LOOKING BACK | Are AI-Generated Kittens Melting the Polar Ice Caps?

We use a lot of generative AI. It's mostly for work, as we've found ChatGPT and Claude to be excellent writing tools and we are unafraid to turn to Gemini and other models as research assistants.

But people in our family and friends circles are definitely using AI for recreation, making fanciful images and videos and stories with pretty much the same technology we use to boost our productivity at work, and we're not going to lie, we've not only consumed and enjoyed non-commercial, AI-generated content, we've also dabbled in making some of it ourselves. We've never made kitten pictures, but, as any long-tenured surfer of forums and social media already knows, the internet has a thing about cats.

If you launch an image generator, people on the internet are going to access it to make a ton of pictures about cats. The images might be nice and entertaining and fun, but they do come at a cost. In fact, every image we generate comes at a cost, particularly, an environmental one. Even the images we generate for good causes are consuming computing power and, as a result, natural resources.

There is something profoundly contradictory about using an artificial intelligence image generator to create a beautiful illustration of an endangered rainforest, a melting glacier or a polar bear stranded on a shrinking ice floe. The resulting image might help raise awareness about climate change, but its creation requires computing infrastructure that consumes electricity, water and other resources, potentially contributing to the very environmental degradation the image depicts.

Of course, producing a conventional photograph, painting or digital illustration also consumes resources. So does publishing a newsletter, maintaining a website, operating a smartphone or conducting a financial transaction. Virtually every modern economic activity has an environmental footprint. The question is not whether artificial intelligence consumes resources, but whether its consumption is justified by the value it creates—and whether society can afford to continue expanding AI infrastructure at its present pace.

Environmental researchers, climate advocates, technology executives, policymakers and institutional investors are examining whether AI's extraordinary growth threatens to overwhelm electricity grids, increase greenhouse gas emissions, strain water supplies and undermine corporate sustainability commitments. 

At the same time, scientists and technology companies argue that AI could become one of humanity's most powerful tools for addressing climate change. Machine learning can improve energy efficiency, optimize electricity grids, accelerate the development of clean technologies, identify environmental threats and help businesses reduce waste. 

Both propositions can be true. Artificial intelligence can simultaneously contribute to environmental degradation and provide technologies capable of mitigating it. 

For financial institutions, the contradiction has particular significance. Banks, asset managers, insurers and wealth management firms are incorporating AI into their operations while simultaneously facing pressure from investors, regulators and clients to manage environmental risks. The resulting tension is especially pronounced for environmental, social and governance (ESG) investors and financial advisors who promote sustainability as a central investment objective. 

If an ESG-focused financial advisor uses generative AI to create marketing images, summarize sustainability reports or construct environmentally conscious investment portfolios, is that advisor advancing sustainability or undermining it? 

The answer depends on what the AI accomplishes, what resources it consumes, and what alternatives are available. 

The Environmental Cost of Artificial Intelligence 

Artificial intelligence's environmental footprint begins long before anyone enters a prompt into a chatbot or image generator. 

Training sophisticated AI models requires enormous computing resources, frequently involving thousands of specialized processors operating simultaneously. Once trained, those models must be deployed in data centers where servers continually process requests from businesses and consumers. 

These operations require electricity to power processors, networking equipment, storage systems and cooling infrastructure. They also require water, land, construction materials and an extensive supply chain of specialized electronics. 

The environmental consequences therefore extend across the entire AI lifecycle, from semiconductor manufacturing and data center construction to model training, inference, equipment replacement and disposal. 

The distinction between training and inference is important. Training a frontier AI model can require a substantial concentration of computing power over weeks or months. Inference, the process through which a trained model responds to users, typically consumes much less energy per individual interaction. However, inference occurs continuously and at enormous scale. 

As hundreds of millions of people use AI tools for writing, coding, research, image generation and other tasks, the cumulative environmental impact of inference can become substantial. 

Moreover, not all AI requests require equivalent resources. Generating a short text response, analyzing a large document, conducting an extended reasoning process, producing a high-resolution image and creating a video involve different computational workloads. 

Consequently, asking how much electricity a single AI prompt consumes is rather like asking how much fuel a vehicle consumes without specifying whether the vehicle is a motorcycle, sedan, delivery truck or passenger aircraft. 

The answer depends on the model, hardware, data center, workload and electricity supply. 

Electricity: AI's Growing Appetite 

The International Energy Agency's 2025 Energy and AI analysis estimated that data centers consumed approximately 415 terawatt-hours of electricity globally in 2024, representing roughly 1.5% of worldwide electricity consumption. Its central projection anticipated that demand could more than double to approximately 945 terawatt-hours by 2030. 

