Intelligence Failure
When it comes to AI, can oil companies be trusted?
An industrial desert stretches for miles through west Texas and New Mexico. Stubborn patches of evergreen brush have crept into the spaces between large expanses of concrete and the rusty fences surrounding them. Instead of trees, a forest of metal drills dots the horizon. The drill sites rumble in a steady drone. The faint scent of sulfur permeates the air. Truckloads of machinery lie in neat rows by every drill as dozens of workers in hard-hats and shiny uniforms shuffle between them. The rhythm and mechanical buzz of activity above ground belie the explosive chaos beneath. This is the Permian Basin: 1.6 million acres of fuel-rich drilling sites in southwest North America, and the pride and joy of one of the world’s largest oil and gas companies, ExxonMobil.
While the rest of the United States has reduced the overall number of oil rigs in operation, oil rigs in the basin have only multiplied, and Texas alone produces more than 40% of all oil extracted by the United States in a year. The abundance of energy resources in the region arose in the early 20th century when geologists discovered a hidden wealth of fossil fuels and gases in the shale rock deposited over 250 million years ago. It was trapped in microscopic pockets and had previously been ignored; all they had to do was figure out how to access it. In recent decades, engineers developed a process called hydraulic fracturing, nicknamed fracking, now one of the most common and lucrative methods for extracting fossil fuels in the United States.

Not far from the basin, in the Texas panhandle, a new collection of data centers is currently under construction to support the rapidly emerging artificial intelligence industry. While the earthy, visceral business of oil and the nebulous realm of the cloud, software, and artificial intelligence may seem worlds — and eras — apart, they are deeply connected in today’s economy; and they incur similar costs. Both consume huge amounts of water, but perhaps most critically they both contribute to rising global carbon emissions and climate change. Like many energy and oil extraction companies, ExxonMobil claims to be working to address the environmental impacts of their operations. Their big move has been to adopt artificial intelligence as part of their strategy. But in reality, artificial intelligence has only compounded Exxon’s capacity for harm. Not only has U.S. oil production grown to a staggering 13.7 billion barrels per day, and natural gas to 114.3 billion cubic feet per day, but AI is paving the way for Exxon to carry out even more fracking operations, continuing to cost local communities dearly and fuel our global carbon emission crisis.
In 2019, ExxonMobil partnered with Microsoft cloud computing to incorporate artificial intelligence across its business, especially in Texas’s Permian Basin. Microsoft’s AI system collects and catalogs a veritable ocean of company data — managerial reports, sensor feedback, third-party reports, and more — transforming chaos into clarity. The company then uses AI’s novel level of computing power to choose the most abundant locations to drill, determine the minimum amount of equipment and personnel needed, and detect damage or leaks long before human supervisors ever could. Now, proposals and permits for new constructions can be backed by AI-produced data, promising smooth operations — which are often mistaken for safe ones.
Following the early success of its partnership with Microsoft (Exxon now produces more barrels at a lower cost than other energy companies), ExxonMobil has invested millions of dollars into producing crucial materials for large computing facilities, and offered its business expertise to help AI companies expand. Exxon is boosting the AI industry, whose operations carry their own significant environmental costs. Extracting, storing, and transmitting the information an AI system needs to operate consumes vast amounts of energy – and it requires even more energy to build the system in the first place. Just the process of training an AI model can produce emissions equivalent to driving more than 100 gas-powered cars for an entire year. The enormous computers built to maintain the necessary computing power devour water at an astounding rate; the machinery requires a constant freshwater flow to remain cool enough to operate. A single, basic conversation with the most common and publicly available chatbot, ChatGPT, may use around 500mL of drinkable water — equivalent to the size of a standard water bottle. The millions of computer chips used by AI must also be mass-produced in carbon emission-heavy industrial plants with materials extracted from environmentally costly mines.
How does one explain Big Oil’s paradoxical embrace of both environmental sustainability and AI? ExxonMobil supports artificial intelligence despite its public commitments to sustainability and reducing carbon emissions because analytical AI appears new and revolutionary. It looks good to the public, and more importantly investors, to be on the cutting-edge. Large, energy-intensive companies often promote innovation as the solution to environmental crises, especially the ones they had a hand in creating. Open AI CEO Sam Altman, for example, has proposed that in order to meet the energy demands of AI systems, we should improve research into nuclear fission power — a project he has invested in — instead of changing how and where AI is used and created. But ultimately, AI is so tempting to large companies not because it is truly innovative, but because it is exceptionally good at maintaining the status quo.
Companies like ExxonMobil can claim they are doing their part to fight climate change by increasing efficiency and reducing waste. However, the introduction of artificial intelligence to environmentally ruinous industries has only increased the toll they take on our planet by exponentially boosting both production and destructive consumer behavior, while at the same time providing a veneer of “data-driven” efficiency and innovation. Moreover, when those in charge use AI to make environmentally costly business decisions, they have the benefit of pointing to a supposedly impartial, data-driven model that makes decisions for them instead of taking responsibility, or even for explaining how a decision was made. They take advantage of technology’s supposed impartiality to claim that their rapid expansion efforts are profitable and environmentally friendly because they are efficient.
