Modelling world agriculture as a learning machine? From mainstream models to Agribiom 1.0

Models of world agriculture and food systems are used widely to predict future scenarios of land and resource uses. Starting with a brief history of world agriculture modelling since the 1960s, this paper demonstrates the hybrid character of these models as well as their limitations in representing real world diversity and options, and presents an alternative modelling experience. As we argue, models are tools of evidence, hence "truth machines", but also tools of government, with a multi-faceted political dimension. For instance, the virtual realities that conventional models build, incorporate value judgements about the future that remain invisible and difficult to challenge. For ease of computation and comparison, they standardise functional forms and parameters, eliding observable diversity and blacklisting policy options such as those based on agroecology and biological synergies. They are designed for prediction and prescription rather than for supporting public debate, which is also a (comfortable) political stance. In contrast, the Agrimonde experience – a foresight initiative based on the Agribiom model – shows that a model of world agriculture can be constructed as a "learning machine" that leaves room for a variety of scientific and stakeholders' knowledge as well as public debate. This model and its partners unveiled some virtual realities, processes and actors that were invisible in mainstream models, and asserted their own vision of sustainable agri-food systems by 2050. Agribiom and Agrimonde improved knowledge, policy-making and democracy. Overall, they highlighted the need for epistemic plurality and for engaging seriously in the production of models as learning machines.