Importantly, those figures encompass data centers generally, not exclusively artificial intelligence. 

Nevertheless, AI is a major driver of anticipated growth. 

This creates a difficult problem for electricity providers. Data centers require substantial amounts of reliable power, often concentrated in geographic regions where transmission infrastructure and generating capacity are already constrained. 

In some locations, meeting additional demand may require natural gas generation or extending the operation of existing fossil-fuel facilities. Elsewhere, new solar, wind, nuclear or geothermal generation may supply the necessary electricity. 

The resulting carbon footprint varies enormously. 

An AI workload operating on a relatively clean electricity grid may produce substantially fewer emissions than an identical workload powered by coal or natural gas. 

However, even a data center purchasing renewable energy certificates does not necessarily operate on carbon-free electricity every hour. The distinction between annual renewable-energy accounting and actual hourly electricity consumption has become central to evaluating technology companies' environmental claims. 

The concern is not merely that AI consumes electricity. It is that its accelerating demand may compete with other efforts to electrify transportation, manufacturing and buildings. 

Water: The Hidden Cost of Computing 

Water presents another significant environmental challenge. 

Many data centers use water-based cooling systems to dissipate heat generated by powerful processors. Depending on the cooling technology, some water evaporates during operation, while additional water may be withdrawn and returned to local supplies. 

Electricity generation can also consume water, creating an indirect water footprint even for facilities that minimize onsite water consumption. 

Semiconductor manufacturing adds another dimension, requiring highly purified water for numerous production processes. 

A useful distinction is between water withdrawal and water consumption. Withdrawal measures water taken from a source, while consumption generally refers to water that is not immediately returned to that source, including water lost through evaporation. 

These quantities should not be treated as interchangeable. 

Nor can AI's water footprint be reliably reduced to a universal number of bottles per chatbot conversation. Such calculations depend heavily on location, season, cooling design, electricity generation and computational intensity. 

For communities experiencing drought or water scarcity, however, the consequences can be tangible. A large data center's demand may create competition with residential users, agriculture and existing businesses. 

The October 14 reporting by the Los Angeles Times on AI-related water strain highlights the importance of accounting for electricity-related water demand, rather than concentrating exclusively on water used inside data centers. 

Materials, Land and Electronic Waste 

Artificial intelligence also requires physical infrastructure that cannot be produced without environmental consequences. 

Advanced semiconductors depend on complex manufacturing processes, specialized chemicals and global mineral supply chains. Data centers require steel, concrete, copper, networking equipment, electrical transformers and cooling systems. 

Constructing those facilities creates emissions before a single AI workload begins operating. 

Their operation may also generate demand for additional transmission lines, substations and power plants. 

Meanwhile, the rapid evolution of AI hardware creates concerns about electronic waste. Accelerators and servers may be replaced as newer generations offer improved performance, potentially shortening equipment replacement cycles. 

The environmental consequences include mining impacts, manufacturing emissions, discarded electronics and the challenges of recovering valuable materials from sophisticated equipment. 

These considerations complicate the argument that increasingly efficient AI processors automatically make AI sustainable. 

A new processor may perform more calculations using less electricity, but manufacturing and deploying millions of those processors still requires substantial resources. 

Efficiency improvements can also encourage additional consumption, a phenomenon economists describe as the rebound effect or Jevons paradox. 

When computing becomes cheaper, businesses may simply use more of it. 

AI Pollution: Beyond Carbon Emissions 

Carbon dioxide is only one component of artificial intelligence's pollution problem. 

Data center expansion can increase emissions of nitrogen oxides, particulate matter and other pollutants associated with fossil-fuel electricity generation. Diesel backup generators can create localized air-quality concerns, while natural gas turbines used to power facilities may expose nearby communities to additional pollution. 

An October 8 report from environmental law organization Earthjustice and the Better Data Center Project examined the growing popularity of so-called bring-your-own-power arrangements, under which data center developers construct or procure dedicated electricity generation. 

The organizations identified 177 gigawatts of planned natural gas generation associated with such projects. They estimated that ozone and particulate pollution from the proposed developments could produce between $5 billion and $10 billion in annual monetized health damages if the projects materialize. These are projections, not measured current damages. 

The report also questioned whether privately powered data centers might circumvent environmental oversight while continuing to impose costs on surrounding communities.  

