Generative AI’s Hidden Impact

The Unseen Environmental Cost of the Artificial Intelligence Boom

As generative AI becomes integrated into our daily routines, its escalating energy demands cast a long shadow over the planet. The race for smarter, faster models comes at a significant environmental price, forcing a critical examination of the technology's sustainability.

Artificial intelligence is swiftly becoming an inescapable part of modern existence. People now see it appearing unexpectedly in search results, proposing to draft correspondence, and assisting learners with their school assignments. This rapid integration is fuelled by technology behemoths locked in a competitive sprint to craft the most sophisticated models and capture the loyalty of users. Every interaction, every prompt answered, carries an invisible environmental toll.

The Growing Energy Appetite

The computational intensity of artificial intelligence directly translates to higher energy consumption, which in turn generates additional planet-warming emissions. A recent government report offered a stark projection. It found that AI could inflate the share of national power that data facilities use from 4.4 percent to a high of 12 percent by the year 2028. To satisfy this surging demand, some power plants will inevitably burn greater quantities of natural gas and coal. This escalating power requirement places a significant strain on existing energy grids and raises urgent questions about the sustainability of the current trajectory of AI development.

The International Energy Agency projects a startling increase in global electricity demand from data centres, anticipating it will more than double by 2030 to a level exceeding the entire current consumption of Japan. Artificial intelligence is identified as the primary driver of this growth. The United States, a major hub for AI development, is expected to see data centres account for nearly half of its growth in electricity demand over the next six years. This unprecedented hunger for power is not just a future concern; it is a present-day reality with profound implications for climate change mitigation efforts.

The Carbon Footprint of a Query

Not all artificial intelligence is created equal in its environmental impact. Furthermore, certain chatbots have a stronger link to greenhouse gas creation than their counterparts. Research released recently delved into this disparity, analysing the capabilities of various generative AI chatbots alongside the atmospheric pollution they create when operating. Researchers discovered a direct, exponential relationship: chatbots with larger and more complex internal structures, frequently known as bigger "brains," consumed significantly more energy. While these more powerful models often provided more accurate answers, this correlation only held true up to a certain point, revealing a law of diminishing returns.

A more nuanced approach is necessary. The most powerful, heavily trained models are not always necessary for straightforward inquiries. More compact, specialised models can perform specific tasks with a high degree of competence. The strategy should be to select the appropriate system for a given job, a way to optimise for both performance and environmental responsibility. This approach encourages a more mindful selection of AI tools, considering their energy consumption as a key factor in the decision-making process for developers and end-users alike.

A Comparative Analysis of AI Models

The research assessed fourteen large language models, a prevalent type of AI known by the LLM acronym. Each model was subjected to a rigorous test consisting of 500 multiple-choice questions and 500 free-response queries covering five distinct subject areas. The researchers meticulously tallied the power that every model used to complete these tasks. They then converted these energy usage figures into their equivalent CO2 amounts using worldwide figures, providing a standardised metric for comparison. An interesting finding emerged from this analysis: inquiries involving logic-heavy disciplines, such as abstract algebra, consistently yielded the most extensive replies. This suggests they likely required additional power to formulate when contrasted with knowledge-based topics like history.

The study also highlighted a significant difference between standard models and those designed to show their reasoning. AI systems that show their reasoning process step-by-step typically consume significantly greater power for each question than their more direct counterparts. The study examined five reasoning systems, which did not provide answers that were substantially more accurate than the other nine models. In a striking example, the system with the highest emissions, DeepSeek-R1, delivered responses with a correctness level similar to models that generated only a quarter of the pollution.

Image Credit - Freepik

The Limitations of Current Research

It is crucial to acknowledge the gaps in the current understanding of AI's environmental impact. The study, for instance, exclusively featured LLMs that were open-source. This means that a number of the most common AI applications from major technology corporations, including OpenAI’s ChatGPT and Google’s Gemini, did not feature in the analysis. The immense energy consumption and carbon footprint of these proprietary models, used by millions daily, remain largely opaque. This lack of transparency from major tech corporations presents a significant hurdle to a comprehensive assessment of the industry's overall environmental toll.

Furthermore, the paper's conversion of energy use to emissions was done using an international CO2 mean. This provides a useful estimate but does not capture the precise pollution produced from operating these systems in specific locations. The carbon intensity of electricity grids varies dramatically from one country to another. A data centre powered primarily by renewable energy in Norway will have a vastly different emissions profile than one in a region heavily reliant on fossil fuels. This geographical variance is a critical factor that requires more detailed, localised research to fully comprehend.

