Comparisons and myths

Does ChatGPT use a bottle of water for every email it writes?

The claim: “ChatGPT drinks a bottle of water for every email it writes.”

Nobody has published it for today’s models. The best-known study (Li et al., UC Riverside) estimated GPT-3 used a 500 ml bottle for every 10 to 50 medium-length answers: 10 to 50 ml each. Company figures are far lower: OpenAI says about 0.32 ml per average question and Google 0.26 ml per median prompt, and neither says clearly whether it counts water used at power plants. A newer figure for an AI-written email also circulates, but Hum couldn’t read it on its page, so it isn’t used.

The bigger pictureWater is felt locally. In a recent year, US data centres used about 17 billion gallons on site, and about 211 billion more through the power plants that supplied them (Berkeley Lab estimates).

ml per answer
  • 10 to 50 ml the study’s estimate per answer, GPT-3
  • 500 ml what the claim says: a whole bottle
Why the claim misleads

The study measured an older model and spread one bottle over 10 to 50 answers. The retold version drops both.

What nobody knows yet
  • No independent per-question water figure exists for current ChatGPT models.
  • The companies don’t say clearly whether they count water used at power plants.
  • The newer email figure is in a news article Hum couldn’t open.
Sources (5)

Hum’s sums

  • 500 ml ÷ 50 answers = 10 ml
  • 500 ml ÷ 10 answers = 50 ml
  • 0.000085 gallons × 3,785 ml ≈ 0.32 ml
Cite this

Li, Yang, Islam, Ren (UC Riverside), Making AI Less Thirsty (GPT-3), as quoted in Hum, “Does ChatGPT use a bottle of water for every email it writes?”, https://humearth.org/q/bottle-of-water (quote checked 18 Oct 2026).

Cite the original source first; Hum is where you found it.

What do Wh, kWh and TWh mean?

Wh is a watt-hour: a 1-watt light left on for an hour. 1 kWh = 1,000 Wh, and 1 TWh = 1 billion kWh. Hum compares each number with something everyday so you don’t need the units.

Send this answer

“A bottle of water per ChatGPT email”: the study behind it said one 500 ml bottle per 10 to 50 answers, for GPT-3. For today’s models, nobody has published it.

Next questions