While the global commentariat debate whether AI will end humanity, the companies are selling themselves as saviours of the world through their foundations. OpenAI Foundation, for example, recently announced a large grant to support better weather prediction for farmers in East Africa and South Asia in order to “open the door to the mass production of forecasts tailored for smallholder farmers across low- and middle-income countries who are trying to decide whether and when to plant, harvest, and fertilize crops”.
Dangers of the tech fix
Previous generations of tech optimists have sold the idea that better prediction will solve farmers’ problems through improved early warning and anticipatory action. Satellite weather forecasts have certainly improved and the processes of data through large language models will no doubt provide even more information, more tailored to particular places and circumstances.
But information is not necessarily knowledge, and more information does not always result in better prediction. Indeed uncertainties – where we don’t know the probabilities of certain outcomes – always exist. In complex, non-linear systems, such as the weather/climate, this will always be so. Downscaling climate models and improving prediction at the more granular scale useful to farmers remains challenging. AI solutions of course make use of the same data but process it in new ways. Yet the uncertainties remain despite the brash claims that AI can solve farmers’ problems during droughts.
This is a bigger problem with the AI tech fix proposed for complex, uncertain problems. Prediction is always elusive and judgement will inevitably be needed. AI models simply analyse existing data, so biases, omissions and uncertainties will be replicated, no matter what the models’ outputs claim. As with any model, if it’s rubbish in, it will be rubbish out; if uncertainty prevails, then this will not disappear even if it can be hidden or assumed away.
Obscuring uncertainties in disaster risk management
The tech promoters will argue of course that AI solutions – just as existing satellite-based meteorological models – are improving. There are more and more accurate data and predictions can make a difference to farmer outcomes, as randomised control trials from India appear to show. This is true, but this equally doesn’t mean that uncertainties are eliminated; far from it. Assuming risk when uncertainty dominates is dangerous providing a confident hubris relying on predictive models when sceptical caution is always needed.
The obscuring of real uncertainties is potentially a real danger. Prediction would be great if you had completely accurate knowledge of the future, but farmers’ distrust of existing early warning systems is well established, and this is for good reason. Equally insurance systems, now being promoted widely as part of AI solutions as tech meets financialisation, would function much better if predictions of future losses could be established. But we equally know that the various types of sovereign and index-based insurance have many pitfalls, as uncertainties intervene in the neat tech-finance formulation.
As part of a grant from the African Development Bank (AFDB) to the Government of Zimbabwe, a nationwide ‘Multi-Hazard Risk Assessment and Mapping’ project is being designed as part of the, ‘Mitigating Fragility through the Africa Disaster Risk Financing Programme in Southern Africa’. Although not captured by the AI hype, the proposed work is framed very much around improving risk prediction through identifying hazards, relating these to exposure (and vulnerability) and mitigating effects through risk management and communication. This is the classic ‘disaster risk reduction’ (DRR) approach, as proposed by the Sendai Framework. The idea is that with this information, the government will be better able to respond through early warning, anticipatory action and the sale of crop insurance products.
The Ministry of Agriculture meanwhile wants to train its full staff base of 14,000 in AI technologies, enhancing digital competence in planning for droughts, diagnosing disease challenges and supporting market connections. From satellites to data processing to online extension advice is the vision. Beyond the logistical hurdles of connectivity, hardware availability and cost, the real challenge is to ensure that the extension workers in the field understand the limits of AI prediction tools and can deliberate intelligently around different sources of information. Not rejecting the promise of AI but engaging responsibly and intelligently.
Models are just models
As many point out the models on which climate forecasting are based are only as good as the data. And much of this remains poor. Models are trained on huge amounts of data, but this may not be derived from African contexts, where different climate, soil and agronomic conditions apply. There is a danger that models produce recommendations that are simply wrong. As one commentary recently argued that “with the rise of AI-generated misinformation, some rural WhatsApp groups may be exposed to fabricated weather alerts, and deepfake videos presented as truth.” This makes it even more critical that extension workers and others being trained for the new digital world are encouraged to deliberate around multiple sources of information, not rely on one source.
The key is to be confident in a world of uncertainty and not rely on a tech fix no matter how sophisticated. Too many become seduced by the appeals of technological solutions but ignore the certainty of uncertainty. What happens if prediction fails and the models do not deliver? With compounding and cascading hazards (where for example drought and heat combine, and effects emerge over time in complex ways), uncertainties inevitably increase. There is no way that early warning, anticipation and insurance can work as assumed under such conditions. We just don’t know the likelihoods of things happening.
Navigating uncertainties
As I argue in my book, Navigating Uncertainty: Radical Thinking for a Turbulent World, (open access), embracing uncertainty suggests a very different approach (see especially Chapters 6 and 7). Where uncertainties exist, then for sure farmers will check out the met office forecasts, but farmers will also triangulate with their own experience, other indigenous indicators whether the direction of the wind or the behaviour of animals, and respond with a sense of caution and judgement. Deliberation around competing sources of information is key when uncertainty prevails, and supporting the capacity for this is as important as making use of AI solutions.
Uncertainties are navigated making use of multiple knowledges and through adaptive learning and incremental improvisation. The standard DRR and digital AI approaches need to be qualified before large sums are spent as officials succumb to the promotion of AI as a tech-fix to save the world. We must remember that even with the best AI models, a tech-fix solution will not be enough and navigating uncertainty, with human deliberation and social engagement, will always be important.
This blog was written by Ian Scoones and first appeared on Zimbabweland
Source: Can AI models help farmers in the face of uncertainty? | zimbabweland
COMMENTS