A water-per-prompt number needs a workload, a date and an accounting boundary. The unit alone cannot tell you what was counted.
THE SHORT ANSWER
Published figures can describe different models, tasks, locations and statistics. Some count cooling water; others also count water used to supply electricity. A smaller number alone does not establish a smaller total footprint.
This is a dated audit of three original claims, checked 16 October 2026. Use it to inspect a claim's source, then ask for the missing details. It is not an estimate of your own AI use.
Three claims. Three sets of boundaries.
01 / MODELLED SCENARIO
Li and colleagues: GPT-3
500 mL across roughly 10–50 medium responses
Version / workload
Paper v5, 26 March 2025. Modelled GPT-3 (175B) requests: up to about 800 input words and 150–300 output words.
Water counted
On-site cooling plus location-based electricity-generation consumption. Manufacturing water excluded from this case study; training calculated separately.
What to watch
The original April 2023 v1 said 20–50 exchanges. Both headlines describe a group of responses, not a 500 mL measurement for one prompt. Results depend on energy assumptions and deployment location.
0.26 mL reported for the median text prompt, May 2025
Publication / statistic
Paper v1, 21 August 2025. A prompt-weighted median of model-level average serving energy, converted to estimated water. It is not individual prompt water metering.
Water counted
On-site freshwater consumption, using 2024 water-use effectiveness. The stated equation has no electricity-generation or manufacturing-water term; this narrower water boundary is inferred from its method.
What to watch
The serving energy method includes CPU/RAM, idle machines and overhead. That broader energy method does not make the water figure a full lifecycle total. Training is excluded.
Original Google-authored study ↗ · §§3.1–3.4, Table 2. We omit its narrower comparison case because the table and prose disagree.
03 / ATTRIBUTED CLAIM
Sam Altman: ChatGPT query
0.000085 gallons per “average query”, reported 10 June 2025
Workload / statistic
The post supplies no model, task length, sampling period, sample size or definition of average.
Water counted
The post does not disclose cooling, electricity or manufacturing boundaries, or whether this is withdrawal or consumption. Undisclosed does not mean excluded.
What to watch
That is about 0.32 mL if US liquid gallons are assumed; the post does not specify the gallon system. Converting the unit does not resolve the missing method.
Look for these five details in the original source. Mark the ones you can actually find. A complete checklist makes a claim easier to inspect; it does not prove the number is correct or comparable.
Can these figures rank providers?
These sources do not hold the workload, reference period, geography, statistic and water boundary constant. Their numerical difference cannot tell us how much came from hardware, task length or accounting choices. A cross-provider ratio would imply a comparison these sources do not establish.
Lower water per task can also coexist with growing total use if activity grows. Neither a per-prompt figure nor a global annual volume settles the impact on a particular water supply.
Independent, non-exhaustive source audit. Claims are attributed to their authors; missing disclosure is not proof a claim is false. Source review: 16 October 2026. No universal personal-prompt coefficient is established here. Corrections record.