Generative AI is often discussed in terms of speed, capability and market competition. Kate Crawford, a researcher at USC Annenberg and Microsoft Research who focuses on the social implications of artificial intelligence, argues that another part of the story remains far less visible: the environmental costs of AI.
Her concern is not simply that AI uses energy. It is that the full picture is hard to see, because much of the relevant information is not public.
The hidden cost behind AI systems
Crawford says AI has high environmental costs that are largely unknown. Generative AI systems need energy, but they also require large amounts of fresh water to cool processors and generate electricity.
That matters because water use is not an abstract technical detail. It connects data centers to local infrastructure, public resources and the communities around them. If the numbers are incomplete, it becomes harder to judge whether AI growth is being managed responsibly.
The source points to several available figures that show why Crawford is raising the alarm. She cites estimates that by 2027, global water use for AI could be equivalent to half of the United Kingdom's consumption.
There is also a local example. In West Des Moines, Iowa, a data center for OpenAI's GPT-4 accounted for about 6 percent of the county's water use in July 2022.
Company environmental reports add another signal. Google and Microsoft's water use increased by 20 percent and 34 percent, respectively, within a year.
Why the data gap matters
Crawford says the full environmental costs of AI are "closely guarded corporate secrets." Current figures are drawn from research, limited company reports and data published by local governments.
That creates an uneven record. Some information exists, but it does not provide a complete view of how much energy and water AI systems require, where those demands are concentrated, or how they may grow as systems become more capable.
The practical problem is accountability. If the public, researchers and policymakers can see only fragments of the picture, then debate over AI's environmental impact is forced to rely on partial evidence. Crawford also argues that "there's little incentive for companies to change."
In plain terms, the issue is not only consumption. It is visibility. Without clearer reporting, it is difficult to compare systems, evaluate data center decisions, or decide what standards should apply to AI developers and operators.
Efficiency over distant fixes
Crawford is not calling for the conversation to wait on new technologies. Instead of relying on new technologies like nuclear fusion, as recently proposed by OpenAI CEO Sam Altman, she calls for pragmatic measures to limit AI's environmental impact.
Those measures include prioritizing energy efficiency, building more efficient models and redesigning data centers. Each of those responses focuses on reducing the environmental burden of systems being built and operated now.
"Rather than pipe-dream technologies, we need pragmatic actions to limit AI’s ecological impacts now," says Crawford.
That framing is important. It shifts the debate away from whether future breakthroughs might eventually ease AI's energy needs and toward what can be measured, improved and reported in the present.
Sam Altman, CEO of OpenAI, has recently warned of an energy crisis in the AI industry, citing the enormous energy needs of future AI systems. Crawford's argument sits beside that warning but points to a broader concern: energy is only part of the environmental footprint, and fresh water use also needs attention.
A proposed reporting framework
Crawford hopes that US Democrats, led by Senator Ed Markey of Massachusetts, can push through the Artificial Intelligence Environmental Impacts Act of 2024, which was introduced on February 1.
The bill would require the National Institute of Standards and Technology to work with scientists, industry and civil society to develop standards for assessing the environmental impacts of AI. It would also create a voluntary reporting framework for AI developers and operators.
Based on the source, the bill's future is uncertain. But the proposal shows one route toward making AI's environmental footprint easier to evaluate: shared standards and a reporting structure, even if voluntary.
For companies, that could mean more consistent expectations around what should be measured. For researchers and lawmakers, it could make the evidence base less fragmented. For the public, it could make claims about AI's environmental impact easier to understand and challenge.
What comes next
The source makes clear that there is no single actor that can solve the issue alone. Crawford says it will take the combined efforts of the AI industry, researchers and lawmakers to truly address the environmental impact of AI.
That is a demanding standard, but it reflects the nature of the problem. Companies operate the systems and data centers. Researchers can examine impacts and methods of assessment. Lawmakers can push for standards and reporting structures.
Generative AI is becoming a larger part of digital life, but its environmental costs remain partly obscured. Crawford's central point is that serious decisions require better information, and that waiting for distant technical fixes is not enough.