AI’s energy appetite matters but its applied use could be as consequential for sustainability, from accelerating research to orchestrating solutions and resources.

In the arc of emerging technologies, figuring out how to use for public benefit what is freshly born tends to lag behind general use. Entertainment before education in the era of user generated content. Gaming before surgery in the era of AR. Experimentation that yields creative applications, but that perhaps could be expedited to impact. Collective focus is on the race to achieve increasingly advanced artificial intelligence, and the perils and boons it could bring. Yet assuming we get there and existential risk is avoided, do we have a civilizational agenda for what wicked problems should such power be put to work to solve?

Many possibilities. One of urgency: stabilizing and regenerating the planet, its lifeforms and systems, and in so doing, supporting human thriving. As we approach planetary boundaries of biodiversity loss to pollutant overwhelm, and threshold into global water bankruptcy, they compound into a precarious social and economic future. Plastics in our bodies, land conflict over water, and the largest public companies facing an estimated $1.2 trillion annual loss from climate change starting 2050.

AI’s environmental impact over the coming decades will exist in the tension between its voracious resource appetite and the solutions it accelerates. Certainly, compute’s energy and water consumption must be dealt with. Yet there is also a latent power, a window of agenda setting open now to shape how AI is put to use. Environmental challenges deserve much more from AI across investment, talent, bureaucratic support, and presence in the public mind. 

Hiding Latencies

Leaders such as Anthropic’s Dario Amodei predict AI will supercharge innovation and deliver a compressed 21st century worth of R&D in the span of five years. AI is already being used to make discoveries in  critical challenges like energy storage and material science, but testing in the real world remains a key bottleneck. To address this, startup Lila Sciences has raised $200M to create both a “scientific superintelligence” and an automated laboratory where researchers can test hypotheses quickly and cheaply and feed the lab results back into the AI model, with early breakthroughs in materials for carbon capture being achieved in months instead of years. Importantly, their lab is physical, giving much needed real world feedback to teach and inform the theoretical mind. As AI to lab loops proliferate, opportunities abound to point them at unsolved technical challenges within climate solutions, from fusion to cold storage to carbon sequestering materials.

Forecasting insolent weather and environmental shifts has long been the work of Earth data, and is getting an AI boost. Precision farming, as with Rare and Salesforce’s Agent Tierra allowing cultivators to dialogue through Whatsapp and make informed choices.

Above, a mesh of satellites in the sky have for decades gifted the ability to spot deforestation, methane leaks, and more. A network of sensors in forests and cities gathering surface and substrate, from motion-detected animal tracking cameras to soil samplings, have offered ground truthed data on ecosystem habitants and their services. Layered with AI, these rich archives of historical and real-time knowledge about the living world are becoming interpretable to less technical people and are opening opportunities for novel and accelerated pattern spotting. The digital mind of AI is assembling its senses. What emerges as it moves beyond human prompting, and begins to digest this wave of data? What insights or hard truths could emerge, and what might Earth wish to say to us with a new translator? Our self destruction could become evident to an entity that understands our goals of human thriving and sees our dissonant behavior. Hallucinations riffing of this intuition could unleash novel solutions for biosphere regeneration.

Or, this expanded knowledge might help Earth plan its own resourcing. Operating the grid, extracting minerals, and gardening the Earth in perfectly choreographed synchronicity like sunflowers, to turn to available resources throughout the day’s arc. As AI’s agentic capabilities expand, they are beginning to permeate nuanced and trade-off laden decisions on the environment with and without humans in the loop. Decision-augmentation looks like the UN-backed Project Resilience, which uses AI to recommend optimal land use policies. Experts predict grid load balance will increasingly be managed autonomously in coming years. For AI to act responsibly on our behalf it will need to be trained to prioritize planetary health the way it does efficiency or cost savings. Consider Amazon’s Autonomous Supply Chain Agents, which pick suppliers and determine restock timing as well as delivery vehicle routes. Each decision has significant environmental impacts, but how sustainability fits into agentic architecture for any group is nascent. 

Accelerants

What would it take for these opportunity areas to come to fruition?

A perennial challenge is funding. New major philanthropic capital arriving as technology companies IPO could be directed towards these areas. For research science in particular, philanthropy will need to mix with government and venture. Another persistently influential force is the sheep dog of government incentives. How might we construct a cocoon of policy infrastructure that coaxes forth the best minds and best quests within applied AI?

We may need to establish a data commons or culture. Strong and large data sets help models train better and faster, but are often costly. Acting on a subset of environmental data, such as surface water imaged from space of the Amazon, without seeing the substrate flow, brings an incomplete picture that could jeopardize the desired outcomes by designing without a systems view. A whole Earth model will likely need to come to fruition through a networked digital organism of organizations. Groups like Nature Data Lab seek to bring this together but given data’s preciousness, designing fair and motivating incentives is quite a task and perhaps begins in paying for outcomes over hoards.

A surprising absence of will, ideas or practical know-how about how to apply AI exists in quiet corners: within academia, museum labs, traditional science. In part this is because of aversion, there are real and valid fears about misuse, IP, and energy detriment that slow experimentation and adoption that must be addressed. The other part of it is imagination, having time, technical support, and inspiration to be curious about how AI might be used towards one's specific tasks. On the flip side, it is equally important for applied AI to not simply be an environmentalist and activist agenda. Technologists and major labs ideally are activated as well. It is imperative that those operating and developing climate and nature solutions have literacy. 

Customization is another key. General purpose AI can lose some impact fidelity if not fit for purpose, and customizing can bring down model development and training cost. A forester doesn’t need sonnet writing, she needs specific care recommendations for a specific fungus on local trees relevant to the day and acre she finds herself inquiring in. Context jumping and translating becomes its own field: For example Deep Forest is an LLM trained on pines, but when applied to India it struggled and required new training. Historically context translation would be potentially prohibitively costly, yet AI has the opportunity to bring down cost through efficiencies, self-learning, and inter-model coaching.

Perhaps more than anything else, we need clear opportunity area goals. If we learn from the wins and downfalls of carbon markets and non-AI climate tech, advance market commitments like Frontier matter immensely for early-stage ideas and their investors to take the leap into scaling. Public agreements to purchase a solution before it exists market-signal to investors that the area is fundable, startups that it is desirable, and corporations that it is competitively strategic to build and acquire.

A combination of solutions and accelerants might just get us to the future we want.