ExxonMobil advertises real “green” initiatives: It has made promises to reduce emissions by 2030 and to be carbon neutral by 2050 (primarily via controversial carbon-capture technology); has plans to use electricity-powered equipment; and has worked to stop flaring (burning methane on site). But its AI systems learn exclusively from company data generated by prior human decision-making, and the success and impact of the models they generate are, in turn, determined by humans in charge. Exxon engineers and executives choose the most profitable tracts of land to drill on, and the AI builds on those decisions to select comparably profitable sites even faster — it’s a self-reinforcing decision cycle. Exxon employees are rewarded for maximizing profits, so AI will make the same kinds of decisions those employees did. But when AI machines make these decisions, they are subject to less scrutiny because they are considered impartial. In this way, the model and its human translators work together to more thoroughly obscure the full impact of their actions.
AI’s sudden prevalence and impact on our lives may make it feel inevitable. That may or may not be true — but if AI is here to stay, is there a way we can use it for the good of our planet? The first step is recognizing that artificial intelligence is not revolutionary: It is regurgitative.

These machines aren’t just programmed with a set of instructions, they are trained. AI models are learning machines, and like learning humans they can be misguided, regardless of intention. They find patterns, ask if they produced a desired outcome, and either readjust or form a new “success” pattern based on the answer they get. In the process, human values can unintentionally become embedded within the model, shaping its decisions and outcomes. ExxonMobil, as a corporation, primarily values profit and shareholder interests. It values waste reduction and efficiency as a function of preserving or increasing profit, which has an additional side-benefit of being seen as environmentally sustainable. This improves the company’s image and marketability, which is great for profits. ExxonMobil will only be able to generate AI models that mimic this behavior unless it undertakes serious, systemic change, or chooses to use AI with the intention of changing how it fundamentally operates and defines success.
ExxonMobil has a lot of influence on the industry. It could use AI to make a real and impactful improvement, if it just asked the right questions. Striking the balance between energy needs and sustainability is infinitely complex. Figuring out how diverse communities and ecosystems may be impacted by Big Oil’s operations, or the best way to introduce renewable energy sources while maintaining reliable service for everybody, are knotty, data-rich problems that AI might be uniquely suited to solving. If Exxon’s leaders were to focus on the right problems — the ones truly impacting our global climate and local environments — they might ask an AI model to collect reliable climate data and connect it to their own data on their work, to help them address the specific areas of harm they contribute to. They could use AI to integrate multiple sources of data on the impacts of fracking and the energy costs of AI, a task that would be difficult for humans to do accurately and quickly enough.
Based on Exxon’s track record, such a course correction is unlikely.
Seen from a different perspective, however, AI technology is one of the few tools that climate scientists and activists currently have in common with large polluters, and some have begun embracing the potential that AI has to tackle complex problems. Conventional efforts to mitigate climate change involve large coalitions of many people and long negotiations. Every climate problem, like fracking, can be massive, with many competing stakeholders and imperfect, often fragile solutions. All of which takes time, an especially limited resource when large polluters continue to wreak rapid, neatly packaged havoc on our ecosystems, free from meaningful consequences. AI may be the key to speeding everything up, to coming to better conclusions faster, and to creating more targeted and equitable solutions — however imperfect.
One example is the Urban Climate AI Lab at the University of Illinois Urbana-Champaign run by Lei Zhao, Assistant Professor of Civil and Environmental Engineering. His team has begun exploring the use of artificial intelligence to understand and predict urban heat waves, a phenomenon exacerbated by climate change that impacts increasing populations of people. Before AI, models couldn’t predict the heat burden in specific locations, meaning we were underestimating the danger that cities were in during major heat waves. As the number of city dwellers is expected to increase in the coming decades, it’s important to understand how they may be impacted, and to allocate sufficient resources to protect them in times of extreme heat. AI models could be made detailed enough to address neighborhood-by-neighborhood needs, mitigating the most serious risks while allocating finite resources to help the most people. Before AI, the data were simply too complicated, with each city having its own unique set of circumstances and data libraries. Human researchers on their own could only do their best with regional averages, inadvertently obscuring the needs of vulnerable communities.
The new AI models, headed up by researchers with expertise in climate science and civil engineering, would be designed with specific solutions in mind and trained using carefully human-reviewed and verified data. Researchers like those in the Zhao Research Group have the crucial understanding that any one aspect of climate change is just the tip of an iceberg, with enormous complexity beneath. They have narrowed their focus to one high-impact phenomenon and one model designed explicitly with harm reduction in mind. Through careful selection, they are aiming AI at more complex data within the purview of their narrower question and are taking the extra steps to verify their work and carefully determine its scope of ability. Their example indicates that AI can fix and improve broken systems rather than just break them faster, provided it is used with the intention of finding true solutions, not just the easy, cheap, or profitable ones.
Responsible citizens and scientists are often overwhelmed by the sheer power of corporate money and influence, and the disinformation generated by well-funded junk science. This is in part because good science takes time, care, and consideration, which companies can more freely ignore. AI cannot fix everything, but using it responsibly can give environmentalists the edge they need to reach the broader public, develop faster short-term solutions to current crises (heat warnings, water access, storm response, etc.), and even promote regulatory policies that curb the harm caused by AI-equipped companies like ExxonMobil.
As with all technology, AI is a tool, like a hammer. A hammer can build and create. But you can also use a hammer to smash things to pieces very, very quickly. When it comes to new technology, we tend to occupy a space in between. You can have the best intentions of a builder, but if you rush your work, place the nails poorly, or don’t know — or care — what you are creating, your construction might end up being a deathtrap. We have more control and influence than we think when it comes to AI and the impact it will have on our future, in part because it explicitly learns from its very human creators. But it matters how, when, and where we use it. Most of all, it matters that we take responsibility, and hold others accountable, for the world we are building with AI.
This piece was written for ESE 477, Advanced Environmental Writing, in Fall 2024.