Your Builder Optimized Your Windows for the Energy Code. Nobody Checked What They'd Do to Your Brain.

Stand in front of a new subdivision anywhere in the American Sun Belt and count the window shapes. Flush vinyl rectangles pushed nearly to the surface of the stucco, no muntins, no hierarchy, no depth. Horizontal ribbons in the living areas. Maybe a randomly placed triangular accent near the roofline because the architect watched a Zaha Hadid documentary once. Every one of those windows was optimized for U-factor, solar heat gain coefficient, visible transmittance, and air leakage rate, then approved by an inspector who verified the energy code was satisfied, signed off, and moved on to the next house without ever considering what those glass openings were doing to the autonomic nervous system of every person who walked up to the front door.

A study published in May 2025 by Nikos Salingaros at the University of Texas at San Antonio (a mathematician, not an architect, which matters) used ChatGPT 4.5 and o3 to evaluate ten pairs of window designs against two independent frameworks: Christopher Alexander's fifteen fundamental properties of living geometry, and a separate set of ten emotional descriptors including calmness, coherence, comfort, and well-being. Drawing on the vast peer-reviewed literature about human visual perception and environmental psychology embedded in its training data, the AI assessed each pair using both criteria. Unanimous, across all ten pairs: both the geometric and emotional evaluations preferred the same designs.

The preferred windows were not expensive, not complicated, and not novel. Vertical rectangles with proportions between 1.4:1 and 2:1 height-to-width, thick frames with deep reveals, symmetrical arrangements with lintels aligned and sills consistent across the facade. They looked like the windows people installed for five centuries, and that builders stopped installing about thirty years ago when energy codes, manufacturing economics, and modernist aesthetics converged on a different set of priorities that had nothing to do with the occupant's neurological response to the geometry of their own home.

Then Salingaros asked the AI to identify the contemporary window typologies that scored highest for anxiety, and it returned seven. Every one is a current industry standard: unitized flush curtain-wall panels, horizontal ribbon windows, frameless triangular apertures, tilted diagonal picture windows, rotated parallelogram windows, barcode ribbons with random mullion spacing, and diagonal scatter-punched openings. These are the windows going into new homes, schools, offices, and hospitals worldwide, in volumes measured in billions of square feet per year, and they are, according to the combined geometric and emotional analysis, the windows most likely to produce heightened anxiety, negative emotional responses, and chronic low-grade stress in the people who live with them.

The Regulatory Blind Spot

According to the U.S. Department of Energy, heat gain and loss through windows account for 25 to 30 percent of residential heating and cooling energy use. For the average American household spending roughly $2,000 a year on energy, with approximately 43 percent going to heating and cooling, that translates to $215 to $258 flowing through the glass annually, and this is the number that the entire window industry, from ENERGY STAR certification to Title 24 compliance in California to the IECC requirements governing residential construction in 49 states, is organized around reducing.

It is, by every measurable standard, the wrong number to optimize in isolation.

Building codes enforce energy performance, structural safety, fire resistance, egress dimensions, and minimum natural light. They do not enforce anything related to how the geometry of a window affects the cortisol levels of the person looking at it. Salingaros writes: "Codes enforce energy and structural safety, not neurophysiology, leaving a huge regulatory blind spot."

This blind spot is not theoretical, and the biometric literature that Salingaros's AI drew upon proves exactly how concrete it is. EEG studies, galvanic skin response measurements, heart-rate variability monitoring, and eye-tracking research constitute decades of accumulated data showing that specific geometric qualities in the visual environment produce measurable physiological stress responses in the people exposed to them. Diagonal and asymmetrical window placements trigger alarm responses in the autonomic nervous system. Flush, frameless glazing eliminates the boundary cues that the visual cortex uses to parse figure from ground. Horizontal ribbons violate the gravitational expectations that the vestibular system constantly calibrates against, and the vagus nerve connects all of this visual input to cardiac control, digestive function, and stress-hormone regulation through pathways that evolved over millions of years and do not care what your window's U-factor is.

What AI Optimization Misses

Meanwhile, the AI tools that architects and builders actually use for window design are getting extraordinarily good at what they do. A 2026 study in Scientific Reports by Adel Nasab and Rabiei at Bauhaus University and Leibniz University built an artificial neural network surrogate model that predicts window energy performance with R-squared values exceeding 0.99 (essentially perfect accuracy) at a speed of less than one-hundredth of a second per evaluation, compared to the 50 to 65 seconds required for a full EnergyPlus simulation. Coupled with a multi-objective evolutionary optimization algorithm called NSGA-III, the system generates Pareto-optimal window configurations that simultaneously balance heating demand, cooling demand, and ASHRAE 55 adaptive thermal comfort.

Brilliant. Three objectives. Not one of them is "will this window make the occupant feel safe."

Here is what makes this particularly frustrating, and what constitutes this article's original contribution: the energy-optimal and psychologically-optimal window specifications are not in conflict. The Adel Nasab framework found that intermediate window sizes, roughly 20 to 40 percent window-to-wall ratio, provide the best energy-comfort balance, while Salingaros's ten rules for empathetic windows specify a 25 to 40 percent solid-void ratio on the main facade, meaning the two independently derived optimization ranges overlap almost perfectly. Both frameworks favor vertical window proportions. Both penalize extremes. The science does not contradict itself, and yet nobody has built the optimization tool that incorporates human neurological response alongside thermal performance and energy demand, which means that every AI-generated window schedule currently in use is solving an incomplete version of the problem while ignoring variables that the biometric literature suggests matter as much to the occupant's quality of life as the number on their utility bill.

