Architecture and AI: The Algorithm Proposes, the Human Decides

Data centres consumed roughly 1.5% of global electricity in 2024, by the International Energy Agency's estimate (Energy and AI report, 2025); the figures and the 2030 scenario are in the article "The AI Race: Who Owns the Chips, the Energy and the Rules?". There is a paradox here that concerns architecture: the tools that promise a building using less energy rest on computers that themselves use a great deal of it.

From imitating form to imitating process

Architecture has long drawn on nature by copying the image: a dome like a shell, a façade echoing tree bark. Generative design is different. The designer defines goals and constraints, such as cost, light and heat; the algorithm generates many alternatives, tests them, and selects the most suitable in a manner close to natural selection. The guest editors of a special issue of the journal AI EDAM (Stouffs and Rafiq, 2015) described this approach as a dialogue between designers and computers, aimed at exploring possibilities and changing how the designer thinks, not at extracting a single optimal solution.

The point of decision therefore shifts. The architect no longer draws the solution and then tests it; he frames the question and judges the answers. Here lies the danger too: the algorithm optimises only what it was asked to. If we ask for the lowest cost and the highest thermal efficiency, it will deliver, and may pass over values that are hard to turn into a mathematical function, such as familiarity of place and collective memory. A home remains an experience of living, not an optimisation problem.

Efficiency on paper and in reality

Marketing talk holds that smart design means buildings that use less energy. Performance studies, however, record a well-known gap. In a paper presented at the CIBSE ASHRAE symposium in Dublin (de Wilde and Jones, 2014), measurements of completed buildings show 'significant gaps' between predicted and actual energy performance, which the authors attribute to occupant behaviour, weather, workmanship errors, control settings and modelling issues. The best calculated design does not guarantee equal performance after handover.

It may be hoped that systems linked to building sensors will narrow this gap by learning from occupants' behaviour and adjusting ventilation and lighting. That is a hope for which we found no decisive evidence, and it has a known price: a building that watches its occupants gathers intimate data on them. That is a privacy door to be settled at design time, not left to default settings, and we set out its principles in our article 'Privacy in the Age of AI'.

The cost of the tool itself

Sustainability accounting is incomplete if we count the building and neglect the tool. According to the International Energy Agency, accelerated servers, which dominate AI workloads, account for almost half of the net increase in data-centre electricity use through 2030, and their consumption grows about 30% a year in the Agency's scenario. This is not a reason to abandon the tool but to count its cost within the function we wish to optimise: from extracting the material to operation and demolition, and from training the model to running it.

A lesson from Najd

We do not start from nothing in the Kingdom. The At-Turaif District in ad-Dir'iyah was inscribed on the World Heritage List in 2010 as an outstanding example of the Najdi style, whose chief material is adobe (mud brick) and which features geometric decoration. An article by the engineering consultancy Cundall notes that adobe walls have high thermal mass, absorbing the day's heat and releasing it at night, and that thick walls, shaded openings and compact forms reduce cooling loads; the description is qualitative, not numerical.

The lesson is not nostalgia. If we trained an algorithm on buildings from temperate, rainy climates, it would likely propose solutions foreign to our hot, water-scarce climate unless local environmental data and values were put into its objective. The best we can offer the machine is to learn from our forebears' wisdom in responding to place, rather than to import its goals ready-made.

How to use it wisely

These are editorial suggestions based on the above:

• The machine proposes, the human decides: no design is accepted merely because the algorithm ranked it first; its logic must be explained (see our article 'Explainable AI: When the Decision Falls Silent').
• Put sustainability in the objective from day one, by counting the building's life cycle and the cost of the design tools.
• Measure after occupancy, since actual performance is the arbiter, not the model.
• Write values into the equation: local climate, family privacy, patterns of gathering, building traditions.
• Retrain the architect to frame questions well, for the quality of the output follows the quality of the question.

The deepest point is that the algorithm is a mirror of how clear or vague our aims are. If we choose well what we ask of it and what we refuse to hand over to it, we get the architecture we hope for: one that honours its occupant and respects the land and the place.