AI Images in Advertising: The Legal Position in Switzerland
AI images in advertising are visuals generated or materially altered by a generative model instead of captured with a camera. Whether a Swiss brand has to label them is a far narrower question than the current debate suggests, and the answer depends on where the campaign runs. Switzerland has no labelling duty for AI-generated advertising today, and none has been decided. The EU AI Act does introduce transparency obligations that apply from 2 August 2026, but not the blanket duty that is widely reported: the marking obligation falls on the companies that build the tools, while advertisers owe disclosure only in specific cases. This guide sets out what actually applies, what is still unresolved, and what both mean for a production setup. None of it is legal advice, and the open questions are marked as open.
Do you have to label AI-generated advertising in Switzerland?
The short answer is no. Swiss law contains no labelling duty for AI-generated advertising, and none has been decided. Parliament examined the question and declined: Motion 23.3563 was closed on 6 May 2025 with the assessment that specific regulation of deepfakes does not currently appear sensible for Switzerland (Curia Vista, Motion 23.3563).
The Federal Council set out a sectoral approach on 12 February 2025 and announced a consultation draft covering transparency, data protection, non-discrimination and supervision by the end of 2026 (BAKOM on artificial intelligence). In its answer of 20 May 2026 it stated that it was still too early to name concrete legislative changes, and the consultation had not been opened. Switzerland signed the Council of Europe AI convention in March 2025 but has not ratified it.
Two things apply regardless of any AI-specific rule. Unfair competition law catches misleading advertising whether or not a model was involved: Art. 3 para. 1 lit. b UWG makes it unlawful to make incorrect or misleading statements (UWG, Fedlex). And the Federal Data Protection and Information Commissioner has held that the use of programs enabling the falsification of faces, images or voice messages of identifiable persons must always be clearly recognisable (FDPIC, 2023, updated 2025). That is the only recognisability requirement any Swiss authority has issued in this area. It sits in data protection law and is limited to identifiable persons. It is not a general advertising label.
Self-regulation has not filled the gap either. The Swiss Commission for Fairness reported for 2025 that it had dealt with transparency and labelling of AI-generated content and that the topic would occupy it again in 2026. It has issued no guideline and published no decision.
What does the EU AI Act actually require from 2 August 2026?
This is where most current reporting goes wrong. The AI Act splits the transparency duty between two roles, and advertisers sit on the lighter side of it.
Providers, meaning the companies that build the tools: Art. 50(2) requires that outputs be marked in a machine-readable format and detectable as artificially generated or manipulated (AI Act, Article 50).
Deployers, meaning the brand or agency running the campaign: Art. 50(4) requires disclosure only where the content constitutes a deep fake.
The Commission FAQ names an advertising company explicitly as a deployer, and clarifies that employees and freelancers such as digital animators, web designers or content creators are not separate deployers (European Commission FAQ on Article 50). Where disclosure is owed, it has to be understandable and perceivable by natural persons without any specific technical tools, at the latest on first contact. A line in the terms and conditions or a note in the website footer does not discharge it.
Whether an ordinary AI product image counts as a deep fake is the open question. Art. 3(60) sets three cumulative criteria (AI Act, Article 3):
Resemblance to existing persons, objects, places, entities or events.
Existence: the subject exists, can plausibly exist, or could have plausibly existed.
A false appearance of authenticity, judged by context and expected audience.
That definition is considerably wider than the everyday use of the word and is not limited to real people. It is also not settled where it stops. The Commission's final guidelines of 20 July 2026 run to 51 pages, are non-binding, and contain no example covering a retouched or CGI-like standard product shot (Commission guidelines on transparency). No court has ruled on it, and binding interpretation would come only from the Court of Justice.
Two timing details are widely misreported. First, there is a grace period: the Commission FAQ states that systems placed on the market before 2 August 2026 must meet the marking and detection obligation only from 2 December 2026, and that content generated before 2 August 2026 does not need to be labelled retroactively. Second, the frequent claim that high-risk obligations begin on 2 August 2026 is out of date. Regulation (EU) 2026/1744 of 8 July 2026 moved them to 2 December 2027 and 2 August 2028 (Regulation (EU) 2026/1744, EUR-Lex). Article 50 was not moved.
Penalties for breaching Art. 50 reach 15 million euros or 3 per cent of worldwide annual turnover, enforced by national market surveillance authorities. The accompanying Code of Practice on Transparency and the EU labelling icons are both voluntary. The obligation behind them is not.
Deployers cannot simply rely on the machine-readable marking embedded in the content by the provider under Article 50(2) to fulfil their disclosure obligation.
European Commission, 2026, FAQ on transparency obligations under Article 50 AI Act
When is a Swiss brand covered at all?
Being established in Switzerland neither puts you inside the regulation nor keeps you out of it. Art. 2(1)(c) extends the Act to providers and deployers in a third country where the output produced by the AI system is used in the Union (AI Act, Article 2).
Covered: you target EU audiences, run an EU landing page, or ship the campaign into an EU market.
Not covered on this basis: the campaign runs in Switzerland only, with no EU delivery.
Unchanged either way: Swiss unfair competition and data protection law, and the platform rules further down.
In practice most Swiss brands with any EU footprint should assume they are a deployer for at least part of their media. Deciding disclosure once at the process level is cheaper and more defensible than deciding it asset by asset, and it removes the question from the hands of whoever happens to be uploading.
