Your Architect Billed You $96,000 for Custom Plans. Federal Law Says Nobody Owns Them.
Picture the closing meeting. Your architect slides the final plan set across the table, eleven sheets, every elevation rendered in soft evening light, the kind of distinctive sawtooth roofline that made you hire this firm instead of buying stock plans. You paid $96,000 for the design. Eighteen months later, a builder three streets over is framing a house with your roofline, your window rhythm, your entry sequence, and your first call is to the lawyer who handled your closing. Your lawyer looks at the renderings, asks one question, and ruins your week.
"Which parts did a person actually draw?"
That question did not matter five years ago, when every line in a residential plan set traced back to a human hand and authorship was a fact too obvious to litigate. It matters now because the most important copyright ruling for residential architecture in a generation says the answer determines whether you own anything at all.
Copyright law, settled before the wave arrived
On January 29, 2025, the U.S. Copyright Office released Part 2 of its AI report, and its conclusion was blunt: generative AI outputs are protected only where a human author determined sufficient expressive elements. Detailed, laborious, ingenious prompting does not count. It considered the argument that a hundred carefully engineered prompts constitute authorship and rejected it, finding that prompts "do not appear to adequately determine the expressive elements produced, or control how the system translates them into an output." What a human selects, arranges, or substantially modifies can be protected. What the machine decided cannot.
Seven weeks later, the D.C. Circuit made it case law. In Thaler v. Perlmutter, decided March 18, 2025, a unanimous panel held that the Copyright Act requires every work to "be authored in the first instance by a human being," as a matter of statutory law, not agency discretion. Judge Millett's opinion walked through the statute and found it riddled with references intelligible only for humans: authors with lifespans, with widows and surviving children who inherit, with signatures and nationalities. Machines appear in the Act only as tools. A computer scientist had listed his "Creativity Machine" as sole author of an AI-generated image, and the court affirmed the Copyright Office's refusal to register it, closing the narrowest possible fact pattern: a work with no human expressive contribution whatsoever. No human author, no copyright, no appeal to equity.
For architects, the most instructive precedent is older and stranger. In 2023, the Copyright Office's Review Board ruled on Kristina Kashtanova's graphic novel Zarya of the Dawn: her human-written text was copyrightable, her selection and arrangement of AI-generated images was copyrightable, but the Midjourney images themselves were not. Swap "graphic novel" for "plan set" and you have the entire legal framework for AI-assisted residential design compressed into a single paragraph, which is more guidance than the AIA's standard contracts currently offer on the subject.
A plan-set audit nobody has run
So let's apply that framework to the thing you actually bought. Take a hypothetical $800,000 custom home, where custom residential design fees conventionally run 8 to 15 percent of construction cost, so that at the 12 percent midpoint the design fee lands at $96,000. Here is what that money purchased, sorted by what federal law will let you defend.
Likely protected, because a human determined the expression: the hand-sketched parti and massing studies; the manual CAD and BIM drafting, dimensioning, and detailing; the structural coordination and site-specific problem-solving, from grading against the actual lot to threading the stair past the actual shear walls, all of which record judgment no model exercised. And, crucially, the architect's curation itself, the rejected schemes, the chosen option, the arrangement of AI-generated studies into a coherent design, which is exactly the "selection, coordination, and arrangement" the Office protected in Zarya.
Likely not protected, because the machine determined the expression: the raw Midjourney elevation renderings passed into the set unmodified; the TestFit-generated floor plan options the architect picked from a menu of hundreds without redrawing; any AI-drafted specification boilerplate reproduced verbatim, none of which any person shaped beyond the picking. Under the January 2025 report, these portions have no author, cannot be registered, and in an infringement suit contribute nothing to the claim.
And here is the perversity: the most visually distinctive elements of the set, the glowing elevations the client fell in love with, the images a copycat builder would actually steal, are the least protectable parts. Meanwhile the dimensioned foundation plan nobody will ever photograph is the most protectable. You paid custom fees for a design whose signature moves may belong, legally speaking, to no one.
What protection is worth when it exists
To calibrate the stakes, look at what enforceable plan copyright actually buys. In AIS Designs Co. v. J.J.S. Custom Built Homes, a Texas federal jury found a builder had infringed a design firm's home plans and the court entered judgment: $278,605.95 against the builder, $570,407.19 against one individual defendant, $6,984 against another, bringing the total recovery to $855,997.14 plus a permanent injunction. That is the enforcement value of a registered, human-authored plan set.
Behind verdicts like that one sits a statutory machine, 17 U.S.C. section 504(c), which allows up to $150,000 per work for willful infringement, a ceiling that turns cease-and-desist letters from requests into arithmetic. But the machinery has an on-switch: under section 412, you generally cannot collect statutory damages unless you registered before the infringement began or within three months of publication, and registration requires a human author. For the AI-generated portions of your plan set, there is no on-switch at any price, and the recovery for those portions is not reduced but zero, because unprotectable expression contributes nothing to an infringement claim.
Residential plan piracy was already a thriving industry when every plan was drawn by hand; design firms employ staff whose entire job is policing copycat builders, a job that gets materially harder when the copier's best defense is a paraphrase of the Copyright Office. Now hand the copycats a citable federal defense. "Your honor, the expressive elements were determined by a machine" is no longer a creative argument. It is a paraphrase of the Copyright Office.