The Geographic Lottery of Emissions

The location of a data centre plays a pivotal role in determining the environmental cost of an AI model. This point was made by Jesse Dodge, who is a senior research scientist with the Allen Institute for A.I. He explained that some regions are powered predominantly by renewable sources, while others rely heavily on burning fossil fuels to generate electricity. This disparity creates a geographic lottery where the same AI operation can have a dramatically different carbon footprint depending on where it is processed. This highlights the importance of strategic data centre placement in mitigating AI's environmental impact.

A study Dr. Dodge headed in 2022 contrasted the disparities in greenhouse gas pollution resulting from training a large language model in sixteen different regions around the world. The findings were revealing. Based on the season, several of the most polluting zones, including the central United States, exhibited a carbon intensity approximately three times greater than that of the lowest-emitting regions, like Norway. This research demonstrates that a significant portion of AI's emissions is not inherent to the technology itself but is a consequence of the energy infrastructure that supports it.

The Inefficiency of Digital Reasoning

Even with the limitations of current studies, the fresh study addresses a void in examining the balance between power expenditure and system correctness. As Dr Dodge noted, it is widely understood that as a model's size increases, its capabilities, electricity usage, and emissions typically follow suit. The increasing trendiness of reasoning systems, which deliver more extensive, detailed answers, is likely further inflating these energy costs. These models, in an effort to mimic human thought processes, generate more extensive text, which directly translates to higher energy consumption.

For certain topics, a large language model must generate more text to arrive at a more precise answer. Lengthier replies and those that use a reasoning procedure create additional pollution. This presents a fundamental challenge for developers and users of AI. The very features that can make these models more helpful and transparent also make them more environmentally damaging. This trade-off between verbosity and sustainability is a key consideration for the future of AI development and responsible usage.

Output Length Over Subject Matter

Sasha Luccioni, a specialist in AI and climate matters, offers a different perspective on the drivers of AI emissions. She argues that the subject matter of a query matters less than the volume of the output, a factor fixed by how the system was developed. Dr Luccioni also emphasised that the sample size of the recent study is not large enough to provide a thorough overview of the emissions from the entire AI ecosystem. She believes that the focus should be on the volume of text generated, not the topic it addresses, when assessing environmental impact.

In her view, the critical factor is not whether a model is tackling complex mathematics or philosophy, but the volume of the provided input and the subsequent output. A year ago, Dr. Luccioni released research that assessed 88 different LLMs. The work similarly determined that more substantial models tended to have greater emission levels. More strikingly, her findings also showed that formulating text with AI—the primary function of chatbots—consumed tenfold the power of basic classification functions, like organizing emails into folders. This suggests that not all AI tasks are created equal in their energy demands.

Image Credit - Freepik

Overlooking Simpler Solutions

Dr Luccioni points out that as generative systems have gained prominence, older and less energy-intensive AI tools have been forgotten. These "old school" applications, including the classic functions of a search engine, can often perform tasks more efficiently. She contends that for many everyday needs, an ordinary individual has no need for the power and complexity of a large language model at all. There is a risk of technological over-engineering, where sophisticated solutions are applied to simple problems, resulting in unnecessary energy expenditure. This highlights a need for greater awareness among users about the different types of AI available.

Dr Dodge echoes this sentiment, adding that people simply looking for factual information are better served by a standard search engine. Generative AI models are known to "hallucinate," or fabricate false information, making them unreliable sources for factual queries. The use of powerful LLMs for simple fact-checking is not only inefficient but can also be counterproductive. Dr Luccioni suggests that society is in a sense "reinventing the wheel." People do not have to employ generative AI to function as a calculating device, she argues, when a simple calculator application can perform the same function with a fraction of the environmental impact.

The Hidden Water Footprint

Beyond the massive electricity consumption, the AI industry has another significant, and often hidden, environmental cost: water. The powerful servers that run AI models generate immense heat and require constant cooling to prevent failure. Data centres often rely on large quantities of fresh water for these cooling systems. One major tech company's data centres, for example, consumed approximately 5 billion gallons of water in 2022, a 20% increase from the previous year. This places a considerable strain on local water resources, particularly in water-scarce regions where many data centres are located.

