Is AI Actually Bad for the Environment? Looking at the Full Picture
local AI writing tool, private AI writing tool
This is the question people actually want answered, and the honest response is: it depends on which part of the picture you're looking at, and over what timeframe. Both the "AI is an environmental problem" case and the "AI could help solve environmental problems" case are backed by real data. Here's both, side by side.
The Case That AI Is a Growing Environmental Cost
Start with the International Energy Agency (IEA), the most authoritative source tracking this globally. In its central projection, emissions from electricity used by data centers are set to roughly double, growing from about 180 million tonnes of CO2 today to around 300 million tonnes by 2035 (IEA). Data centers are explicitly flagged by the IEA as one of the fastest-growing sources of emissions in the entire energy sector right now (Carbon Credits).
There's also a documented gap between stated efficiency gains and actual outcomes. Google's own disclosures show it reduced data center emissions by 12% in 2024 through clean energy purchases and operational improvements, yet its absolute electricity consumption from data centers still grew by 27% in the same year, and its overall greenhouse gas emissions rose 51% between 2019 and 2024 (Brookings). Efficiency is real, but right now, growth is outrunning it.
The Case That AI Could Be a Net Environmental Positive
Here's the part of the picture that gets less attention. The same IEA report estimates that if existing AI applications were widely adopted across other industries, they could reduce global energy-related emissions by an amount equivalent to roughly 5% of the total by 2035, around 1.4 billion tonnes of CO2 per year. That reduction would be three to five times larger than the emissions data centers themselves are projected to produce in that same year (IEA).
That's not a marketing claim, it's the IEA's own modeling, and it comes from concrete applications already in use today: AI-powered satellite monitoring that detects methane leaks in oil and gas operations so they can be repaired faster, AI systems that improve efficiency at fossil-fuel power plants, and machine learning models that make renewable energy sources like wind power more predictable and valuable to the grid (IEA).
The Critical Catch in That Good-News Number
Here's where the IEA's own analysis adds an important asterisk: that 5% emissions-reduction potential isn't guaranteed. The report states plainly that there is currently no momentum ensuring the widespread adoption of these beneficial AI applications, and that their real-world impact, even by 2035, could end up being marginal if the barriers to adoption aren't addressed (IEA). Those barriers include limited access to data, missing digital infrastructure, regulatory restrictions, and social or cultural resistance to adoption.
In other words: the upside case is real and well-documented, but it's a potential, not a guarantee, while the cost side of the ledger, rising electricity and water use, is already happening right now.
A Third Path: Reducing How Much Needs the Cloud at All
There's a different kind of solution that doesn't depend on industry-wide adoption timelines or policy decisions: reducing how much AI processing needs to happen in a power-hungry data center in the first place. The World Economic Forum reports that AI chips designed specifically for on-device processing, running an AI model directly on a phone or computer instead of a remote server, can use 100 to 1,000 times less energy per task than the equivalent cloud-based version, because these chips are built to prioritize efficiency over raw computing power (World Economic Forum).
This isn't a fix for every kind of AI task, the comparison depends on model size and task complexity, and some academic research has found cloud processing remains more energy-efficient for certain mobile workloads where data transfer is cheap relative to local processing (arXiv). But for the routine, everyday tasks most AI use actually consists of, drafting, brainstorming, simple Q&A, on-device processing is something individuals and companies can act on directly, today, without waiting on the broader adoption question the IEA's climate-benefit numbers depend on.
So, Is AI Bad for the Environment?
The most accurate answer is: AI's current environmental cost is real and growing quickly, and remains a modest share of global emissions today; AI's potential environmental benefit is larger on paper, but depends entirely on adoption that hasn't happened yet. Treating either side of that as the whole story oversimplifies what the actual data shows.
Sources
Carbon Credits, AI's Energy Hunger: Data Centers Set to Use Power Equal to Japan's by 2035
Brookings, Global energy demands within the AI regulatory landscape
World Economic Forum, How on-device AI could help us to cut AI's energy demand
arXiv, Cloud-based or On-device: An Empirical Study of Mobile Deep Inference