What is the environmental impact of SRHD?

In architecture, we are used to thinking about impact in physical terms. We count embodied carbon. We worry about operational energy. We think about water, waste, maintenance and durability. So if we are going to use AI in our own work, it is only fair that we ask the same question of ourselves: what is the environmental cost?

That matters for Spec Rep Help Desk because digital services can feel weightless when they are not. AI runs in real buildings, filled with real servers, drawing real electricity and, in many cases, consuming real water for cooling. The difficulty is that there is no universally agreed single figure for the impact of one AI interaction. The footprint varies with the model used, the length of the exchange, the amount of computation involved, the efficiency of the data centre, and the electricity grid where the computing actually happens. For that reason, the most honest approach is not to pretend we know the exact footprint of every query. It is to model a reasonable range and explain the assumptions clearly.

For this exercise, the AI service is hosted in the United States. That means the electricity use, carbon intensity and data-centre water consumption are tied primarily to US infrastructure rather than to the local grid of the person using the tool. The comparison figures used to help readers interpret the results can still be familiar everyday references, but the underlying environmental footprint is driven mainly by the energy and cooling context where the AI is actually run.

We have therefore looked at three user scenarios and tested each against a likely case and a deliberately pessimistic case. The likely case assumes relatively efficient AI use and a typical US hosting context. The pessimistic case assumes heavier computation, more emissions-intensive electricity, and higher water intensity. It is not a prediction of what happens every day. It is a stress test designed to show how much the footprint can vary depending on underlying conditions.

Scenario 1 is a short conversation of five exchanges. In the likely case, that interaction is modelled at about 2 watt-hours of electricity, around 0.68 grams of CO2e, and roughly 13 millilitres of water. In the pessimistic case, the same five-exchange conversation rises to about 18.75 watt-hours, 18.5 grams of CO2e, and 178 millilitres of water. The point is not that every short AI chat has exactly this footprint. The point is that a short text interaction is usually environmentally small, but not literally weightless. If the underlying infrastructure is efficient, the footprint is very modest. If it is inefficient or emissions-intensive, it can rise materially.

Scenario 2 is a more involved conversation of 10 exchanges. In the likely case, that comes to about 4 watt-hours of electricity, 1.36 grams of CO2e, and 26 millilitres of water. In the pessimistic case, it rises to about 37.5 watt-hours, 37.1 grams of CO2e, and 356 millilitres of water. Again, these figures should be read as modelling outputs, not universal constants. What they show is that the environmental impact of AI scales with use. The longer and more computationally intensive the interaction becomes, the more efficiency matters.

Scenario 3 looks at a regular user having one 10-exchange conversation on roughly 230 working days per year, allowing for annual leave and public holidays. In the likely case, that annual use comes to about 0.92 kilowatt-hours of electricity, 0.31 kilograms of CO2e, and 5.9 litres of water. In the pessimistic case, it rises to about 8.63 kilowatt-hours, 8.53 kilograms of CO2e, and 81.9 litres of water.

To help put that in perspective, 0.92 kilowatt-hours is roughly equivalent to about 9 hours of use from a 100-watt desktop computer. 5.9 litres of water is about 39 seconds of showering at 9 litres per minute. And 0.31 kilograms of CO2e is roughly in the order of 2 kilometres of travel in a typical internal combustion passenger car. The pessimistic case is much higher, rising to the rough equivalent of about 86 hours on a desktop computer, about 9 minutes in the shower, and about 48 kilometres of driving. Even at that level, it is still modest compared with the environmental consequences of most building-level decisions.

That helps keep the issue in proportion. In the likely case, the footprint of a lightweight AI support tool is small. In the pessimistic case, it becomes more noticeable, but it is still not in the same league as the consequences of real project decisions. That is where the discussion becomes more interesting for architects.

It is also important not to treat all AI use as equivalent. A short text exchange is one thing. AI video generation is another. Very short AI-generated video can require far more electricity than a text interaction. So the environmental case for AI should not be judged in the abstract. It should be judged by the specific task being performed and whether the value created is worth the resources consumed.

Now consider the built environment context. Say a residential project involves 200 square metres of plasterboard lining, and a specification change achieves an embodied carbon saving of just 0.05 kilograms of CO2e per square metre. That is a total saving of 10 kilograms of CO2e. On that basis, one modest specification improvement on one modest project would outweigh many years of likely-case use of a lightweight AI support tool by one person, and would still exceed roughly a year of use even under the deliberately pessimistic case.

That is the real point. The environmental question around AI is not simply whether it has a footprint. It does. The more useful question is whether the footprint is justified by the value created. If a digital tool helps an architect identify even one small lower-carbon substitution, improve the quality of a specification, reduce waste, or avoid rework on a real project, the environmental saving from that decision may easily outweigh the impact of using the tool itself.

At Spec Rep Help Desk, that is the standard worth applying. Be honest that AI has a footprint. Be equally honest that, for a lightweight text-based tool, that footprint is likely to be modest at the level of an individual user. Then focus on the bigger test: whether the tool helps deliver better decisions in the built environment. Because if it does, the environmental ledger may be far more favourable than first impressions suggest.