Is AI bad for the environment?
Mostly no at the individual level, increasingly yes at the system level. One AI prompt uses about 0.24 Wh — nine seconds of TV. But U.S. data centers hit 4.4% of national electricity in 2023 and are projected at 6.7–12% by 2028. The problem is the buildout's speed and location, not your usage.
Why — the first-principles explanation
Almost every fight about this question is really two different questions wearing the same clothes. "Is my use of AI bad?" and "Is the AI industry's buildout bad?" have opposite answers, and mixing them produces either smug dismissal or panic.
On the first: the per-use numbers are genuinely small. Google's August 2025 measurement puts a median Gemini text prompt at 0.24 Wh, 0.26 mL of water, and 0.03 gCO2e — less energy than watching about nine seconds of television. Even multiplied by heavy daily use, that stays far below the footprint of one car trip or one steak. And it is falling fast: Google reports a 33x drop in energy per prompt and a 44x drop in carbon in a single year. Anyone telling you to feel guilty about sending a message is aiming at the wrong target.
On the second: the aggregate is real and moving fast. Berkeley Lab's report to Congress found U.S. data centers used 176 TWh in 2023 — about 4.4% of all U.S. electricity, up from 1.9% in 2018 — and projects 325–580 TWh (6.7–12%) by 2028. The IEA puts global data centres at ~415 TWh in 2024 (~1.5% of world electricity), heading to ~945 TWh by 2030. Note the honest uncertainty in those ranges: nobody knows, and the IEA's 2035 scenarios span 700 to over 1,700 TWh. The width of the range is the real finding.
The harm mechanism is not the total, though — it's concentration and timing. Data centers cluster in a handful of counties, so a load that's a rounding error nationally can be the dominant new demand locally, forcing utilities to keep old fossil plants online or fire up gas peakers. Same for water: 66 billion liters spread over America is nothing; drawn from one stressed aquifer it's everything. Meanwhile AI also helps the environment in places — grid optimization, materials discovery, climate modeling, methane leak detection — but those benefits are diffuse and unproven at scale while the costs are concrete and immediate. Honest verdict: AI is not an environmental catastrophe and not an environmental hero. It's a fast-growing industrial load whose damage depends almost entirely on where it's built and what powers it — which are policy and siting decisions, not personal ones.
An example that makes it click
Think of a new gym opening in your neighborhood. Does one person doing one workout hurt the town? Obviously not — a few lightbulbs and a shower. That's your AI prompt: nine seconds of TV. But if two hundred gyms open on the same block in three years, all drawing from the same substation and the same water main, the town has a genuine problem — and it has nothing to do with any individual's workout being wasteful.
That's exactly AI's shape. Berkeley Lab's numbers say data centers went from 1.9% of U.S. electricity in 2018 to 4.4% in 2023, possibly 12% by 2028. Tripling in a decade. But it's not spread evenly like sunshine — it lands in Loudoun County, or Phoenix, or a small town in Georgia, all at once. The right question was never "should I use this chatbot?" It's "where is this thing being built, and what's turning the turbine?"
Key facts
- U.S. data centers consumed 176 TWh in 2023 — about 4.4% of total U.S. electricity, up from 76 TWh (1.9%) in 2018 — and are projected at 325–580 TWh (6.7–12%) by 2028 (Berkeley Lab / DOE, December 2024).
- Global data centres used about 415 TWh in 2024, roughly 1.5% of world electricity, projected to reach ~945 TWh (just under 3%) by 2030 (IEA, 2025).
- A median Gemini text prompt uses 0.24 Wh, 0.26 mL of water, and 0.03 gCO2e — less energy than about nine seconds of television (Google, August 2025).
- Google reports a 33x reduction in energy and 44x reduction in carbon per median prompt over a 12-month period.
- U.S. data center emissions were about 61 billion kg CO2e in 2023, at 0.34 kg/kWh — slightly below the 0.35 kg/kWh U.S. grid average (Berkeley Lab).
- IEA scenarios for 2035 span roughly 700 TWh to over 1,700 TWh, an uncertainty range wider than today's entire consumption.
▶ The 60-second explainer (script)
Is AI bad for the environment? You're actually asking two different questions, and they have opposite answers. Question one: is your usage bad? No. Google measured a median Gemini prompt at 0.24 watt-hours — that's less energy than nine seconds of television. Point two-six milliliters of water, five drops. And it's improving fast: Google reports energy per prompt fell 33-fold in twelve months. Feeling guilty about typing a message is aiming at the wrong target entirely. Question two: is the buildout bad? That's much more serious. Berkeley Lab found U.S. data centers used 176 terawatt-hours in 2023 — 4.4 percent of all American electricity, up from 1.9 percent in 2018. By 2028 they project six-point-seven to twelve percent. And notice that range — it's enormous, because nobody actually knows. The IEA's 2035 scenarios span 700 to over 1,700 terawatt-hours. But here's the thing that actually matters: this load doesn't spread evenly. It lands in a handful of counties, all at once, forcing utilities to keep old fossil plants running and pulling water from specific aquifers. So the honest verdict? AI isn't a climate catastrophe, and it isn't a climate hero. It's a fast-growing industrial load, and whether it's harmful depends on where it gets built and what powers it. Those are policy decisions. Not yours.
What authoritative sources say
People also ask
Should I stop using AI to help the environment?
It would achieve almost nothing. At 0.24 Wh per prompt, a year of heavy use is dwarfed by a single short flight or a few tanks of gas. Your electricity source, transport, and diet all matter thousands of times more.
Is AI worse than crypto mining?
They're different shapes. Crypto's energy is largely proportional to price incentives with no efficiency ceiling; AI's per-task energy is falling sharply — 33x in a year at Google — even as total demand rises.
Does AI help the environment at all?
In specific places — grid optimization, weather and climate modeling, materials discovery, methane leak detection. But these benefits are diffuse and hard to verify, while the energy and water costs are immediate and measurable.
Why do estimates vary so much?
Different boundaries (chip-only vs full infrastructure, on-site vs power-plant water), different hardware generations, and different assumptions about growth. The IEA is explicit that substantial uncertainty exists even about today's consumption.
The same question, asked other ways
- Why is AI bad for the environment?12,100/mo
- How is AI bad for the environment?9,900/mo
- How does AI affect the environment?4,400/mo
- How bad is AI for the environment?2,400/mo
- How does AI hurt the environment?1,300/mo
- Does AI harm the environment?880/mo
- Is AI really bad for the environment?880/mo
- How does AI harm the environment?720/mo
- How is AI affecting the environment?720/mo
- Is using AI bad for the environment?720/mo
- What does AI do to the environment?720/mo
- What is AI doing to the environment?720/mo
- Why AI is bad for the environment?720/mo
- How does AI impact the environment?590/mo
- Is AI actually bad for the environment?590/mo