Using AI for a deeper understanding of climate disruption

What if you could predict the cascading effects of extreme weather events on global supply chains, or pinpoint greenwashing patterns by looking at where data manipulation may have originated?

These are just some of the novel insights that ClimateGPT, the “first open source AI platform dedicated to addressing the impact of climate change”, could potentially uncover, according to its co-creators, Daniel Erasmus, CEO of Erasmus.AI, and Jonathan Dotan, founding director of Starling Lab and CEO of EQTY Lab, which provides ‘responsible’ AI tooling.

Recently previewed at Davos, ClimateGPT — an ensemble of task-specific large language models synthesised on interdisciplinary (economic, social, natural sciences) research on climate change — is designed to bring greater trust and transparency to the pressing challenges of accurate and authenticated climate data. The 7-billion parameter foundational model which features more than 300 billion climate-specific tokens (climate data that is encoded), is endorsed by the Endowment for Climate Intelligence, a grouping of climate scientists and AI engineers, which calls itself a “centre of excellence for climate AI”.

The result of four years of research, building and testing, Mr Dotan says ClimateGPT has been fine-tuned and trained to understand key concepts around climate. “It’s not enough just to say, ‘Find me the [most] relevant information,’” he explains. “[The model] needs to actually understand key concepts around water shortages, or the ability to think about the impact of weather on an economy.” 

Having a model that has deep knowledge and representation of climate, its vulnerabilities, but also potential solutions, financing aspects and its impact on human activity, means you can better understand the vulnerabilities of various assets to climate scenarios, or how actors or stakeholders within specific scenarios are likely to be influenced, says Mr Erasmus.

Disruption interconnection

“It’s really a decision-making and research support tool,” he explains, “that tries to bring a deeper understanding of the novel interconnections that climate disruptions will create. That’s the challenge. Any asset holder these days looks at their portfolio and says, ‘What are the unknown climate liabilities that we don’t know about or the scenarios that we haven’t thought of?’ The hope is that this model can make some contribution in terms of augmenting decision-making capabilities.” 

Increasingly, the reality is sinking in that we are already in a 1.5°C world, says Mr Erasmus, and we need to prepare for this in terms of the social and institutional implications, and harm mitigation. “What we hope to contribute, in some small way, is to help inform all of the decisions that are undertaken and shift that delta and bring it a little bit closer, because better decisions and actions matter here.”

Mr Erasmus says ClimateGPT has an understanding of extreme weather events at a global scale. “But in an interconnected world, a drought in the Horn of Africa or a landslide in Bangladesh has cascading effects. Those cascading effects are represented in the knowledge in the model. There are also dashboards that can help understand and price risk differently within specific geographic areas and predict cascading effects on global supply chains.”

Transparent and efficient

But is ClimateGPT just another ‘black box’ that spits out useful information and insights, but does not allow users to really understand how it arrived at those conclusions? Mr Dotan and Mr Erasmus say there is a strong focus around transparency of the model. ClimateGPT includes things like a “knowledge graph”, which shows how the model interconnects different ideas about cause and effect relationships. Key references and specific reports used to obtain answers are also displayed.

An AI lineage explorer shows how the model was constructed, and the different data sets used. “Think of it as like a fingerprint of each of the ingredients that went into the creation of the model,” says Mr Dotan. “That’s really important, because you want to instil trust in this process so you know exactly where the ingredients came from, and how they were prepared.” 

Integrating the model with public blockchains like Hedera also ensures the highest levels of data integrity, says Mr Dotan, as it provides an independent form of ‘notarisation’ of the information stored on the ledger, which he says is ‘tamper proof’. Hedera describes itself as “energy-efficient” distributed ledger technology.

“One of the biggest challenges in preparing a model of this kind was not just simply to come up with something that was very impressive from a computational standpoint,” says Mr Dotan. “We also wanted it to be [energy] efficient,” he says. 

They said they scoured the globe to find ‘green’ power to train the model, finding the most energy-efficient GPUs (graphics processing units) in upstate Washington that use hydro power, solar-powered storage in Las Vegas, and servers in Ireland powered predominantly by wind, that could host the model.

Staying responsible

ClimateGPT’s founders say they have done everything possible — including seven pilots with world experts on responsible AI — to think about things like inclusion, ethical interfaces, trust and safety issues to build an AI model that is not only energy efficient, but also responsible.

But Mr Dotan says it would be foolish to think that they could control every use of the model. “That’s a community effort that requires vigilance and a set of shared norms around where we want this AI to evolve over time,” he says.

One of the 14 tests ClimateGPT underwent, says Mr Dotan, was custom-designed around climate misinformation or disinformation. “One of the big problems would be if our model amplified the known and most perilous forms of contrarian claims around climate.”

The good news, he says, is that because the effort to create distortion in the climate community is so co-ordinated, a taxonomy of the most common climate myths and disinformation claims has been identified by online data researchers. This data was used to judge the quality of the model and filter out any common misinformation/disinformation claims.

ClimateGPT is only made available to qualified users, says Mr Dotan, and applications are vetted to ensure people are coming to it with the right intentions.