This challenges a popular proposed solution to AI's electricity problem: allowing technology companies to generate their own power rather than burdening existing electrical grids. 

Generating electricity independently may reduce some grid constraints, but it does not eliminate pollution. Indeed, depending on the technology, it may increase it. 

Chemical pollution presents another concern. Semiconductor manufacturing uses solvents, acids and other substances requiring careful handling. Data center cooling equipment may use refrigerants with substantial global warming potential, while some emerging cooling technologies have raised questions about the chemicals they employ. 

Meanwhile, the construction and operation of large facilities can produce noise, waste heat, wastewater and other localized environmental impacts. 

The distinction between global and local consequences matters. A facility's greenhouse gas emissions contribute to worldwide climate change, but its noise, water consumption and air pollution may disproportionately affect a relatively small neighboring community. 

Those residents may receive few direct benefits from the AI applications operating inside the facility. 

This creates an environmental justice problem: Who enjoys the benefits of artificial intelligence, and who bears the environmental costs? 

The Growing Environmental Backlash Against AI 

The environmental debate is no longer confined to climate researchers and technology critics. It is increasingly influencing public opinion, local politics, infrastructure development and investment decisions. 

In an AP-University of Chicago survey of 3,424 U.S. adults released September 16, the share of Americans extremely or very concerned about AI's environmental consequences increased from 41% in 2025 to 53% in 2026. 

Among adults ages 18 to 29, that figure jumped from 38% to 60%. 

Importantly, opposition to unrestricted data center development crosses partisan lines. Respondents from both major political parties expressed support for requirements that data centers use cleaner energy and contribute to electricity infrastructure upgrades.  

That bipartisan concern distinguishes the data center debate from many other environmental controversies. 

A community might support artificial intelligence generally while opposing construction of a massive data center nearby. Residents may welcome the technology's economic benefits while questioning whether they should subsidize electricity infrastructure serving multinational technology corporations. 

The October 14 Los Angeles Times investigation illustrated the controversy's growing intensity. Citing research by Data Center Watch, the newspaper reported that at least 45 data center projects worth nearly $68 billion had been blocked or delayed between April and June 2026.  

For technology investors, this opposition represents more than a public relations problem. 

Environmental resistance can translate into permitting delays, construction cost increases, litigation, regulatory restrictions and uncertainty about future access to electricity and water. 

These risks can materially affect the economics of AI infrastructure investments. 

For wealth managers and institutional investors, they should be considered alongside more conventional measures of profitability, growth and valuation. 

Can AI Help Save the Environment? 

The strongest argument for accepting AI's environmental footprint is that the technology might ultimately prevent more environmental damage than it creates. 

Machine learning is already useful in applications ranging from weather forecasting and environmental monitoring to energy demand prediction and industrial optimization. 

AI can help electricity grid operators anticipate fluctuations in renewable generation, identify equipment failures and improve the coordination of energy resources. 

In buildings, intelligent control systems can adjust heating, ventilation and air conditioning to reduce unnecessary consumption while maintaining comfort. 

In agriculture, AI-enabled precision farming can help optimize irrigation, detect crop diseases and reduce excessive application of fertilizers and pesticides. 

Environmental researchers can use machine learning to analyze satellite imagery, monitor deforestation, identify biodiversity changes and model climate-related hazards. 

AI may also accelerate scientific research into batteries, alternative materials and other technologies relevant to decarbonization. 

For insurers and financial institutions, these applications create opportunities to improve climate risk assessment, catastrophe modeling, infrastructure investment and environmental reporting. 

However, the existence of promising applications does not establish that AI is environmentally beneficial on a net basis. 

An August 11 Guardian report highlighted research suggesting that AI's contributions to fossil-fuel production could outweigh the emissions reductions associated with its applications in renewable energy. 

The researchers examined how AI improves productivity across different energy industries. Their modeling indicated that efficiency gains in oil, natural gas and coal production could increase emissions more than AI-enabled improvements in clean energy would reduce them. 

The study challenged the assumption that improving the efficiency of energy production necessarily benefits the climate.  

Artificial intelligence is not inherently committed to environmental sustainability. It can optimize the production of renewable electricity, but it can also optimize oil extraction. 

It can improve public transportation systems, but it can also make advertising more effective at encouraging resource-intensive consumption. 

It can reduce industrial waste, but it can also increase manufacturing productivity and accelerate production. 

The environmental outcome depends on how the technology is deployed and what economic activities it encourages. 

The Economic Growth Problem 

Research released by the MIT Sloan School of Management on August 12 raised an even more fundamental challenge to optimistic predictions about AI and sustainability. 

