AI Generated Your Floor Plan in 11 Seconds. It Doesn’t Know Which Direction the Sun Rises.

A researcher at Ostim Technical University in Ankara asked three AI tools to draw floor plans for sustainable housing. She gave them generous prompts, specifying climate zones, passive solar strategies, and daylight requirements. Out of 31 generated plans, only eight were coherent enough to reconstruct in AutoCAD for simulation. One tool, LookX, produced outputs so illegible that every single plan was thrown out before testing could begin.

Of the eight survivors, not one consistently oriented living spaces toward the sun. Not one.

That study, published in AI EDAM by Cambridge University Press in July 2025, is the first rigorous attempt to measure whether AI-generated floor plans perform as architecture or merely resemble it. Tuğçe Çelik tested plans from ChatGPT, Microsoft Copilot, and LookX across five climate zones, then ran daylight simulations using Velux Daylight Visualizer on equinox and solstice dates. Her conclusion deserves quoting directly: AI tools exhibit “a gap between generative representation and environmental logic.”

They draw rooms. They do not understand light.

A Million Users, Zero Sun Paths

Maket, one of the leading residential AI floor plan generators, recently crossed one million registered users. Founded in Montreal in 2020 and backed by $3.4 million CAD in seed funding from Amiral Ventures, the platform lets you type room counts and square footage targets, and a generative model produces dimensioned layouts in seconds for $30 per month on a Pro account. Competitors include Planner 5D, Snaptrude, TestFit, and an expanding field of startups racing to automate what architects spend weeks doing manually.

These tools are genuinely useful for exploring layout variations quickly, and Maket in particular has added features like agentic editing, where you describe changes in plain language and the AI adjusts the plan in real time, and a regulatory assistant that parses uploaded zoning PDFs to answer setback and lot-coverage questions. At $30 a month versus $400 for a Revit license, the accessibility argument is obvious.

Not one of them simulates daylight, models solar orientation, or knows whether your kitchen faces west, which in Phoenix means your afternoon cooking happens in a greenhouse, or east, which in Seattle means your morning light disappears by 10 AM in November and does not come back until March. When Çelik’s study specifically asked AI tools to account for climate and passive strategies in the prompts, the resulting plans still ignored where the sun would be at any given hour of any given day in any given city, which means that explicit instruction made no difference.

What Orientation Actually Costs You

Building orientation is not an aesthetic preference. It is a measurable energy variable that the Department of Energy and ASHRAE have studied for decades. A home with its long axis oriented east-west, with primary glazing facing south in northern climates or north in southern ones, reduces heating and cooling loads by 10 to 20 percent compared to an identical floor plan rotated 90 degrees. In hot-arid climates, the penalty for wrong orientation climbs higher.

Run the numbers on an average American household. EIA data puts median residential energy expenditure at roughly $2,000 per year. A 15% penalty from suboptimal orientation means $300 annually in excess heating and cooling costs that would not exist if the floor plan had been drawn with the sun in mind. Over a 30-year mortgage, that is $9,000 in energy bills attributable to a design decision that took 11 seconds and never once consulted an ephemeris.

An architect would catch this, and so would a first-year architecture student, because orienting a building to its site is among the earliest lessons in any design curriculum, predating computers, predating CAD, predating mechanical air conditioning itself by centuries. Vitruvius wrote about it in the first century BCE, recommending that dining rooms face west to catch evening light and that winter apartments open to the south for warmth, and twenty-one centuries later the AI tools that a million people are using to plan their homes still treat the sun as if it does not exist, generating rooms in arrangements that bear no relationship to the cardinal directions or the latitude of the lot or the path the sun traces across the sky from the day the foundation is poured to the day the last owner sells the house.

What the Academic Literature Actually Shows

Çelik’s study is not isolated. A separate research team at Tongji University developed a GAN-based daylight performance predictor that can evaluate floor plan daylighting 267 times faster than traditional simulation methods, achieving a mean squared error of 4.2 and structural similarity index of 0.98 against ground truth. Remarkably accurate prediction. But the researchers built this tool precisely because existing AI floor plan generators produce layouts as flat images without the geometric and physical properties needed for performance evaluation. You cannot simulate daylight on a JPEG. The plans lack depth, material assignments, window specifications, and wall thicknesses, all information an architect includes instinctively but a generative image model has no framework to encode.