Your contract assumes a human author
The standard AIA owner-architect agreement, B101-2017, allocates copyright in the instruments of service on the assumption that the architect is the author. That assumption was safe when the only authors in the building were people. Frantz Ward's construction group has catalogued five specific ways AIA contracts fail to address AI on real projects: standard of care, responsibility for design documents, intellectual property ownership, third-party technology risk, and insurance alignment. On ownership, their warning is direct: questions of who owns AI-generated content remain open, especially when third-party AI platforms are involved, and the party that selected the tool will likely own the liability if the tool fails.
Even the AIA's own contracts publication saw this coming years before the case law caught up. In May 2024, "Who Owns Building Design in the Age of AI?" acknowledged the ownership question was unresolved and added a darker possibility: if AI-generated designs are arguably derivative of the training data, meaning other architects' designs, then existing law might leave ownership with the creators of the training images. Read that twice. Your "custom" AI-generated facade might not merely be unprotectable. In the worst case, it might belong to someone else's training set.
None of this has produced a standard contract fix. As of this writing, the AIA has not issued AI-specific amendments. Those 2017 revisions predate the generative wave entirely, which is precisely the situation the Frantz Ward lawyers suggest curing with supplementary conditions: disclose AI use, define human oversight, allocate responsibility for AI errors, and state who owns and may use AI-generated content.
Steel-manning the opposition
Three objections deserve honest weight, because the honest version of this story is narrower than the alarming one.
First, almost no residential plan set is purely AI-generated. Architects modify, detail, coordinate, stamp, and take professional responsibility for the documents, and that human overlay may well be enough to protect the set as a whole under the Office's "AI as a tool" doctrine. Admittedly, the audit above sorts individual elements while a court might look at the set and see a human-authored compilation, and for most practicing architects that holistic reading is probably correct.
Second, copyright registration was never the real enforcement mechanism in residential plan piracy. Contracts do most of the work: single-use licenses, liquidated damages clauses, practical obscurity. A builder who signed your license agreement cannot copy your plans just because the renderings lack a human author; breach of contract does not require a copyright. Still, the copyrightability gap matters at the margin, and the margin is where the valuable cases live, which is exactly where a citable federal defense does its damage.
Third, the strongest pro-protection argument has never been tested. In Thaler, the court expressly declined to reach the claim that Thaler himself was the author by virtue of creating and directing the AI, because he waived it below. A future litigant, an architect who can document directing every expressive decision through an AI tool the way a photographer directs a camera, could win where Thaler lost, since the D.C. Circuit left that door open and nobody has walked through it yet.
Concede all three, and the core still holds: at the margin where designs are valuable enough to steal, the infringer's best defense is now federal guidance, the contract's authorship assumptions are stale, and the window to fix the paperwork is open right now, while only a minority of the profession has adopted the tools.
What this article did not prove
Several honest boundaries. This is legal analysis, not legal advice, and the Copyright Office itself insists the authorship inquiry is fact-specific, so no audit of a hypothetical plan set predicts any real registration outcome. That 8 to 15 percent fee range is industry convention, not statute; actual fees vary by market, scope, and reputation, and the $96,000 figure is illustrative arithmetic, $800,000 times 12 percent, not a surveyed average. That waived authorship theory in Thaler remains untested, and a future decision could narrow everything argued here. No data exists on how many residential plan sets involve generative AI or how many have faced authorship challenges at registration; the Office does not break out denials that way, and research found no case of a residential architect actually being denied a plan registration on AI-authorship grounds. Finally, the AIA's reported 6 percent regular AI-use figure among architects comes via secondary reporting on the AIA's March 2025 survey, and adoption is moving fast enough that any number here has a short half-life.
What to do before you sign
If you are commissioning custom plans, add a supplementary condition to the owner-architect agreement requiring the architect to disclose which deliverables were AI-generated and to identify which portions they consider human-authored. You cannot protect what you cannot inventory.
Ask for the optioneering record, keeping the rejected AI schemes, the marked-up iterations, and the selection memos as evidence of human selection and arrangement, which is the protectable part under Zarya logic. A firm with no such record has a thinner claim than one that documents its judgment.
Register the human-authored portions promptly, since registration within three months of first publication preserves eligibility for statutory damages under section 412, without which even a slam-dunk copying case pays only the actual damages you can prove dollar by dollar; registration is cheap, while unregistered rights are a theoretical comfort nobody can price.
If you are buying AI-generated stock plans, price in the thin protection. Compete on execution, service, and speed rather than on the un-copyability of the design, because the design, or at least its most distinctive expression, may be legally un-copy-protectable.
And a final caution that runs the other direction: do not upload proprietary or client-confidential plans into public AI tools without reading the platform's terms on training data. This ownership question has two jaws, one being that your AI-generated plans may be unprotectable, and the other being that your human-authored plans, fed into someone else's model, may train your competitor's next "custom" design.
Codes do not care about your feelings, and neither does the Copyright Act. It cares whether a human determined the expression. For a growing share of residential design, the honest answer is becoming uncomfortable, and the time to put the answer in writing is before the plans are drawn, not after the copycat pours the foundation.