Discussion

Agricultural AI needs a way for farmers to challenge its advice

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#11434

Agricultural AI can offer useful forecasts and planting advice, but a wrong recommendation may leave farmers carrying the cost. A new essay argues that access is not enough: farming communities need a practical way to question advice and get errors investigated.

Watch Desk analysis

What happened

In an essay published on 5 November, Tuhu Nugraha draws on a recent programme on climate-resilient smallholder livelihoods in Bandung to make the case for accountability in agricultural AI. He argues that farmers need systems grounded in local conditions, and institutions able to hear complaints and respond when advice does not fit what people see in the field.

The essay gives examples of what access can look like. Digital Green’s FarmerChat lets farmers ask questions by voice or text and share photographs. A peer-reviewed case study found expanded access to personalised agricultural information, but did not establish that the service improved crop yields. A June 2026 analysis by the Center for Global Development and Digital Green found that, in one telecom partnership, waiving data charges was associated with 31% more queries; rigorous impact trials were still under way.

It also cites International Telecommunication Union estimates that 17% of rural residents in low-income countries used the internet in 2026, compared with 41% of urban residents. Nugraha proposes practical safeguards for higher-risk services, including independent agronomic review, accessible complaints through existing extension networks and contracts requiring providers to investigate substantiated technical failures. These are proposals, not rules already in force.

Why it matters

A digital service can deliver advice without making it usable, reliable or contestable. A farmer may have little connectivity, limited ability to assess a model, or no clear route to challenge a recommendation. Where advice could affect irrigation, credit or insurance, the stakes are higher than an answer that simply misses the mark.

The essay’s useful distinction is between making advice available and building the capacity to scrutinise it. Local testing, trained extension workers and a process for investigating complaints all take time and money. Treating those costs as part of the service, rather than as an optional extra after a pilot, could make failures easier to find and responsibility harder to shrug off.

Our read

This is a strong argument for judging agricultural AI by more than how many people can access it. FarmerChat’s reported use shows what more accessible interaction can look like, while the evidence cited does not yet establish improved yields. That distinction matters: a service can be promising without being proven, and “the model said so” is not a satisfying appeals process.

Nugraha’s essay in Modern Diplomacy is an argument for better oversight, not evidence that a particular AI system has caused crop losses. Its most practical suggestion is to build review and complaints into higher-risk services before a bad outcome leaves a farmer with nowhere to turn.

What to watch

  • Whether the trials of FarmerChat establish effects on farming outcomes, not only service use.
  • Whether agricultural AI procurement includes independent review and accessible complaints.
  • Whether local extension networks get the resources to investigate errors.
  • Whether safeguards differ sensibly between general advice and recommendations affecting water, credit or insurance.

Discussion spark: For agricultural AI advice that could affect a farmer’s livelihood, who should carry the cost of independent review and a workable appeals process: the provider, the public sector or both?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.