The innovation that makes the difference is not the latest tool

A professional is not measured by what they add, but by how they care for what is entrusted to them. We all race to adopt the latest tool — it is this year's conversation, and last year's too — but that race is not what defines us. And in architectural visualisation that distinction has never mattered more.

It is worth remembering where we come from. This craft was born to take the physical model further, to be able to think forms that were almost unthinkable by hand. The Guggenheim Bilbao is the example that comes to everyone's mind: Gehry raised those titanium curves by scanning a physical model into the computer, with software that came from the aerospace industry. One critic went as far as saying the building could not have been built without that technology. It was not rendering yet — it was design and geometry — but that is where the border between what can be imagined and what can be built began to bend.

Since then everything has become more digital and, at the same time, less tangible. It is a young profession, barely three decades old, that we have largely taught ourselves: the software ran ahead of the school. There is nothing to hide in that. We are, in a way, a first generation learning as we go, and on top of it a tsunami has come our way that we all know by its initials, AI (cue, mentally, the violins from Psycho). It is normal to carry questions. What is not normal is to act as if they did not exist. So let us talk about something almost nobody talks about: how the confidentiality of projects is being handled — or rather, how it is not.

What reaches our hands is not ours. A render is someone else's project that we bring closer to reality. Inside goes the client's work, yes, but also the work of many people who never even hired us: a brand's furniture, a specific material, an architecture studio's plans (someone's private house), a building that is not yet public. When we handle that image, we handle the trust of that whole chain. And protecting it is the first part of the job, not an add-on.

Regulation is starting to arrive, and it is needed. In Europe there are already two concrete pieces: the 2019 Copyright Directive, which lets you reserve your work against data mining, and the 2024 Artificial Intelligence Regulation, which obliges the large models to respect that reservation and to publish what they were trained on. But a professional does not wait to be forced — and besides, the law hands you the tool; it does not use it for you.

Robert Kneschke learned this the hard way: a German photographer who one day found his images inside LAION, the gigantic dataset that has been used to train half of generative AI. They had been collected by crawling the web. Kneschke sued, and lost — first in 2024 and again on appeal in 2025. The Hamburg court upheld the collection on two grounds: it fell under a research exception and, above all, the prohibition he had put up (a notice, in plain language, in the terms of the website where he sold his photos) did not count, because it was not in a format a machine could read. Translated: a no-trespassing sign written for people is not read by a robot. You have to say it in its language.

There is, however, a nuance in our favour that few people know. The Spanish Intellectual Property Law, in its article 43.5, says that when you assign your images for a project, that assignment does not reach uses that did not even exist when you signed. Training an AI with a render commissioned years ago was not on the table: that permission was never given — it is still yours. It protects us looking backwards. From here on, though, the use is known, so that gap is yours to fill, by writing it into the contract.

And let nobody think this is distant theory. Two problems are worth separating. One is what goes in: Anthropic agreed to pay one and a half billion dollars to authors and publishers for having trained on books it downloaded pirated. OpenAI, in its lawsuit with the New York Times, was ordered by a judge earlier this year to hand over twenty million conversations so that what was inside could be traced. The other problem is what comes out: Disney, Universal and Warner have sued Midjourney because its AI generated characters of theirs as recognisable as Elsa or the Minions. That case is still open, but it is the closest to our trade, because it is about images. The pattern repeats: once the work enters the machine, the damage cannot be undone. It is paid for, or exposed, but not erased.

And you have to understand the tools you use, because even the protection has cracks. The law lets you put a seal inside the image, in its metadata, to say: do not train on this. But it is fragile: it takes no more than a platform compressing the photo to make it lighter for that seal to vanish. Not out of malice — out of optimisation. The organisation that sets the standard verified it: out of fifteen platforms, only one preserved that information. And deleting that data is forbidden, yes, but only when it is done on purpose, and an automatic compression has no intent. Care, then, cannot rest on a single seal: it rests on judgement, on the contract, on making the effort to stay informed and on learning what it is you are using.

I do not have a closed answer here, and I suspect nobody has one yet. We are the first generation of a new craft whose ground has shifted under its feet. The sensible thing is not to run faster nor to stop, but to learn without losing focus: to respect the data, the work and the trust of those who leave them with us. Because in the end, the innovation that makes the difference is not the latest tool. It is being the studio that can be trusted. And that, curiously, is a job no machine does for you.

Bibliography and sources

Spain, Intellectual Property Law (TRLPI, RDLeg 1/1996), art. 43.5: the assignment does not extend to modes of use non-existent or unknown at the time of signing. BOE

Directive (EU) 2019/790 on copyright, art. 4: text and data mining exception and reservation of rights by machine-readable means. EUR-Lex

Regulation (EU) 2024/1689 (AI Act), art. 53: obligations of general-purpose AI models (respect for the reservation and training-data summary). artificialintelligenceact.eu

Robert Kneschke v. LAION: LG Hamburg, 27 September 2024 (310 O 227/23) and OLG Hamburg, appeal of December 2025 (5 U 104/24): the reservation of rights must be machine-readable. Bird & Bird

Bartz v. Anthropic: $1.5 billion settlement with authors and publishers over the use of pirated books (2025). NPR

The New York Times v. OpenAI: court order to hand over 20 million ChatGPT conversations (January 2026). National Law Review

Disney, Universal and Warner v. Midjourney: lawsuit over generation of protected characters (June 2025, ongoing). Time

Directive (EU) 2001/29 (InfoSoc), art. 7: protection of rights-management information. EUR-Lex

IPTC, Social Media Photo Metadata Test (2016): of 15 platforms, only Behance preserved the metadata. IPTC

Guggenheim Bilbao and CATIA (Guggenheim Foundation). guggenheim.org

This article is informative in nature; it does not constitute legal advice.