MIT professor John Sterman and researcher Jennifer Turliuk examined AI's direct emissions, its effects on economic growth and its potential to accelerate climate-related technological innovation. 

Their analysis suggested that AI's greatest climate threat might not be the electricity consumed by data centers, but the additional economic activity AI enables. 

Using the En-ROADS climate simulation framework, the researchers modeled a scenario in which AI eventually increases global economic output by 25%. Under their assumptions, projected warming by 2100 increased from 3.3 degrees Celsius to 3.6 degrees Celsius above preindustrial levels. 

Those figures describe modeled scenarios, not inevitable outcomes. 

The researchers argued that improvements in computing efficiency would be insufficient without policies that reduce emissions throughout the economy. 

In other words, making AI environmentally efficient does not necessarily make the economy environmentally sustainable.  

This distinction is particularly important for financial services. 

An AI system that improves a company's productivity may increase its profitability while simultaneously expanding its overall environmental footprint. 

For investors, financial performance and environmental performance are not always aligned. 

Sustainable Artificial Intelligence: Can Technology Clean Up After Itself? 

The emerging field of sustainable artificial intelligence seeks to reduce the environmental costs of developing and deploying AI systems. 

Sometimes described as green AI, the approach emphasizes energy-efficient models, resource-conscious computing, transparent environmental accounting and infrastructure designed to minimize waste. 

Sustainable AI involves more than powering data centers with renewable electricity. 

It also means questioning whether every application requires a large frontier model, whether computational workloads can be reduced, and whether existing hardware can be used more efficiently. 

Researchers are developing smaller language models, model compression techniques, quantization methods and other approaches that reduce computational requirements. 

Organizations can also route straightforward tasks to smaller models while reserving more resource-intensive systems for complex problems. 

Other strategies include scheduling flexible workloads during periods of abundant renewable electricity, improving cooling systems, extending equipment lifetimes and recovering waste heat. 

An October 8 MIT News profile of associate professor Christina Delimitrou offered an example of AI being used to improve the efficiency of the infrastructure that supports AI itself. 

Delimitrou and her colleagues are applying machine learning to improve the management of cloud computing resources, reduce unnecessary software overhead and make better use of existing servers. 

The objective is to deliver more computational output from existing hardware, potentially reducing the need for additional data centers.  

For businesses, however, efficiency gains must be accompanied by credible environmental measurement. 

On September 29, sustainability publication Trellis reported that Sustainable AI Group had introduced CLEER, a tool designed to estimate energy consumption and greenhouse gas emissions associated with proprietary AI models. 

The methodology uses token consumption and estimates derived from publicly studied models to assess the environmental footprint of enterprise AI activity. 

Etsy, private equity firm Hg and tax software provider Sovos were identified among early users. 

The approach offers organizations a means of incorporating AI-related emissions into broader environmental accounting, although estimates remain imperfect when model providers do not disclose detailed operational information.  

For corporate sustainability teams, such developments could help make AI consumption a measurable business activity rather than an invisible component of technology spending. 

Could AI Ever Be Truly Sustainable? 

The answer depends on the definition of sustainability. 

An AI system could conceivably operate using carbon-free electricity, highly efficient cooling, responsibly sourced equipment and extensive materials recycling. 

Its operational greenhouse gas emissions might approach zero. 

However, eliminating every environmental impact would be considerably more difficult. 

Hardware manufacturing requires materials and energy. Infrastructure occupies land. Equipment eventually deteriorates. Even renewable electricity generation has an environmental footprint. 

A more practical objective is minimizing AI's environmental costs and ensuring that its measurable benefits outweigh those costs. 

That requires examining the full lifecycle of AI systems, including indirect economic effects. 

It also requires distinguishing carbon neutrality from environmental sustainability. A company might offset its reported emissions while continuing to consume scarce water, contribute to electronic waste or disrupt local ecosystems. 

Carbon accounting alone cannot capture every ecological consequence. 

The MIT Sloan analysis suggests that technological efficiency and voluntary corporate initiatives, while valuable, may be insufficient without broader environmental policies and incentives.  

AI may become dramatically more sustainable than it is today. Whether it can become environmentally neutral in every meaningful sense is another question. 

The Financial Services Industry's AI Sustainability Dilemma 

For banks, insurers, asset managers and wealth management firms, AI's environmental impact is becoming both an operational concern and an investment consideration. 

Financial institutions use AI for fraud detection, credit analysis, risk management, customer service, portfolio research and administrative automation. Some applications offer substantial economic benefits, while others provide relatively modest improvements in convenience. 

