Artificial intelligence promises many potential breakthroughs. Among these is the ability to sift through large, diverse landscapes of data, to identify patterns that might not otherwise be visible. Another side of this is the delivery of real-time evolving forecasts and predictive insights that can make adaptive management more of an everyday practice.

The core challenge in both of these areas is that AI systems are not strictly designed to reflect what is happening, but rather to guess at what might happen. This means three very important things must be considered, when we talk about using AI systems to parse large landscapes of data and to provide actionable insight:

1. AI-based data anlayses need to be hand-checked by people who understand potential glitches in the cross-referencing, the structural peculiarities of relevant human and planetary systems, and thematically focused advisers to end users.
2. Local insights are indispensable. Even if every detail of an AI-generated predictive data analysis is exactly right, this will be impossible to know if the data are not checked against local conditions, needs, and capacities.
3. The real-world needs of end users should drive design and output of predictive insights. This means large, global service providers or tech giants are not the appropriate arbiters of what constitutes climate intelligence, nor can their AI systems decide what is needed by whom.

For climate intelligence to play an active role in everyday decision-making, a multi-layered ecosystem of locally focused service providers needs to help ensure data translates into local experience in a way that improves lives and livelihoods.

Consider the potential serious consequences of getting things wrong on any of these points:

Let’s imagine, as a thought experiment, that AI systems cross-reference 437 different pools of data, each with different timescales and frequencies of observation, each addressing outcomes at a different scope across a different core group of sectors of human activity. In each of the 437 cases, there are several items that don’t line up well with the others. This quickly adds up to millions of opportunities for mathematical error in each calculation, and the AI system need to “decide” how to smooth over those errors.

Imagine the predictive insights have to do with food production, and a cascade of AI “judgments” leads to the insight that something that has never grown locally is the crop most likely to succeed. A boom market emerges, as everyone rushes to make this resilience adjustment, only there are several reasons why the insight cannot possibly be right. This cross-referencing needs to be hand-checked, and the whole system calibrated to produce relevant, actionable insights, or an entire growing region could experience harvest collapse and send economic ripple effects across the world.

Local insights would make it much easier to identify potential errors, especially if the AI system in question appears to be making judgments based on commodity market trends, without properly accounting for geography. By focusing on what farmers need to produce, what local and regional markets favor, and the economic and ecological constraints that affect their local behavior, errors can be minimized, and the predictive insights can be made more relevant, rooted, and actionable.

If the needs of end users are not considered, AI systems could be designed to favor revenues to large-scale AI service providers. The AI systems themselves might be set to identify this potential, and to focus on what kind of information will make a larger number of people dependent on the singular service provider. Imagine farmers trying to find out how to respond to an unprecedented climate-related anomaly, like temperatures or crop pests they have never faced before, and finding only vague non-answers available, because the only pool of information they can access was set up to be global and general, not local and specific.

Again, there is an opportunity for costly insight failure in such a system. It may not make sense for any global company or any government agency to staff the local insight gathering needed to ensure complex data cross-referencing and real-time updated predictive insights are locally relevant, rooted, and actionable. This is where an ecosystem of micro-, small-, and medium-sized enterprises (MSMEs) attuned to this need can serve a vital need.

To simplify: If we don’t hand-check models and outputs, if we don’t consider local insights, and if we don’t design around user needs and priorities, we could see AI systems further degrade regional and global food security at the worst possible time. If we do consider all three of these, we can get better data-driven predictive insights to more people, to ensure food systems are more secure and conducive to better lives and livelihoods, even as climate disruption threatens to undo all of the progress we have made over the last century.

What starts to become clear, when we look at the question of AI for climate intelligence is that AI systems are best applied as computational tools, which enhance our ability to process information and empower human beings, both individuals and institutions, to make better choices. They must not be the decision-makers, and we must not pretend they will relieve us of the responsibility to make our own choices.