AI and the future of our food

By Erin Blakemore

February 28, 2022 at 9:00 a.m. EST

nature

machine intelligence

Risks relating to data acquisition, access, quality and trust.

Agricultural data ranges from the molecular to the landscape scale, and spans domains from agronomy and plant breeding to remote sensing and agricultural finance. National and international agricultural research institutions collect copious amounts of data, which could in principle support ML models. However, these data are too often not discoverable, interpretable or reusable.

CGIAR, a global consortium of agricultural research institutes, has in recent years espoused FAIR (findable, accessible, interoperable and reusable) data principles. Although there is progress in increasing findability through standardization, syntactic and semantic interoperability remains elusive due to lack of common data formats and structure protocols, as well as disordered or unused standards.

Reliability and relevance of agricultural data are additional concerns. A decades-long focus on staple crops such as wheat, rice and corn has outweighed research efforts concerning crops of crucial importance to the poorest producers and subsistence farmers, including quinoa, cassava and sorghum14.

Similarly, the people and practices at the centre of Indigenous farming systems are often under-represented in data, despite their contribution to local food security and dietary diversification - 15,16.

For instance, typical agricultural datasets have insufficiently considered polyculture techniques, such as forest farming and silvopasture. These techniques yield an array of food, fodder and fabric products while increasing soil fertility, controlling pests and maintaining agrobiodiversity17. Partial, biased or irrelevant data may result in poorly performing agricultural DSS, thereby eroding smallholders’ and Indigenous farmers’ trust in digital extension services and expert systems, eventually compromising food security.

Risks from narrow optimization and unequal adoption. While optimizing for yield, past agricultural technologies contributed to new pest complexes, loss of biodiversity and pollution3,4. These risks are broadly known, yet may be difficult to avoid if agriculture is further intensified through AI, and yield is prioritized over ecological integrity.

Expert systems and autonomous machines could improve the working conditions of farmers, relieving them of manual, routine tasks18. However, without deliberate and inclusive technology design, socioeconomic inequities that currently pervade global agriculture, including gender, class and ethnic discriminations19,20, and child labour21 will remain external to ML models applied in agriculture. This is no minor concern; over 98 million children work in farming, fishing, forestry and livestock, in an intensity that deprives them of their childhood and development opportunities22. Agronomic expert systems that remain agnostic to agricultural labour inputs, namely disadvantaged communities and children in employment, will ignore and thereby might sustain their exploitation.

Furthermore, small-scale farmers who cultivate 475 of approximately 570 million farms worldwide and feed large swaths of the so-called Global South23 are particularly likely to be excluded from AI-related benefits. Marginalization, poor Internet penetration rates24 and the digital divide25 might prevent smallholders from leveraging such advanced technologies, widening the gaps between commercial farmers and subsistence farmers. The dissemination of AI is also likely to raise concerns around the potential effects on farmers’ work, identity, agency and ownership rights, including of intellectual property26. In such circumstances, there are clearly risks that large and small farmers will profit unequally, and smallholders get locked into proprietary systems they do not fully understand27.

Risks from deploying AI and ML at scale. Adoption and use of earlier successive waves of technologies for agricultural 

intensification tended to be led by larger commercial farms with more capital to invest and ability to harvest marginal gains in productivity over larger areas28. This increases the likelihood that commercial farmers may be the first to harvest the benefits of AI-driven productivity, with the potential of widening the divide between large farmers and smallholders.

Concomitantly, as AI becomes indispensable for precision agriculture, we can expect an increasing reliance of commercial farmers on a small number of easily accessible ML platforms, such as TensorFlow and PyTorch. Under these conditions, farmers will bring substantial croplands, pastures and hayfields under the influence of a few common ML platforms, consequently creating centralized points of failure, where deliberate attacks could cause disproportionate harm.

In particular, these dynamics risk expanding the vulnerability of agrifood supply chains to cyberattacks, including ransomware and denial-of-service attacks, as well as interference with AI-driven machinery, such as self-driving tractors and combine harvesters, robot swarms for crop inspection, and autonomous sprayers29. The 2021 cyberattack on JBS30, the world’s largest meat processor, foreshadows potential risks that come from the introduction of digital technologies into agrifood supply chains. A 2021 ransomware attack on NEW Cooperative, which provides feed grains for 11 million farm animals in the United States31, further emphasizes this emerging cyber-crime landscape.