A comprehensive metrics paper published in Buildings (MDPI, May 2025) proposed dedicated evaluation criteria for AI-generated residential floor plans, covering spatial quality, circulation efficiency, room proportions, and adjacency logic. Their framework exists because no standard evaluation method existed before. We have been generating AI floor plans for years and only now developing ways to measure whether they are any good.

These are not fringe concerns. Cambridge University Press, Springer, and MDPI are publishing this work because the gap between what AI tools produce visually and what buildings require functionally is wide enough to have become a research field of its own.

When AI Floor Plans Help and When They Hurt

If you are a homeowner exploring whether 1,800 square feet can accommodate three bedrooms, a home office, and an open kitchen without the hallway eating your living room, an AI floor plan generator is a fine place to start. Treat it like a napkin sketch that draws itself. Maket, Planner 5D, and their competitors are excellent at what illustrarch’s 2026 review calls “feasibility and pre-design,” the phase where you need ten layout options to find the two worth developing.

If you are about to hand one of those plans to a builder and say “build this,” stop.

No AI floor plan tool currently on the market accounts for solar orientation, seasonal daylight variation, prevailing wind patterns, or the thermal behavior of the envelope under real weather conditions. Maket’s zoning assistant can parse an uploaded PDF, but it does not cross-reference overlapping regulations or verify that its interpretation matches the actual code, a limitation the illustrarch review documented explicitly. No tool models what happens when afternoon sun hits a west-facing wall of glass in July.

For a homeowner spending $400,000 on new construction, the $30-per-month AI tool generated a plan that might cost $9,000 in excess energy over the life of the loan. A licensed architect, charging $10,000 to $25,000 for full residential design services, would orient that plan correctly on the first iteration and specify window placement, overhang depths, and glazing types calibrated to the site's latitude and microclimate. Do the math. The AI plan is not cheaper.

Why This Narrative Might Be Wrong

AI floor plan tools are iterating rapidly, and the gap Çelik identified may close faster than academic publishing cycles can track. Maket’s agentic editing feature, where you describe changes conversationally and the AI applies them, represents a fundamentally different interaction model than the text-to-image diffusion approaches Çelik tested. Purpose-built architectural AI that works from spatial graphs rather than pixel generation could solve the performance blindness problem within a product cycle or two, and several teams are working on exactly this.

Çelik’s sample was also small: 31 generated plans, eight valid reconstructions. Her methodology was rigorous, but the dataset was limited to three tools, one researcher, and five climate zones. A larger study with Maket, Snaptrude, and TestFit, tools designed specifically for floor plan generation rather than general-purpose image models like ChatGPT and Copilot, might produce different results. Purpose-built tools encode more architectural constraints than generic image generators, and comparing them fairly requires testing them on their own terms.

And $9,000 over 30 years, while real, is not catastrophic, averaging to roughly $300 a year, less than a dollar a day, a number that might not survive a cost-benefit analysis against the $10,000 to $25,000 fee for hiring a licensed architect when the homeowner has already decided to build a conventional tract home in a mild climate. For homeowners in coastal California or the Pacific Northwest, where heating and cooling loads are naturally lower, the orientation penalty shrinks to perhaps $150 annually. Not wrong, but thinner.

What I Did Not Prove

I calculated the $9,000 figure using EIA’s national median energy expenditure and a 15% orientation penalty drawn from DOE Building America research in heating-dominated climates. Actual penalties vary enormously by climate zone, home size, insulation levels, glazing specifications, HVAC efficiency, and local utility rates. A passive-house-certified envelope largely neutralizes orientation effects; a poorly insulated tract home amplifies them. My calculation assumes a conventional 2,000-square-foot home built to code-minimum insulation standards with standard double-pane windows, which describes the majority of new construction but not all of it.

I also did not test any AI floor plan tool myself. My performance claims rest on Çelik’s peer-reviewed study and the Tongji University research. Maket, Planner 5D, and their competitors may have added orientation-aware features since these studies were conducted. If they have, neither their marketing materials nor their documentation mentions it, which is its own kind of data point.

Lastly, architecture involves far more than solar orientation. Structural integrity, egress compliance, accessibility, plumbing stack alignment, and roughly a hundred other constraints shape a buildable plan. Daylight is one axis of performance among many, and I focused on it because it is the most measurable, the most studied, and the one with the clearest dollar figure attached to failure. Other axes matter more in some contexts. But this one you can calculate.