Does AI's energy use harm polar bears?
Indirectly and very slightly, along with every other use of electricity. Data centers used about 415 TWh in 2024 — roughly 1.5% of world electricity — and emissions of about 180 million tonnes CO2, under 1.5% of energy-sector emissions. Your individual prompts are negligible. The growth rate is the real story.
Why — the first-principles explanation
The causal chain is real, so let's follow it honestly instead of shouting a number at you. Polar bears hunt seals from sea ice. Sea ice extent falls as the Arctic warms. Warming tracks cumulative CO2 in the atmosphere. So anything that burns fossil fuel adds a tiny increment to that total, and AI runs on electricity that is still substantially fossil-fired. The chain connects. The question is only how much — and that's where nearly every article you'll read goes wrong in one direction or the other.
Start with scale. The IEA — the most credible neutral source here — puts data centers at about 415 TWh in 2024, roughly 1.5% of global electricity, with emissions near 180 Mt CO2. All data centers. That's Netflix, banking, email, cloud storage, and AI combined. AI is a fast-growing slice of that slice. So if you deleted every AI query on Earth tomorrow, you would not measurably change Arctic sea ice. Aviation, cement, steel, and cars each dwarf it. Anyone telling you your chatbot habit is killing bears is off by orders of magnitude.
Now the other direction, because the industry's numbers deserve the same scrutiny. Google reported in August 2025 that a median Gemini text prompt uses 0.24 watt-hours and 0.03 grams of CO2e — about nine seconds of TV. That figure is real but carefully bounded, and researchers said so immediately. It excludes training entirely. It covers text only, not image or video generation, which cost far more. "Median" hides a long tail of expensive queries. It uses market-based carbon accounting — crediting renewable energy purchases — which lowers the reported footprint by roughly two-thirds versus location-based grid averages. And Google doesn't publish total query volume, so nobody outside can multiply up to a real total.
That's why estimates diverge so wildly: people are drawing different boundaries and reporting the result as if it were one fact. Per-prompt or per-data-center. Training or inference. On-site water or the water evaporated at the power plant. Market-based or location-based carbon. Every one of those choices swings the answer by multiples, and almost nobody states which they picked.
The defensible summary: AI is currently a small contributor to a very large problem, and it is the fastest-growing one. The IEA projects data centers to roughly double to 945 TWh by 2030 — near Japan's entire consumption — while still staying under 1.5% of energy-sector emissions through 2035. Both halves matter. The honest worry isn't your prompt. It's a specific power plant getting kept open for a specific data center, on a specific local grid.
An example that makes it click
Imagine a lake being filled by a hundred fire hoses. Someone hands you a drinking straw and asks whether your straw is flooding the valley. Technically, water leaves your straw and reaches the lake. The chain is real. The contribution is a rounding error.
The catch: this straw is doubling in size every few years, and nobody's turning off the hoses. So the correct answer isn't "straws don't matter" and it isn't "put down the straw, you monster." It's: keep your eye on where the hoses are, and notice that the straw is on a trajectory to become a hose. That's an argument about power plants and grid planning — not about whether you should feel guilty asking a chatbot for a recipe.
Key facts
- Data centers consumed around 415 TWh in 2024 — about 1.5% of global electricity — growing ~12% annually since 2017; the US accounts for 45%, China 25%, Europe 15% (IEA, Energy and AI, 2025).
- Data center electricity use is projected to more than double to about 945 TWh by 2030, roughly Japan's total consumption today, and around 1,200 TWh by 2035 in the IEA Base Case.
- Data center CO2 emissions are about 180 million tonnes today, projected to reach 300 Mt by 2035 (Base Case) or 500 Mt (Lift-Off Case) — remaining below 1.5% of total energy-sector emissions.
- Google reported in August 2025 that a median Gemini text prompt uses 0.24 Wh, emits 0.03 gCO2e, and consumes 0.26 mL of water — with energy per prompt falling 33x and carbon 44x over the prior 12 months.
- Google's per-prompt figure excludes model training, covers text prompts only, and uses market-based carbon accounting that lowers the reported footprint by roughly two-thirds versus location-based grid averages; Google does not disclose total query volume.
- Hugging Face researcher Sasha Luccioni has called for a standardized 'AI energy score' comparable to Energy Star ratings, noting Google's report is not a substitute for standardized comparisons.
▶ The 60-second explainer (script)
Does AI harm polar bears? Indirectly, and very slightly — along with every other use of electricity. Here's the honest version. The chain is real: polar bears hunt from sea ice, sea ice melts as the Arctic warms, warming tracks cumulative CO2, and AI runs on electricity that's still largely fossil-fired. So the question isn't whether — it's how much. The IEA says all data centers used about 415 terawatt-hours in 2024. That's roughly one and a half percent of world electricity, and that includes Netflix, banking, email, and cloud storage — AI is a slice of that slice. Delete every AI query on Earth tomorrow and you would not measurably change Arctic sea ice. Aviation, cement, and steel each dwarf it. Now the other side. Google says a median Gemini text prompt uses 0.24 watt-hours — nine seconds of TV. True, but carefully bounded. It excludes training entirely. Text only, not images or video. Median hides the expensive tail. And it uses market-based carbon accounting, which cuts the reported number by about two-thirds versus the actual grid. That's why estimates disagree so wildly — people draw different boundaries and report it as one fact. The defensible summary: AI is a small contributor to a very large problem, and it's the fastest-growing one. Doubling to 945 terawatt-hours by 2030. The thing to watch isn't your prompt. It's which power plant stays open for which data center.
What authoritative sources say
People also ask
Should I feel guilty about using ChatGPT?
Not on climate grounds. A median text prompt is roughly nine seconds of television. Your commute, your thermostat, and one flight each dominate a lifetime of chatbot use by a wide margin.
So the 'AI is boiling the planet' headlines are wrong?
Overstated, usually by comparing AI's total to an individual's footprint, or by quoting training costs as if they recurred per query. The growth rate is genuinely concerning; the current level is small.
Why do water numbers vary so much?
Boundaries. On-site cooling water is a small number. Add the water evaporated at the power plant generating the electricity — off-site, and usually excluded — and it multiplies. Both are reported as 'AI's water use'.
What actually threatens polar bears?
Sea ice loss driven by cumulative greenhouse gas emissions across the whole economy — energy, transport, industry, agriculture. AI is a small and growing share of the energy slice, not a distinct threat.
What's worth watching instead?
Local grid decisions: whether a fossil plant is kept online or a new one built to serve a data center, and whether new demand is matched with new clean generation. That's where AI's emissions are actually decided.
The same question, asked other ways
- How does AI affect polar bears?720/mo