Rapid diffusion of intelligent machines in multi-component, multi-agent systems, such as agriculture, may exacerbate nondeliberate, accidental risks as well32. For instance, if monocultures— where a single genotype of a plant species is cultivated on extensive lands—are irrigated, fertilized and inspected by the same suites of algorithms, a model error or poorly calibrated sensors may lead to excessive fertilization and soil microbiome degradation, at the risk of large-scale crop yield failures. Furthermore, unanticipated, cascading system failures have been shown to arise when the interactions between intelligent agents, specifically in human–machine hybrid systems32, happen faster than humans are able to respond33. As digital tools begin to permeate all aspects of agriculture, and agrifood supply chains, the risk of such ‘flash crashes’ of the type seen in other domains may increase.

Governance mechanisms Data stewardship, ownership and cooperatives. We identify the need for FAIR data frameworks and improved standards for transparency, ownership rights and oversight, across all phases of the agricultural data value chain, including data generation, acquisition, storage and analysis.

Specifically, farmers sharing information on soil type, composition and nutrient availability, land surface phenology, choice of crops, amount of fertilizer used, crop rotations, historical crop yield records and actual yield should all follow open-science data-sharing requirements, specifying the repository and dataset.

Addressing ownership issues by democratizing data access and use via standards-compliant repositories is likely to be a foundational aspect of this approach, enabling more open, multi-stakeholder science and technology development34.

In this context, data-stewardship tools that facilitate agricultural data lakes are essential as global agriculture contends with a deluge of data from multiple sources of varying types. These tools must protect farmers’ proprietary rights, ensure data can be trusted, determine how data can be used and enable effective data mining.

The use of industry standards such as ontologies and controlled vocabularies in data lakes should support data mining across disciplines, heterogeneities and sources that differ in modality, granularity, structure and scale35. For example, CGIAR’s Platform for Big Data in Agriculture provides tools and workflows36 to generate FAIR data with support from several platform-mediated communities of practice and

including the development and use of ontologies to improve semantic interoperability. Data cooperatives, or platforms owned and controlled by their members, are a recent model and potential response to the need for more transparent and democratic governance of farm and farmers’ data. Several examples in the United States include the Ag Data Coalition (ADC) and the Grower Information Services Cooperative (GiSC).

Some data cooperatives such as ADC offer secure data repository solutions where farmers can store their data and decide which agencies or research entities to share it with. Others, such as GiSC, offer ‘data pools’ with shared data resources and analytics services to provide peers with improved insight into their farming practices.

Similar approaches are being tested in emerging economies. For example, Digital Green is developing FarmStack, a data-sharing platform and peer-to-peer data-sharing standard for farmers in India, with offerings like those of ADC. Yara and IBM are collaborating to enable farmers to securely share data and determine who uses the data and how37. A key challenge emerges from the tension between a democratized access to data and data monetization. On one hand, if ML systems profit from data contributed by farmers, farmers should be fairly compensated for generating these data. Furthermore, monetization should incentivize growers to share more and better-structured data.

On the other hand, various AI systems provide benefits without financial gains, and would be limited if the cost of access to data is prohibitive. One option to consider is a licensing structure that differentiates between commercial and non-commercial use of data. Another alternative is to share data only amongst groups who all stand to benefit from sharing, such as smallholders in polyculture systems. Data cooperatives could provide a governance structure for exploring different options and making decisions that align with farmers’ best interests.

Responsible innovation. The risks delineated above emphasize the need to develop agricultural AI systems and services with sensitivity to context, giving consideration to prospective social and ecological ramifications, and 

placing the data owner at, or close to, the centre of design efforts. Table 1 adapts a responsible research and innovation approach38 to agricultural AI and suggests interventions in the public and private sectors to ensure anticipatory, reflexive, inclusive and responsive development. For instance, anticipatory design of agricultural AI would involve considering and assessing safety concerns beyond data privacy. These might include unsustainable use of chemical inputs, or overexploitation of agroecosystems. Reflexive AI development should invite deliberative collaborations of rural anthropologists, applied ecologists, ethicists and data scientists in co-creating new ML models that safeguard biodiversity and are context sensitive, ensuring AI ethical principles are translated into practice39. Inclusive, participatory, human-centred design should value agricultural paradigms other than industrial farming, including Indigenous knowledge systems.

Civil-society frameworks and fora that give voice to vulnerable and marginalized communities, and circulate their concerns, can support these aims.

Staged, risk-aware deployment in digital sandboxes.

We suggest that initial deployment of AI for agricultural purposes take place in low-risk hybrid cyber-physical spaces, which we refer to as ‘digital sandboxes’, where multiple stakeholders can be engaged in rapid and supervised prototyping and piloting of novel ML techniques, and associated technologies. In such cyber-physical space, models and machines could be assessed under local and closely monitored circumstances. This model is not entirely new.