Who owns an AI-generated image, and is it protected?
Two different questions get mixed here constantly. Ownership is contractual and comes from the tool's terms. Protection is statutory and is genuinely unresolved.
On protection, Swiss copyright requires a human creation. Art. 2 para. 1 URG defines works as intellectual creations of literature and art with an individual character (URG, Fedlex). The Swiss Federal Institute of Intellectual Property takes the position that output could be protected provided it can be classified as the expression of a human intellectual creation, and says these questions are currently disputed from a legal perspective (IGE on copyright and AI). No Swiss court has decided the point. Motion 24.4596 on clarifying the copyright act was referred in December 2025 without a draft. In Germany, the Higher Regional Court of Düsseldorf held on 2 April 2026 that purely software-controlled output attracts neither copyright nor photographic protection.
On ownership, the terms differ more than most teams assume:
Midjourney: you own the assets you create, but companies and their staff above one million US dollars in annual revenue need a Pro or Mega plan for that to hold. Assets come as is, without warranty of title or non-infringement, and there is no indemnity (Midjourney Terms of Service).
OpenAI: assigns all right, title and interest in the output to you, while noting the output may not be unique (OpenAI Service Terms).
Adobe: output is customer content, but Adobe disclaims any warranty that it will not violate third-party rights and notes the output may not be protectable (Adobe generative AI terms).
Google Cloud: generated output is customer data and Google asserts no ownership rights (Google Cloud generative AI terms).
The indemnities matter more than the ownership clauses, and they are narrower than they look. Adobe's cover applies only on certain paid tiers, is capped at 10'000 US dollars per output or claim, and falls away on any modification, any combination, and depending on the context in which the output is used. OpenAI and Google both exclude trademark claims arising from use of the output in trade or commerce. That is precisely the advertising case.
On liability, the most relevant European authority so far is the Regional Court of Munich I, which held on 11 November 2025 in the GEMA proceedings against OpenAI that memorised content in a model and its reproduction constitute copying, and that the operator rather than the user is responsible. In the United States, the Northern District of California allowed Getty's Lanham Act claims against Stability AI on 23 April 2026, treating AI images in commercial distribution as goods. There is no Swiss and no EU decision on AI images in advertising, which is worth stating plainly rather than filling with analogy.
What do the ad platforms require?
Platform rules are stricter than Swiss law and, for ordinary advertising, stricter than the AI Act. They are also where a breach costs you something immediately, in the form of a rejected ad.
TikTok is strictest: an AIGC label or a clear disclaimer is mandatory, and undisclosed AI-generated content leads to the ad being rejected or restricted. Adjustments to light, colour or background are exempt (TikTok advertising policy).
Google Ads offers an optional asset label for content created or edited with AI since July 2026, and notes that regulations in the EU, India and New York require disclosures for certain AI assets, while using the label does not guarantee compliance (Google Ads policy).
Meta imposes no disclosure duty on the advertiser in ordinary advertising. It labels content itself and added automatic detection of third-party AI from 1 June 2026. Mandatory self-declaration applies only to political and social issue advertising, which Meta stopped serving in the EU in 2025 (Meta on AI information).
LinkedIn has no AI disclosure duty, only a C2PA icon on signed assets (LinkedIn ads policy).
The practical consequence is convenient: build to TikTok's standard and you already satisfy everything the others ask, plus most of what Art. 50(4) would demand if a given asset turned out to be in scope. That makes TikTok the cheapest place to set an internal rule, and it removes the need to litigate the deep fake definition in-house for every campaign. How this plays out in production is something we cover in our guide to AI content creation in paid social.
Want AI content with the compliance question already answered?
At Collective Agency we build art-directed AI content systems where disclosure and rights are part of the setup rather than an afterthought. If you want a straight read on what applies to your markets and how to handle it in production, get in touch. For the wider Swiss picture on rights, see our overview of intellectual property law in Switzerland. This article is not legal advice: for a binding assessment, talk to your legal counsel.
The output could be protected, provided it can be classified as the expression of a human intellectual creation. These questions are currently disputed from a legal perspective.
Swiss Federal Institute of Intellectual Property (IGE), 2025, Copyright questions in the training and use of artificial intelligence
Summary
Switzerland has no labelling duty for AI-generated advertising, and none has been decided: Parliament declined specific deepfake regulation in May 2025, and the Federal Council consultation announced for end of 2026 has not opened. The EU AI Act does apply from 2 August 2026, but the marking obligation falls on tool providers under Art. 50(2), while advertisers as deployers owe disclosure only for deep fakes under Art. 50(4). Whether an ordinary AI product image qualifies is unresolved, with no guideline example and no court decision. This guide covers what applies, when a Swiss brand is in scope at all, who owns the output under the major tool terms, why the indemnities are narrower than they appear, and what the ad platforms require, where TikTok is stricter than the law.
Related Articles
How ChatGPT Ads Work: Formats, Costs and Where They Run
Published
By
Dino Reichmuth
AI Content Creation Dominates the Feed
Published
By
Dino Reichmuth
External Marketing Consulting: When It's Worth It and What to Look For
Published
By
Lars Kienle
Display and Video Advertising in Switzerland: Formats, Costs, and Programmatic
Published
By
Lars Kienle