Thirty-one floor plans. Three AI tools. Five climate zones ranging from hot-arid Phoenix to cold-continental Moscow. A Turkish architecture researcher named Tuğçe Çelik fed climate-adaptive prompts into ChatGPT, Microsoft Copilot, and LookX—an architecture-specific generative model—and asked them to produce sustainable housing plans.
Of the thirty-one outputs, twenty-three were too incoherent to reconstruct. LookX, the only tool marketed specifically for architects, couldn’t produce visible windows.
The eight surviving plans were rebuilt in AutoCAD and run through Velux Daylight Visualizer on equinox and solstice dates. The peer-reviewed results, published in Cambridge University Press’s AI EDAM journal in July 2025, were blunt: “None of the models consistently integrated solar orientation or seasonal lighting considerations.”
The AI drew rooms. It didn’t know where the sun was.
What Orientation Actually Costs
Orientation is the cheapest design decision in residential construction. It costs nothing—literally zero additional materials, zero additional labor. You just rotate the plan on the lot. Homes re-oriented to face south without any additional passive solar features save 10–20% on heating, according to the Building Performance Association. Optimized designs with proper fenestration push that to 40%.
A Kabul climate study published by SCIRP in 2025 quantified the fenestration math across orientations and glazing types. South-facing windows with a 25% window-to-wall ratio and single glazing achieved 25.7% energy savings versus a windowless façade. Double glazing at 35% WWR: 35.2%. Triple glazing at 55% WWR: 36%. North, east, and west-facing windows at the same ratios increased total energy demand.
The counterintuitive finding: double-glazed south-facing windows outperformed triple-glazed windows in other orientations. You get more from putting cheap glass in the right wall than expensive glass in the wrong one.
The Dollar Math for a 2,000-Square-Foot Home
The EIA’s Residential Energy Consumption Survey puts average annual heating cost for a U.S. single-family home at roughly $960. In climate zones 4–6 (the northeast corridor, upper Midwest, Mountain West), it’s closer to $1,200.
If proper orientation saves a conservative 15% on heating—splitting the difference between the BPA’s 10–20% range—that’s $144–$180 per year. Call it $160.
Over a 30-year mortgage at 3% annual energy cost inflation:
| Scenario | Year 1 Savings | 30-Year Cumulative |
|---|---|---|
| Conservative (10% heating reduction) | $96–$120 | $4,500–$5,700 |
| Moderate (15% reduction) | $144–$180 | $6,800–$8,500 |
| Optimized with passive fenestration (25%+) | $240–$300 | $11,400–$14,200 |
Calculation: cumulative cost = Year 1 savings × [(1.0330 − 1) / 0.03]. That multiplier is 47.58. The moderate scenario: $144 × 47.58 = $6,851. For $180: $8,564.
A floor plan generated in three seconds, placed on a lot without considering compass bearing, can cost the homeowner $7,000–$8,500 over the mortgage. For a subdivision of 200 homes, that’s $1.4–$1.7 million in collective energy waste. The builder saved maybe $2,000 per home by not hiring an architect to think about orientation. The homeowner pays it back four times over.
Where the Tools Actually Stand
Çelik’s study tested text-to-image diffusion models—ChatGPT, Copilot—not purpose-built architecture platforms. Fair point. So where are the dedicated tools?
Higharc, the highest-profile AI homebuilding platform, replaces static plans with generative digital models. Buffington Homes reports 15 days faster in their soft schedule and attributes $10 million in additional 2023 revenue to the speed gains. The marketing material mentions design automation, cost estimation, and sales customization. Solar orientation, lot-specific fenestration optimization, climate zone-responsive window placement? Absent from the feature list.
Maket generates residential floor plans from text descriptions. Its marketing positions AI as replacing the expensive early-stage architect. What it doesn’t replace is the architect’s understanding that the same plan facing north versus south is two different buildings from an energy perspective.
Cove.tool does perform orientation-aware energy analysis—but it’s a downstream performance simulation platform, not a generative design tool. It evaluates plans after they’re drawn. The problem is upstream: the AI generates the plan without orientation awareness, the builder stamps it on 200 lots, and nobody runs the energy simulation because the plan was “AI-optimized.”
Production Building Was Already Doing This Wrong
Production homebuilders were ignoring orientation long before AI entered the conversation. The economics of tract housing favor one floor plan stamped across every lot in the subdivision, regardless of compass bearing. The framing crew knows the plan. The electrical rough-in is identical. The permit application is a duplicate. Rotating the plan 90° for a south-facing lot means redrawing, resubmitting, and retraining—and nobody’s paying for that.
AI tools didn’t create this problem. But they’re encoding it into software and calling it innovation. When a builder used to stamp the same plan on every lot, at least nobody was claiming the plan was “optimized.” When AI generates it, there’s an implicit authority—the machine considered things. It didn’t. It generated a spatial arrangement that looks like a house. It has no concept of cardinal direction.
What Would Fix This
Technically, it’s not that hard. An orientation-aware AI plan generator would need three inputs the current tools don’t use: lot compass bearing (available from any GIS parcel database), climate zone (IECC lookup by ZIP code), and solar path data (freely available from NOAA or NREL). With those three inputs, an algorithm can weight window placement, room positioning, and wall-to-glazing ratios before the plan is generated—not after.
The 2024 IECC already tightened the screws. The Simulated Building Performance path now demands 15–20% energy cost savings over the reference design, up from 5% in the 2021 code. That’s a fourfold increase in the performance threshold. As states adopt the 2024 code, orientation-blind plans will face a harder time clearing energy compliance.
The data is free. The computation is trivial. The reason it doesn’t exist in these tools yet is that no one is asking for it—because the builders buying the software never optimized orientation in the first place, and the homebuyers don’t know what they’re losing.
What This Analysis Didn’t Prove
The Çelik study tested text-to-image diffusion models, not purpose-built architectural generators like Higharc or TestFit. It’s possible—though not demonstrated—that dedicated tools handle orientation better. We weren’t able to independently test Higharc’s or Maket’s outputs for orientation awareness because neither offers a free trial with exportable plan data.
The energy cost estimates use national averages from the EIA. A 2,000-square-foot home in Tucson and one in Minneapolis will see dramatically different returns from orientation optimization—the heating savings matter most in climate zones 4–7, and less in hot-dry climates where cooling dominates and east/west shading matters more than south-facing solar gain.
The 10–20% heating savings figure comes from passive house research, which represents best-case design intent. A code-minimum production home with standard windows will capture less of the theoretical savings than a Passive House-certified build. We used 15% as a conservative midpoint, but the actual capture depends on glazing quality, overhang design, and thermal mass—variables no AI plan generator currently considers.
Finally, orientation is one variable among many. Shading from adjacent structures, tree canopy, setback requirements, and street grid geometry all constrain how much a plan can be rotated in practice. In dense infill lots, the compass bearing may be dictated by the parcel, not chosen by the designer.