How does copyright work in architecture now that AI exists? You can no longer plagiarize as you once did

In 2025, architecture no longer struggles only with forms: it struggles with data. Generative models promise concepts in hours instead of weeks, turning “style” into an algorithmic lever. This power is the new frontier of creativity, but it hides an ethical and legal short circuit. If AI can imitate the fluid curves of Zaha Hadid or the fragmented volumes of Frank Gehry, where does inspiration end and where does algorithmic plagiarism begin? The issue is not aesthetic: it’s legal, economic, and it is already redefining the very value of a brand in architecture.

The problem arises from the way AI learns. When you enter a prompt such as “residential project in the style of BIG,” the algorithm doesn’t invent but analyzes trillions of existing images, drawings, renderings, and photos of real architecture, to distill the aesthetic grammar of that particular studio. The result is a “stylistic copy” produced in seconds. This erosion of a firm’s visual identity is, for studios like Zaha Hadid Architects, a direct threat to their premium value and their intellectual property. If the look and feel becomes an algorithmic commodity, the concept of architecture as a luxury begins to crumble.

What the law says in Europe and the United States

For once, the regulatory framework is almost keeping pace with technology. In Europe, the AI Act imposes unprecedented transparency obligations on providers of generative models: they must declare that a piece of content was created with AI and publish a summary of the training data, in addition to complying with copyright law. The key provisions for “general-purpose” models will fully enter into force on August 2, 2025, with penalties of up to 7% of global turnover for non-compliance. It’s a paradigm shift that enforces traceability of the creative process, transforming copyright compliance from an ethical option into a procurement requirement.

Across the Atlantic, the approach is more straightforward: the U.S. Copyright Office has reiterated that works autonomously generated by AI are not protected by copyright; a human contribution that is creative and verifiable is required. This means one crucial thing for designers: if you delegate the entire formal outcome to AI, you lose exclusive rights to the output. Protection exists only if the professional directs and substantially transforms the algorithmic output. Meanwhile, the most controversial issue remains the training of models. The lawsuit of Getty Images against Stability AI (currently in court in the UK, with global implications) is the industry’s thermometer: Getty accuses the massive, unauthorized use of its archive to train Stable Diffusion. For architecture, which increasingly works with reference images and renderings, this precedent will determine how far one can go in using protected datasets without licenses.

Regarding authorship, current law protects specific architectural works but not style as such. In the U.S., after the Architectural Works Copyright Protection Act, floor plans and built works are protected, but adherence to a pre-existing style reduces the scope of originality. You cannot copy a concrete project, but you can work in the style of a master. This is where AI complicates everything: if it produces an output stylistically indistinguishable from a recognizable brand, commercial and reputational disputes are inevitable. Meanwhile, the most advanced firms are not waiting for court rulings. Zaha Hadid Architects has long declared an extensive use of AI and computation, also in collaboration with Nvidia, as an accelerator of the design process, governing tools with clear policies on datasets and prompts. The lesson is pragmatic: AI is a lightning-fast assistant, but the real challenge lies in defining and defending one’s creative identity.

So how can the industry protect itself?

Three measures are becoming standard. First: Transparency and Provenance. The adoption of the C2PA/Content Credentials protocol (promoted by Adobe, Google, Microsoft) makes it possible to tag files with signed metadata that document origins, edits, and AI use, creating a verifiable forensic trail. Second: European Compliance by Design, requiring model providers to supply documentation proving licenses and rights management. Finally: Redefining Originality in the Workflow. The principle is clear: the more AI decides, the fewer rights you have. Design must include human traces: ethical curation of datasets (not just copying, but informed selection), documented creative instructions, and manual modeling steps that truly influence the formal result. Ultimately, AI reduces style to texture, and texture is replaceable. The reputational insurance comes from the human process before the image itself.

For this reason, the project becomes stronger than the image. AI generates convincing copies, but the answer is neither to reject it nor to idealize it, rather, to train the profession in a new ethics of provenance and a discipline of authorship. With clearer rules, court cases that will set precedents, and traceability standards, architecture can use AI to accelerate ideas without outsourcing its identity. The question is no longer whether AI will copy a style, but whether the architect will be able to demonstrate and defend what remains inimitable, the human and conceptual part of the project.