Key Voices in the Debate
A who's-who of the AI-water discourse, with each voice's position summarised in 2–4 sentences. Useful for: knowing who to cite, knowing who is overcited, and noting where the strongest critics and defenders actually agree.
Researchers
Shaolei Ren — UC Riverside
Lead author of the 2023 Making AI Less "Thirsty" paper that started the modern discourse. Position: AI water use is real, regionally variable, and badly disclosed. Has publicly distanced himself from the "500 mL per query" framing of his work. Policy ask is better measurement and transparency, not "stop using water." The most-cited and most-misrepresented voice in the field.
Pengfei Li / Jianyi Yang / Mohammad Islam
Co-authors with Ren. Less publicly visible but co-signatories on the central methodology.
Defenders / counter-voices
Andy Masley
Independent writer (former physics teacher), runs andymasley.com. Author of "The AI water issue is fake" — the single most thorough debunking of the inflated framings, including a forensic critique of the Empire of AI book. Position: national-scale AI water use is a non-issue (~0.008% of US freshwater); local stress is real and should be regulated; the discourse is mostly innumerate. Heavily reliant on Ren's actual numbers, just read correctly.
Sean Goedecke
Software engineer / writer. Published "Talking to ChatGPT costs 5 mL of water, not 500 mL" (Oct 2024) — the cleanest independent re-derivation of the per-query figure using Ren's own methodology with corrected assumptions. Position: the original 500 mL claim was a unit-conversion error compounded by an outdated model baseline; realistic per-conversation figure is ~5 mL.
Sam Altman — CEO, OpenAI
Posted in The Gentle Singularity (Jun 2025) that a typical ChatGPT query uses 0.000085 gal (0.32 mL) of water and 0.34 Wh of electricity (DCD writeup). Has called external water-alarm claims "completely untrue, totally insane" (CNBC, Feb 2026) while acknowledging that energy demand is the real concern and pushing publicly for nuclear / wind / solar. Position: water is a non-issue at per-query scale; energy is the real problem and is solvable with clean generation. Self-interested, unaudited, but the OpenAI numbers are now the de facto industry baseline.
Simon Willison
Influential independent technologist / blogger. Cross-posted Masley's piece favourably (Oct 2025). Useful as a credibility multiplier — he is broadly trusted in the technical community and his endorsement signals the argument has crossed from "advocate niche" to "informed default."
Critics
Karen Hao — author, Empire of AI (2025)
Long-form journalist whose book became the dominant mainstream reference for the AI-as-environmental-crisis case. Per Masley's analysis, the book contains at least one water-related figure that is wrong by a factor of 1,000, plus the standard 500-mL-per-query misreading. Position: AI infrastructure represents a colonial-style resource extraction project. Substantively contested on the numbers but rhetorically dominant in the broader discourse.
Washington Post (2023, "A bottle of water per email")
Not a person, but the single most-cited journalistic vector for the inflated 500 mL framing. The piece's calculation required five separately questionable assumptions to hold simultaneously. Still routinely cited.
Bloomberg / NYT / Guardian environment desks
Have run multiple pieces echoing the inflated framing. Quality varies; the Guardian in particular has been more careful than many to distinguish per-query from aggregate, and to flag the local-stress angle separately.
UC Riverside news office
Worth flagging: the press release accompanying Ren's own paper used the "500 mL" framing in a way that arguably contributed to the misreading. The paper itself is more careful than its own publicity.
Industry & official
Microsoft
Largest single source of voluntary water disclosure in the industry. 2022 sustainability report kicked off public scrutiny by reporting a ~34% YoY jump in water use coinciding with Azure's OpenAI buildout. Has since committed to "zero-water-consumption" cooling for new sites starting 2024.
Publishes per-data-center water use in annual environmental report. 2023 disclosure: 6.1 B gal of water (operational); has shifted siting policy to disfavour water-stressed regions.
IEA
The 2024 Electricity 2024 and 2025 data center reports are the most-cited official aggregate-energy projections; water disclosure in those reports is thinner but improving.
NREL / Macknick et al.
Authors of the canonical NREL TP-6A20-50900 review of per-source water consumption and withdrawal factors. Their median values for recirculating cooling configurations (coal 479 gal/MWh, gas CC 205, nuclear 672, solar PV 1, wind 0) anchor every credible scope-2 calculation in this wiki.
Where the strongest voices on each side actually agree
Reading Masley, Goedecke, Ren, and the more careful critics together, there is a substantial shared core:
- The viral 500 mL figure is wrong.
- Per-query water use is small in absolute terms (5 mL is the conservative high; 0.3 mL is the corporate low).
- National-scale aggregates are small but growing.
- Scope-2 dominates and is poorly disclosed.
- Local stress in arid siting regions is real and warrants regulation.
- Decarbonising the grid is the largest single water-savings lever.
The disagreement is about emphasis and rhetoric, not arithmetic. That itself is the editorial's strongest setup.