It has precedence, for instance, in biotechnology frameworks governing and enforcing biosafety protocols in genetic, genomic and genetically modified organism research40. Digital sandboxes that report on possible failures of nascent technologies would ensure that experimental practices such as autonomous pest and pathogen diagnosis and control systems are precise as well as safe and well-secured.

At the same time, anonymizing data relating to failed deployment attempts and sharing it with agricultural AI communities will allow lessons to be learned and accelerate safe and secure innovations.

The Hands Free Hectare project (https://www.handsfreehectare. com/) at Harper Adams University in the UK, where autonomous precision agriculture interventions are tested and validated, is one example of such a cyber-physical space in a European context; the AI Lab at Makerere University (https://air.ug/) in Kampala, Uganda, where ML anticipates the spread of plant diseases, demonstrates how the approach works in an African context. This approach has several ancillary benefits. For instance, digital sandboxes that operate in open-science partnerships that link public, private and non-profit institutions can create the context for prototyping AI applications safely. They can also help inform rules and regulations for rolling applications out responsibly. Whereas regulatory rigidity may prevent prototyping of novel ML techniques, government agencies could give special, interim exemption to experimentation and learning spaces such as digital sandboxes before developing targeted, customized regulatory frameworks. Moreover, multi-stakeholder approaches to experimentation and learning, such as digital sandboxes, can create opportunities to apply responsible innovation principles in technology design.

Conclusion

Widespread deployment of AI in agriculture is both valuable and expected. Nonetheless, the history of technological modernization in agriculture strongly suggests that a focus on increased productivity carries potential risks, including intensifying inequality and ecological degradation. Agricultural AI must avoid the pitfalls of previous technologies, and carefully navigate and ameliorate their predicaments by implementing comprehensive risk assessments and anticipatory governance protocols. From data collection and curation to development and deployment, general principles of responsible and participatory AI should be tailored to the distinct challenges facing agriculture, at local and global scales. Failure to do so may ignore and thereby perpetuate drivers of nutritional insecurity, exploitation of labour and environmental resources depletion. Previous mis-steps notwithstanding, technological modernization in farming has achieved much. Past successes, too, should inform and inspire the use of agricultural expert systems and intelligent machines. Accordingly, it is essential that a balanced approach towards innovation is practiced, and that risk assessments and responsible research and development procedures do not stifle innovation in a system so fundamental to human wellbeing. Finally, the emerging risk landscape discussed here is also applicable to agricultural systems that provide non-food products; a similar approach should therefore be considered in the production of fibres, fuels, pulp, paper, oils, resins, cosmetics, rubber and plastics.  (Received: 20 August 2021; Accepted: 2 January 2022; Published online: 23 February 2022)

References

1. The State of Food Security and Nutrition in the World 2020: Transforming Food Systems for Affordable Healthy Diets (FAO, IFAD, UNICEF, WFP, WHO, 2020). 2. Cole, M. B., Augustin, M. A., Robertson, M. J. & Manners, J. M. The science of food security. NPJ Sci. Food 2, 14 (2018). 3. Matson, P. A., Parton, W. J., Power, A. G. & Swift, M. J. Agricultural intensification and ecosystem properties. Science 277, 504–509 (1997). 4. Quinton, J. N., Govers, G., Van Oost, K. & Bardgett, R. D. The impact of agricultural soil erosion on biogeochemical cycling. Nat. Geosci. 3, 311–314 (2010).