Research suggests that a single conversation of about 20 to 50 queries with an AI chatbot can consume half a litre of water. The training phase of a large model could use as much water as is required to produce 370 cars. This thirst for water extends to the manufacturing of the specialised computer chips essential for AI, a process that requires thousands of gallons of ultrapure water for each microprocessor. As the AI industry continues its rapid expansion, its demand for water is projected to grow, potentially reaching between 4.2 and 6.6 billion cubic metres in 2027 alone.

The Two Phases of AI Energy Use

The environmental impact of AI can be broadly divided into two main phases: training and inference. The training phase, where models learn from vast datasets, is incredibly energy-intensive. A 2019 study found that training a single AI model can emit over 626,000 pounds of CO2, equivalent to the lifetime emissions of five cars. More recent research on a popular large model estimated its training consumed 1,287 megawatt-hours of electricity, producing carbon emissions equivalent to driving 112 petrol-powered cars for a year. This initial energy investment is substantial.

The inference phase, however, may ultimately have a greater environmental impact. Inference is the process where a trained model responds to user queries and makes predictions. While a single inference uses less energy than the entire training process, these models are used millions or even billions of times a day. Google has estimated that inference accounts for about 60 percent of the total energy used in its AI operations. As AI becomes more deeply integrated into daily life, the cumulative energy consumption from these countless individual interactions is expected to grow exponentially, potentially dwarfing the initial training costs.

The Hardware Problem: Chips and Supply Chains

The environmental impact of AI begins long before a single line of code is run. The very hardware that powers this revolution, particularly high-performance graphics processing units (GPUs), carries a significant manufacturing footprint. The fabrication of these complex chips is an energy-intensive process that relies on a global supply chain. This chain includes the mining of rare earth minerals and other raw materials, often in countries with lax environmental regulations. The extraction processes themselves can lead to deforestation, soil erosion, and water contamination, creating a hidden ecological cost that is rarely factored into the final carbon accounting of an AI model.

Furthermore, the semiconductor foundries where these chips are made are among the most resource-intensive industrial facilities on Earth. They consume vast amounts of electricity and require billions of gallons of ultrapure water to clean silicon wafers between manufacturing steps. The transport of materials and finished products across the globe adds another layer of carbon emissions to this already costly process. As the demand for more powerful AI escalates, so does the pressure on this resource-heavy supply chain, amplifying its environmental consequences.

Corporate Responsibility and Transparency

Major technology corporations, as the primary developers and deployers of AI, bear a significant responsibility for its environmental impact. Companies like Google, Microsoft, and Amazon have made public commitments to power their data centres with 100% renewable energy and regularly publish sustainability reports. These are positive steps, but they are often met with criticism for a lack of complete transparency. The complex methodologies used for calculating carbon footprints can make it difficult to independently verify claims, and the full impact of their global supply chains often remains obscured.

Critics also point out that power purchase agreements for renewable energy do not always mean a data centre is running on clean power 24/7. In many regions, the grid still relies on fossil fuels when the sun is not shining or the wind is not blowing. True accountability requires more granular, real-time data on energy sources and a greater willingness to disclose the total lifecycle emissions of their AI products, from the mining of raw materials for hardware to the final deletion of a model.

Image Credit - Freepik

The Rebound Effect in AI

There is a paradoxical risk that as AI becomes more energy-efficient, its overall environmental footprint could actually increase. This phenomenon, known as the rebound effect or Jevons paradox, occurs when technological efficiency gains lead to a surge in consumption that outweighs the initial savings. For example, if a new AI model is twice as efficient, it might not lead to a 50% reduction in energy use. Instead, its lower operating cost could spur developers to create more complex applications or encourage users to interact with it more frequently, driving total energy consumption even higher.

This effect is a serious concern in the context of generative AI. As models become more capable and accessible, they are being integrated into an ever-expanding array of products and services. What was once a specialised tool for researchers is now a feature in word processors, search engines, and creative software, used by hundreds of millions of people. This explosion in usage, fuelled by increasing efficiency, could easily negate any environmental benefits gained from more streamlined algorithms, leading to a net increase in global energy demand.

