Artificial intelligence (AI) tools are common enough now that we’ve all likely experienced an AI hallucination. I personally wasted plenty of time trying to track down very plausible (but made up) scientific paper references that early ChatGPT coughed up when asked for “evidence”. Even without direct experience, the widespread reporting of many of the (quite bizarre) extremes of this phenomenon have primed people to be quite skeptical of AI.

This was clear in the feedback we received to our recent AI report and follow-up podcast, with Rhishi Pethe. It got me thinking about the significant challenges people see for farmers and their advisors in building trust (or not) in AI.

Regardless of how close we are to a complete solution for AI hallucination, the level of skepticism we’re seeing brings up another important question: to what degree does focussing on the not-yet-perfect elements of AI create more resistance to experimentation and ultimately limit adoption?

Farmer Data Fears, Take Two

We’ve seen the same fear dynamic happen with farm data, and likewise asked about the degree to which these fears are justified, and the degree to which the focus on them has possibly held back adoption.

I am not discounting the seriousness of AI hallucination. Nor am I suggesting that we should adopt the tools now for all use cases equally. 

But I do believe a degree of care is warranted. Technology changes extremely fast, and farmers and agribusinesses need viable tools to increase profitability and shore up resilience. We must lean into nuanced conversations about AI, and properly analyze its future potential, relative to, but not rooted in, its current state.

No Garbage In, No Garbage Out

Another area where AI has strong potential - but has not yet delivered commercially in precision agriculture - is data quality. In other words, overcoming the garbage-in, garbage-out challenge. 

A common example of the current state is the collection of planting or as-applied data. The equipment might collect it, but then a user error in the cab can easily cause the recorded data to be incorrect, diminishing the utility and requiring someone to get involved to clean the data set. 

Even when large, and high quality data sets have been accumulated, this has often resulted in ever bigger silos of data that present intractable challenges of interchange and interoperability for farmers and solution providers alike.

Overcoming this may be one of the early killer applications for AI. With its heritage in the fundamentals of transformer technology, AI models are uniquely well placed to assist in the automated translation of similar (but not the same) data sets. Deep understanding of these data sets will also provide efficient solutions for identifying and removing low-quality inputs.

In the seeding data example above, an AI would have a better contextual understanding and access to relevant history, enabling it to detect and correct more readily. 

AI won’t just excel at cleaning and translating to improve interoperability, it also promises far greater automation of the real-time interchange of data via APIs. There is significant momentum in the wider market for AI-driven software development, and it is probable that AI-based tools will significantly improve the speed and lower the cost of integrating systems via APIs. The end result here could be far greater utility for farmers, and lower costs combined with greater flexibility and responsiveness for vendors.

To unlock this potential, we must tackle the many open questions about where large, high quality training data sets are going to come from. And we must find an ethical and commercial balance between privacy, private good, and technological progress.

We cannot let our poor past experiences overly influence our estimate of what AI might be capable of. 

A Mile Wide and a Mile Deep

One of our key points in the report was about how AI has the potential to transform how humans interact with technology, and how this might drive new levels of adoption. This was not a point about the subject matter of the interaction (i.e., yield maps), but more simply that the conversational interaction style that AI has driven may suit farmers far better than highly graphical user interfaces that predominate in software based products today.

There are also many significant challenges in the more complex area of the subject matter itself, and how good an AI-driven product will be as a subject matter expert for farmers.

There is (justified) skepticism about whether farmers will even want, or trust AI enough, to replace their current trusted human advisors with an AI-powered equivalent. Comments in response to the report pointed out that many of the current AI-driven tools are either targeting advisors as  their initial audience, or otherwise advocating a ‘human-in-the-loop’ approach to introducing AI. This is a sensible initial approach, and will help to ensure that current risks of inaccuracy and hallucination are mitigated by human curators.

Again here though, there is a risk of underestimating the potential of AI in creating entirely new ways of working. Human advisors are also capable of errors of omission and commission. Incentive schemes and regular human biases will always have an impact on the quality and objectiveness of advice. 

One area of huge promise is the sheer scale that LLMs will be able to operate at. We can imagine a highly trained model that would in effect be a mile wide and a mile deep - far surpassing the ability of human cognition in a deep understanding of an agricultural production system

This model promises massive augmentation of human advice. Not to have the human acting as a chaperone or curator of an LLM, but in effect as a superhuman, uniquely powered by a highly specialized AI model - serving more customers, with greater efficiency, accuracy, and precision.

It’s a Matter of Trust

AI doesn’t deserve to be trusted implicitly, but by the same token, shouldn’t start from a position of distrust. To maximize the potential advantages for agriculture, we must lean in.

‍

Like with many other technologies, agriculture will not benefit from adopting AI in isolation from other industries. Our ability to trust and benefit from AI will directly benefit from advancements driven in the wider tech market. Privacy, accuracy, and business model innovation will all advance as core aspects of artificial intelligence as its adoption grows across a wide range of industries.

There’s an opportunity for the agriculture industry to engage proactively with AI developers and researchers to ensure that the unique needs and challenges of the sector are understood and addressed. By collaborating closely, we can work towards AI solutions that are trustworthy, reliable, and truly beneficial for farmers and the wider agricultural ecosystem. This will require ongoing dialogue, experimentation, and a willingness to embrace change while maintaining a critical eye on the technology's limitations and potential risks.

‍