank the benefits and risks of big data and AI technologies and strategies for responsible innovation and governance using a Likert scale ranging from 1 to 5. Their rankings will be converted into variable weights using the rank-order centroid (ROC) method. ROC performs better than other methods of ranking because it minimizes the maximum error of each weight by identifying the centroid of all possible weights. We will then create an overall index for each group of participants so that shared objectives and strategies to support the responsible innovation of big data and AI in PA are clearly identified.Content from foresight workshop will be analyzed using: (1) topic modelling to identify stakeholders' interests and concerns related to innovation and governance of big data and AI in precision agriculture; and (2) network analysis to map collaborations and relationships between stakeholders across the food system value chain who may want to strategize to develop effective governance approaches. We will use a method called Latent Dirichlet Allocation (LDA), which is a probabilistic model designed to identify latent or hidden topics in a collection of documents (corpus).The LDA model assumes a document is a mixture of topics and a topic is a collection words with probabilities attached to them. Therefore, the topic proportions will be specific to a document but topics are shared by a whole collection of documents. Following this model, a transcript from a single foresight workshop participant is imagined as a mixture of topics, with each topic being a mixture of words. Moreover, network analysis will be used to identify clusters of participants who share similar characteristics, such as: perceived risks/benefits of big data and AI, desired strategies, and experience with developing or using PA.We will design a digital serious game using the Unity Development Platformand host the game online using WebGL. Serious games will be designed and employed to determine how farmers and farm advisors' respond to different technology and policy incentives. The game will simulate a 4×3 experimental design. The first dimension of our 4×3 experimental design will simulate AI information provision to producers from an AI platform with (a) zero, (b) low, (c) medium, and (d) high level of uncertainty. The second dimension will simulate three data ownership conditions: (a) "farmer owns data", (b) "Agritech owns data" and (c) "farmer can purchase their data from agritech for dollars." Therefore, the game design will have four conditions for AI uncertainty and three conditions for data ownership, for a total of twelve conditions that all players will engage with in the game. The serious game will be used to compare and contrast the behavioral preferences of farmers and farm advisors across the 2 factors and 3-4 levels.To pre-test our serious game, we will distribute it–with a 4×3 full factorial design–to 1000 participants on Amazon Mechanical Turk (MTurk). Participants will complete a survey before and after playing the game. The survey will include questions about respondents' risk preferences, trust in AI (in general), and suggestions for improving the game play. Once the game has been pre-tested by about 1,000 MTurk workers, we will use the data to refine the game design and finalize it for wider dissemination.Both difference-in-difference time series regression models and unsupervised machine learning models will be applied on the experimental data generated from the serious games to identify conditions under which data ownership and levels of AI uncertainty are more or less trusted.We will create a science museum exhibit and aneducational block-based coding game where the players (children) will be asked to complete two levels. For the novice coders (level 1), the players will be asked to complete simple coding (e.g. arrange blocks in some order) to fly a drone over a farm. The player will be to take off, maneuver, and land the drone safely to complete the level. The player will get the opportunity to view results from the different sensors on the drone. The next level will gamify a water conservation problem under climate change scenarios. In wetter conditions, the player will fly the drone and precisely apply irrigation to areas of the field that require it urgently. The cost to the environment (in terms of gallons of water saved) and farm economy (e.g. yield) will be shared with the participants and they will be asked to make trade-offs where needed. Once the game is over, the player will be asked to evaluate the game, answer questions about technology and sustainabilityEvaluation:A program evaluator from the Virginia Tech Center for Educational Networks and Impact will assess the impact of research and education activities. The project evaluator will apply a developmental evaluation approach to project evaluation. Developmental evaluation engages the project team in an ongoing learning process using data as it is collected and analyzed as well as the observations of project team members. This approach requires a preliminary evaluation workshop to identify team-generated questions that can guide the reflective processes embedded in team meetings, quarterly team reflections, and yearly evaluation summits. Two evaluation tools will supplement these processes. First, we will use a journey mapping process to detail potential solutions and their solutions as the team identifies specific work tasks to understand the social and ethical implications of big data and AI in PA and strategize responsible innovation. Second, the project evaluator will work with the team to create ongoing ripple effect maps related to how the project evolved and which outcomes and impacts can be attributed to the project. Data to support the ongoing learning process will include both formative – did we do what we said we would do – and summative data, that is, did our efforts lead to the change we predicted.