An institution seeking to reduce its environmental footprint should evaluate these applications differently. 

AI that improves disaster risk modeling or detects financial fraud may deliver significant benefits relative to its environmental costs. Using an enormous generative model to produce dozens of nearly identical marketing images may be more difficult to justify. 

The distinction becomes especially important for ESG investment professionals. 

An environmentally focused advisor should not necessarily avoid AI. Indeed, machine learning may help identify environmental risks, analyze corporate disclosures, compare emissions data and evaluate sustainable investment opportunities. 

But advisors should apply the same scrutiny to AI vendors that they apply to companies in environmentally focused portfolios. 

That includes examining electricity consumption, emissions disclosures, water use, renewable energy procurement, supply chain practices and the credibility of sustainability claims. 

Research from S&P Global illustrates the challenge. In its third-quarter 2026 sustainability analysis, the company reported that 54% of firms in its examined sample had not implemented AI governance policies. It also found that among 498 assessed companies using AI for sustainability, only 30% confirmed that they quantified the technology's environmental impact. 

The problem is not simply that businesses are using AI. It is that many cannot demonstrate whether their AI applications are delivering measurable environmental benefits.  

Financial institutions are also becoming important providers of capital for AI infrastructure. 

An August report citing ING Research described data centers as a growing destination for sustainable financing, including green bonds and loans supporting renewable electricity procurement, cooling upgrades and grid improvements. 

However, financing an AI data center through a green financial instrument does not automatically make the entire facility environmentally sustainable. Investors must distinguish the environmental characteristics of specific financed projects from the broader impacts of the infrastructure they support.  

This creates opportunities for financial advisors to examine not only AI-related equity investments but also the environmental characteristics of infrastructure funds, utility investments, municipal bonds and private credit portfolios exposed to data center expansion. 

Should We Stop Using AI? 

For environmentally conscious consumers and financial professionals, the most practical question is whether they should continue using AI despite its environmental costs. 

A categorical prohibition would be difficult to justify. Society does not generally evaluate other technologies by demanding that they produce no environmental impact whatsoever. 

Instead, businesses and consumers routinely weigh benefits against costs and seek less damaging alternatives. 

AI should be evaluated similarly. 

A financial advisor might reasonably use AI to analyze hundreds of corporate sustainability disclosures, identify inconsistencies in emissions reporting or improve the efficiency of a complex research process. 

The same advisor might question whether generating 50 elaborate illustrations for a single newsletter article creates enough value to justify the computing resources involved. 

The relevant questions are whether AI is necessary, whether a less resource-intensive method would work, and whether the results justify the additional consumption. 

The answers will vary. 

A single image generated for a newsletter is unlikely to meaningfully alter global emissions by itself. But billions of individual decisions collectively influence demand for AI infrastructure. 

This is precisely why environmental responsibility cannot rest exclusively with individual users. 

Technology providers, corporate customers, investors, utilities and governments all influence the infrastructure that determines AI's environmental footprint. 

During September's New York Climate Week, SAP described sustainability as increasingly inseparable from the business decisions governing AI development. The company highlighted growing expectations for transparency about electricity use and clean-energy procurement ahead of the November COP31 climate conference. 

SAP Chief Sustainability and Commercial Officer Sophia Mendelsohn argued that sustainability considerations must be incorporated directly into the data and decisions AI systems optimize. Otherwise, AI may pursue efficiency and profitability while overlooking environmental consequences.  

That observation points toward a broader conclusion. 

Artificial intelligence is not inherently environmentally destructive, nor is it inherently sustainable. It is an extraordinarily powerful technology whose consequences depend on the economic incentives, infrastructure and human decisions governing its development. 

For financial advisors, the challenge is to make those consequences visible and incorporate them into investment analysis, vendor selection and everyday business practices. 

For AI developers, the challenge is to create systems whose environmental costs decline even as their capabilities improve. 

And for the person generating an illustration of a kitten, or a melting glacier, the question is not simply whether the image consumes electricity. 

It is whether the image serves a worthwhile purpose, whether its environmental cost can be reduced, and whether its creation contributes something of value.

The world probably will not be destroyed by one more AI-generated image. But the accumulated consequences of billions of seemingly insignificant decisions are precisely what make environmental sustainability such a difficult problem. 

The irony is that artificial intelligence might eventually become one of our most powerful tools for solving that problem—provided we do not make it considerably worse in the meantime. 

---

Researched by DWN Staff