AI for Good: The Other Side of the Coin

Despite its significant environmental costs, artificial intelligence also holds immense potential to help tackle climate change and other ecological challenges. It is crucial to acknowledge this other side of the coin for a balanced perspective. Scientists are using AI to create more accurate climate models, providing better predictions of extreme weather events and long-term climate trends. This information is vital for policymakers and communities to develop effective adaptation and mitigation strategies. In the energy sector, AI is being used to optimise power grids, improving efficiency and making it easier to integrate variable renewable energy sources like wind and solar.

Elsewhere, AI-powered systems are monitoring global deforestation in near real-time, helping to combat illegal logging and protect vital ecosystems. In agriculture, precision farming techniques guided by AI can reduce the use of water, fertilisers, and pesticides, leading to lower emissions and less environmental degradation. From designing more efficient materials to discovering new ways to capture carbon, AI is a powerful tool. The challenge lies in ensuring that the environmental benefits of these applications outweigh the footprint of the technology itself.

Policy and Regulation on the Horizon

As awareness of AI's environmental impact grows, governments and international bodies are beginning to consider regulatory action. Currently, the industry is largely self-regulated, with companies setting their own targets and reporting standards. However, this may change as policymakers look to establish a level playing field and ensure corporate accountability. Potential measures could include mandatory reporting standards for the energy consumption and carbon emissions of AI models, similar to the energy efficiency labels found on household appliances. This would provide consumers and businesses with the information they need to make more sustainable choices.

Other potential regulatory avenues include the implementation of carbon taxes specifically for data centres or the creation of government incentives for the development and adoption of "Green AI" technologies. International cooperation will be essential to establish global standards, preventing companies from simply moving their energy-intensive operations to regions with weaker regulations. While the policy landscape is still in its early stages, the growing public and political pressure suggests that the era of self-regulation for AI's environmental footprint may be coming to an end.

The Future of AI Hardware

The quest for more sustainable AI is driving innovation in the very hardware that underpins the technology. Researchers and engineers are actively exploring new computing paradigms that could drastically reduce the energy consumption of AI models. One of the most promising areas is neuromorphic computing, which involves designing chips that mimic the structure and function of the human brain. The brain is incredibly efficient, performing complex tasks with a tiny fraction of the power required by a conventional supercomputer. Neuromorphic chips aim to replicate this efficiency for AI applications.

Other emerging technologies include optical computing, which uses photons (light) instead of electrons to process information, potentially offering much higher speeds with lower energy use. Advances in semiconductor materials and chip architecture are also yielding more efficient designs. While these technologies are still in various stages of research and development, they offer a glimpse into a future where the computational power of AI is not so tightly coupled with massive energy consumption. This hardware innovation will be a critical component of building a truly sustainable AI ecosystem.

A User's Guide to Sustainable AI

While developers and policymakers hold significant power, individual users and businesses can also take concrete steps to mitigate their AI-related environmental footprint. The first step is awareness. Understanding that generating an image or a lengthy text document with AI consumes a tangible amount of energy can foster more mindful usage. For simple tasks, consider whether a powerful generative model is truly necessary. Often, a traditional search engine or a simple calculator app is a far more energy-efficient tool for the job.

When using generative AI, being precise with prompts to produce shorter, more concise answers can significantly reduce the energy consumed. Avoid open-ended queries that encourage long, rambling responses. Businesses can adopt policies that favour the use of smaller, specialised AI models over large, general-purpose ones for specific tasks. As the industry evolves, look for and support companies that are transparent about their environmental impact and are actively investing in sustainable practices. Collectively, these small changes can contribute to a larger shift towards a more responsible use of this powerful technology.

The Social Cost of AI's Energy Thirst

The voracious appetite of data centres for energy and water creates not just environmental problems but social ones as well. The construction of massive data centre campuses can place immense strain on local resources, particularly in rural or arid regions. This can lead to direct competition with local communities and agriculture for access to water and electricity, sometimes driving up prices for residents. In some cases, the promise of new jobs has been used to justify generous tax breaks and utility agreements for tech companies, only for the highly automated facilities to create fewer local employment opportunities than anticipated.

Furthermore, the need for a constant, reliable power supply has led some energy providers to delay the retirement of fossil fuel plants or even build new ones to service data centres. This not only undermines climate goals but can also subject nearby communities, often those with lower incomes or from minority backgrounds, to increased air and water pollution. The global boom in artificial intelligence must not come at the expense of local communities, and a just transition requires that the social costs of the industry's resource consumption are fully acknowledged and addressed.