Miraki Mind

AI and Copyright: Switzerland Plans Licensing

Training AI on protected works would need permission in Switzerland, with remuneration and a reversed burden of proof. Here is what is planned and what it means for brands.

Miraki MindAI-assisted editorial · 3 November 2026 · 6 min read

On 26 October 2026, Lena Leuenberger of the Swiss Federal Institute of Intellectual Property (IGE) showed at the CLTR conference in Bern how Switzerland wants to rebuild its copyright law for AI. The lawyer Martin Steiger summed up the talk in German the next day. In short: training and operating AI with protected works would need permission, providers would also have to pay for training abroad, and in a dispute the burden of proof would flip. Parliament gave the mandate with the Gössi motion (24.4596), in a watered-down version.

What is actually new

The question is old, the answer is new. Whether AI training counts as copyright-relevant reproduction at all has not been decided by any Swiss court yet, and experts disagree. In February, the IGE still left two paths open: a statutory exception that allows training without consent but against remuneration, or central licensing through a collecting society. Now, according to Steiger’s summary, there is a concept with three pillars:

Clear rights: Training and operating AI, explicitly including retrieving content to produce an answer (retrieval), fall under the exclusive rights of authors. In principle, that requires permission. Personal use should remain possible and be adapted to AI.
Extended collective licences: Collecting societies should be able to license large volumes of works at once. Anyone who wants to can opt out and negotiate on their own. On top comes a remuneration claim against providers who offer systems in Switzerland that were trained on works of Swiss rights holders or use them, no matter where the training took place. If you already hold a suitable licence, you don’t pay twice.
Transparency and burden of proof: The umbrella associations of providers and rights holders should negotiate what gets disclosed. If they can’t agree, providers must publish a summary of the content they used. And if a rights holder makes it plausible that their works were used, that counts as proven until the provider proves otherwise.

Training on protected works should need permission. Providers should disclose what their models were trained on. If a rights holder makes the use plausible, the provider must prove otherwise.

What doesn’t exist yet: a legal text. According to Leuenberger, the work is at an early stage, and the consultation could start in spring 2027. The IGE’s FAQ still says “by the end of 2026”. The ancillary copyright for media depends on this too: in June, Parliament sent it back to the Federal Council so that AI gets included.

What this means for a production

On set, nothing changes for now, no drama. The new duties hit the providers of models and systems, not the studios that work with them. Indirectly, though, they do. Leuenberger herself points out that Switzerland is a small market, and too many special rules could lead providers to restrict features for Switzerland or to lock users here out altogether. For a production that relies on cloud models for image and motion, that would be a real planning risk. And whatever providers pay ends up in the price somewhere.

The other side: a duty to disclose what a model was trained on makes a question answerable that, honestly, has no clean answer today.

What this means for a brand

First off: this is not legal advice. The revision targets providers. Until it arrives, today’s law applies to brands, and the IGE’s FAQ is fairly clear about what a campaign should watch out for:

In der Schweiz ist urheberrechtlich nur geschützt, was von einem Menschen stammt.
IGE, FAQ KI und Urheberrecht

Rights in the output: An image from a banal prompt is not protected. Protection is possible if a human uses AI as a tool and makes a creative contribution.
Liability on use: If an output contains protected content, whoever uses it infringes copyright, even without knowing. The IGE recommends checking outputs beforehand, for example with a reverse image search and a look for watermarks, signatures and recognisable faces.
Style: A style is not protected. If an output comes too close to an original, it is an unauthorised adaptation. If it can be confused with an artist’s work, it can infringe their personality rights.
Terms of use: They can prohibit commercial use. And according to Leuenberger, powerful providers often shift responsibility onto users via their terms.

The reversed burden of proof makes it easier for rights holders to go after providers. That raises the pressure to license cleanly, and for brands that is an opportunity too: a model with documented provenance is easier to stand behind. Labelling is a separate duty and remains unaffected (more on that here).

Our take and the limits

The direction is clear and easy to read for a production: using protected works with AI will require permission, and providers must disclose, pay and, in case of doubt, prove. In the short term, this changes little about a brand’s own diligence; in the medium term it could turn models with clean provenance into a selling point. Whether providers play along or skip Switzerland is the big open question.

Plenty is also open in the details: what the tariffs look like, what “making it plausible” means in practice, how this fits with the EU. And one limit of this piece: we know the three pillars from Steiger’s summary of an oral talk, not from an IGE document. It only becomes binding with the draft in consultation.

What is actually new

What this means for a production

What this means for a brand

Our take and the limits

Sources