Ten Rules That Cost Nothing Extra

Salingaros asked ChatGPT o3 to synthesize the geometric and emotional findings into practical design rules, and the AI derived them from neuroscience and psychophysiology rather than from photographs of old buildings, a distinction the paper takes care to establish by directly asking the model where the rules came from. When pressed on whether the rules simply reproduced traditional window design, the AI responded: "The outcome 'looks traditional' because millennia of empirical building practice converged, often unintentionally, on geometries that human biology finds legible, calming, and socially intelligible. The pipeline was theory to biometric data to design rule. Traditional precedents merely happen to satisfy the rules; they were not the data source."

The ten rules, condensed for a builder's spec sheet: vertical rectangles with height-to-width ratio of 1.4 to 2.0; nested panes or muntins creating three to five levels of visual hierarchy within each opening; frames at least one-fifth to one-third the width of the glazing, with reveals deeper than 75 millimeters; bilateral symmetry within each bay, with lintels and sills aligned horizontally across bays; rhythmic alternation along the facade (narrow, wide, narrow) rather than monotone repetition; glazing recessed 50 to 150 millimeters into the wall, creating a graded shadow band; window-to-wall ratio of 25 to 40 percent; micro-scale texture on frame surfaces such as beveled muntins and shadow lines rather than flush extrusions; color or material contrast between frame and wall without mirror-like glare; and sill height positioned so that seated eye level, approximately 1.5 meters, aligns with the lower third of the glazing when viewed from outside.

Of these ten, at least six require zero additional materials cost. Vertical proportions, symmetrical placement, aligned lintels, rhythmic alternation, appropriate window-to-wall ratios, and sill height are decisions made on the architectural drawing, not at the window factory. Deep reveals cost more than shallow ones, and real divided-light muntins cost more than applied grilles or bare glass, but the cost differential is measured in tens of dollars per window, not hundreds, and it buys something that no energy rating can quantify: a facade that the human nervous system recognizes as safe.

The Strongest Counterargument

The Salingaros study used ChatGPT as its analytical engine, and this deserves honest scrutiny rather than reflexive acceptance. Large language models do not experience architecture in any meaningful sense; they synthesize textual patterns from training data, and the paper's claim is not that the AI "felt" anything about the windows but that it accurately retrieved and weighted the findings of thousands of peer-reviewed biometric studies, a claim that is reasonable for text-based analytical synthesis but considerably weaker for direct perceptual judgment. The paper itself acknowledges this: "Unlike direct neurological or physiological measurements, such as biometric sensors or eye-tracking studies which register emotional and physiological states, ChatGPT's emotional evaluations are secondhand abstractions. Ideally, therefore, these AI-generated estimations for the beauty-emotion cluster should be complemented by empirical validation."

The ten schematic window pairs were also deliberately simplified into austere line drawings rather than photographs of real buildings, isolating first-order geometric variables from confounding factors like material, color, and urban context. This was methodologically intentional, but it is a limitation that matters for residential applications, because a flush aluminum window in a well-proportioned modern home surrounded by mature landscaping and warm-toned materials may produce a different stress response than the same window in a bare stucco subdivision, and the study did not test that difference.

None of this invalidates the underlying biometric literature, which is extensive, replicated across multiple sensing modalities, and entirely independent of AI. What the study adds is a framework for connecting that literature to specific, actionable design rules, and whether the framework survives replication with human subjects in real architectural contexts is an open question worth answering before municipalities consider writing neurophysiology into building codes.

What This Means for Your Project

If you are building or renovating a home and your architect hands you a window schedule, ask one question that nobody in the approval chain is currently required to ask: what do these windows look like from the sidewalk? Not the U-factor, not the SHGC, but the proportions, the depth, the rhythm, the symmetry. These are free variables that cost approximately the same as the alternatives and are not currently optimized by any commercially available AI design tool, yet the evidence from neuroscience, environmental psychology, biometric sensing, and AI-driven analysis of that combined literature increasingly suggests they matter as much to the people living in the house as the number on the energy bill.

Your builder has a spec for every measurable property of your window glass. The one property nobody specs is whether looking at it makes you want to come home.

Limitations

This analysis relies primarily on two studies, the Salingaros AI evaluation and the Adel Nasab energy-optimization framework, which were not designed to be cross-referenced, and the overlap between their recommended window-to-wall ratios (25 to 40 percent for psychology, 20 to 40 percent for energy) is suggestive but has not been validated in a combined optimization that accounts for both domains simultaneously. Neither study used real residential facades in situ. The biometric literature on facade perception, while extensive, disproportionately samples European and North American observers, and cultural variation in window preferences remains understudied. Energy cost calculations use national DOE averages and will vary significantly by climate zone, fuel mix, and local utility rates. The Salingaros study's AI methodology has not yet been independently replicated by a second research team.