AI water usage versus everyday things

Published estimates for an ordinary AI text prompt range from about 0.26–0.32 millilitres in recent operator disclosures to roughly 10–25 ml in an earlier academic model. Against a 500 ml bottle, that is around 1,560–1,900 prompts at the lower figures, or 20–50 responses under the research scenarios.

These comparisons describe volume, not equal environmental impact. The AI figures allocate water used by shared cooling and electricity systems; a bottle is direct product volume. Location, scarcity and accounting boundary matter more than the mental image. Use everyday comparisons to understand scale, not to declare an exact water cost for your message.

Start with the three published estimates

Google measured 0.26 ml for a median Gemini Apps text prompt in May 2025. Sam Altman stated an average ChatGPT query used 0.000085 US gallons, about 0.32 ml, without publishing the method.

The 2023 paper Making AI Less “Thirsty” gave the higher comparison: 500 ml for roughly 20–50 medium-length responses, or around 10–25 ml each under its modelled locations and times.

Do not average these into a neat middle. They cover different products, infrastructure and boundaries. A range that retains its labels is more informative than a blended number no source actually found.

Compared with a 500 ml bottle

The arithmetic is simple. Dividing 500 ml by Google’s 0.26 ml gives about 1,923 median prompts. Dividing by Altman’s 0.32 ml equivalent gives about 1,563 stated-average ChatGPT queries. Rounding appropriately, that is roughly 1,900 and 1,560.

Under the academic estimate, the authors already supplied the comparison: 20–50 responses per bottle. The gap between 20 and 1,900 is not normal measurement noise. It reflects different systems, assumptions and definitions.

This is why the bottle-per-message claim is false. Even the high research scenario put one bottle across many responses. The low provider disclosures put it across well over a thousand ordinary text prompts.

Compared with a shower

The US Environmental Protection Agency says standard showerheads use 2.5 gallons per minute, about 9.5 litres, while WaterSense-labelled models use no more than 2.0 gallons per minute. One minute of a standard shower is therefore about 9,500 ml of direct household water flow.

That volume is far larger than any of the ordinary-text prompt figures above. But turning it into an exact number of “AI prompts per shower” would amplify incomparable assumptions. A shower measure is withdrawal at the tap; AI water can include allocated consumption at a cooling tower and power station. One may occur in a different catchment.

The useful conclusion is order of magnitude: an individual short text prompt is small compared with everyday household water flow. Scale and location keep it from being irrelevant.

Compared with a tap or drink

A 250 ml glass is a defined example rather than a claim about how much everybody drinks. It contains about 780 of Altman’s stated-average 0.32 ml queries or about 960 of Google’s 0.26 ml median prompts. Under the academic range, it corresponds to around 10–25 responses.

Again, the ratios are arithmetic illustrations. They do not show what happened at a specific facility. A long reasoning request, generated image or tool-using task may not resemble the text prompts behind the low values.

Direct household use is also easier to observe. The allocated water for digital infrastructure depends on cooling, electricity and shared capacity. Our guide to why data centres use water explains that system.

Volume is not the same as water impact

Water withdrawal counts what is taken from a source; consumption counts what is not promptly returned, often because it evaporates. A shower sends water into a wastewater system. A cooling tower releases it to the atmosphere. Both have treatment and infrastructure effects, but the pathways differ.

Source and scarcity matter. Reclaimed water used for cooling has different social trade-offs from potable supply. Consumption during drought can be more damaging than the same volume in a wet period. Indirect power-generation water may occur far from the data centre.

This is why “water equivalent” should not become “environmental equivalent”. A simple litre ratio cannot capture catchment stress, quality, timing or ecological effect.

Small per-use figures can create large totals

It is tempting to look at 0.26 ml and conclude there is no issue. At the level of one ordinary prompt, the allocation is indeed small. At service scale, messages are numerous, workloads vary and facilities support more than text chat.

System demand also includes training, idle capacity, non-AI cloud services and hardware supply chains. Berkeley Lab’s US report models data-centre water in aggregate because local infrastructure must serve the total, not an idealised median prompt.

The reverse temptation is to multiply the highest old scenario by every imagined future query. That assumes all requests share its model, location and boundary. Honest planning uses scenario ranges and updates them as measured data improves.

Make comparisons without being misled

Whenever you see an AI-versus-household ratio, check the source year, product, statistic and task. Ask whether the water is direct, indirect, withdrawn or consumed. Preserve “median” and “average”. Avoid a ratio when the denominator is a vague “query”.

For personal use, choose a smaller capable model, keep context relevant and avoid disposable generations. For providers and organisations, absolute facility water, catchment stress and seasonal peaks are more actionable than a bottle graphic. See how much water ChatGPT uses for service-specific caveats.

Treechat’s methodology applies a fixed factor to estimated message energy, so its water receipt is modelled rather than measured. It is intended to make resource use visible, not to claim that an exact fraction of your drinking water disappeared. Everyday comparisons are most honest when they make uncertainty visible too.