5. Singh, R. B. Environmental consequences of agricultural development: a case study from the Green Revolution state of Haryana, India. Agricult. Ecosyst. Environment 82, 97–103 (2000). 6. The State of the World’s Plant Genetic Resources for Food and Agriculture (FAO, 2010). 7. Semchuk, K. M., Love, E. J. & Lee, R. G. Parkinson’s disease and exposure to agricultural work and pesticide chemicals. Neurology 42, 1328–1328 (1992). 8. Campbell, J. In Topical Research Digest: Human Rights and Contemporary Slavery 131–141 (Univ. Denver, 2008). 9. Nguyen, H. Q. & Warr, P. Land consolidation as technical change: economic impacts in rural Vietnam. World Dev. 127, 104750 (2020). 10. Nilsson, P. The role of land use consolidation in improving crop yields among farm households in Rwanda. J. Dev. Stud. 55, 1726–1740 (2019). 11. Du, X., Zhang, X. & Jin, X. Assessing the effectiveness of land consolidation for improving agricultural productivity in China. Land Use Policy 70, 360–367 (2018). 12. Schmitz, A., & Moss, C. B. Mechanized agriculture: Machine adoption, farm size, and labor displacement. AgBioForum 18, 278–296 (2015). 13. Wilde, P. Food Policy in the United States: An Introduction (Routledge, 2013). 14. Tadele, Z. Orphan crops: their importance and the urgency of improvement. Planta 250, 677–694 (2019). 15. Lugo-Morin, D. Indigenous communities and their food systems: a contribution to the current debate. J. Ethn. Food 7, 6 (2020). 16. Akinola, R., Pereira, L. M., Mabhaudhi, T., de Bruin, F. M. & Rusch, L. A review of indigenous food crops in Africa and the implications for more sustainable and healthy food systems. Sustainability 12, 3493 (2020). 17. Jose, S. Agroforestry for ecosystem services and environmental benefits: an overview. Agrofor. Syst. 76, 1–10 (2009). 18. Talaviya, T., Shah, D., Patel, N., Yagnik, H. & Shah, M. Implementation of artificial intelligence in agriculture for optimisation of irrigation and application of pesticides and herbicides. Artif. Intell. Agricult. 4, 58–73 (2020). 19. Palacios-Lopez, A., Christiaensen, L., & Kilic, T. How much of the labor in African agriculture is provided by women? Food Policy 67, 52–63 (2017). 20. Alkon, A. H., & Agyeman, J. (eds) Cultivating Food Justice: Race, Class, and Sustainability (MIT Press, 2011). 21. Edmonds, E. V. & Pavcnik, N. The effect of trade liberalization on child labor. J. Int. Econ. 65, 401–419 (2005). 22. Child Labour in Agriculture (International Labor Organization, 2021). 23. Lowder, S. K., Skoet, J. & Raney, T. The number, size, and distribution of farms, smallholder farms, and family farms worldwide. World Dev. 87, 16–29 (2016). 24. Mehrabi, Z. et al. The global divide in data-driven farming. Nat. Sustain. 4, 154–160 (2021). 25. Hennessy, T., Läpple, D. & Moran, B. The digital divide in farming: A problem of access or engagement? Appl. Econ. Persp. Policy 38, 474–491 (2016). 26. Klerkx, L., Jakku, E., Labarthe, P. A review of social science on digital agriculture, smart farming and agriculture 4.0: New contributions and a future research agenda. NJAS Wageningen J. Life Sci. 90–91, 100315 (2019). 27. Wolfert, S., Ge, L., Verdouw, C. & Bogaardt, M. J. Big data in smart farming–a review. Agric. Syst. 153, 69–80 (2017). 28. Levins, R. & Cochrane, W. The treadmill revisited. Land Econ. 72, 550–553 (1996). 29. Sontowski, S. et al. Cyber attacks on smart farming infrastructure. In 2020 IEEE 6th Int. Conf. on Collaboration and Internet Computing (CIC) 135–143 (IEEE, 2020). 30. Cyber-attack hits JBS meat works in Australia, North America. Reuters https://www.reuters.com/technology/cyber-attack-hits-jbs-meat-worksaustralia-north-america-2021-05-31/ (1 June 2021) 31. Sharma, A. $5.9 million ransomware attack on farming co-op may cause food shortage. Ars Technica https://arstechnica.com/information-technology/ 2021/09/5-9-million-ransomware-attack-on-farming-co-op-may-causefood-shortage/ (21 September 2021) 32. Rahwan, I. et al. Machine behaviour. Nature 568, 477–486 (2019). 33. Johnson, N. et al. Abrupt rise of new machine ecology beyond human response time. Sci. Rep. 3, 2627 (2013). 34. Gold, E. R. The fall of the innovation empire and its possible rise through open science. Res. Policy 50, 104226 (2021). 35. Majumdar, J., Naraseeyappa, S. & Ankalaki, S. Analysis of agriculture data using data mining techniques: application of big data. J. Big Data 4, 20 (2017). 36. CGIAR GARDIAN Data Ecosystem https://gardian.bigdata.cgiar.org (CGIAR Platform for Big Data in Agriculture, 2021). 37. Yara and IBM. IBM https://www.ibm.com/services/client-stories/yara (accessed 18 August 2021). 38. Stilgoe, J., Owen, R. & Macnaghten, P. in The Ethics of Nanotechnology, Geoengineering and Clean Energy 347–359 (Routledge, 2020). 39. Theodorou, A. & Dignum, V. Towards ethical and socio-legal governance in AI. Nat. Mach. Intell. 2, 10–12 (2020).