AI Coffee Go

Every new AI tool raises two questions: where does the data end up, and who is liable when something goes wrong. This episode explains what "whitelisted" means, namely tools checked internally and cleared by IT or the data protection officer. A tool that is not on the list is not automatically forbidden, but the risk then sits with you.

In the practical check: Adobe Firefly, GDPR-compliant and without training on user data, but only on enterprise accounts. Midjourney, creatively strong, but without EU servers, without a data processing agreement and with prompts public by default, which makes it risky for client work. DALL·E through the API with training excluded, but on US servers. Kling AI as a Chinese provider without GDPR compliance. ComfyUI hosted locally, with no data in the cloud, but the full liability with you.

The golden rule: before using any tool, ask where the data lands and who is liable.

With Karolin Sannwald as the legal voice, this episode asks whom the labelling duty under Art. 50 of the EU AI Act applies to, what has to be labelled and how. Under the regulation a deepfake is any AI-generated or manipulated content resembling real people, places or events that could falsely appear authentic, whether or not a specific real person is recognisable. Text only has to be labelled when it is of public relevance and no human final check took place, meaning real proofreading rather than skimming. The label itself has to be clearly recognisable to an average person, effective and working across platforms; the exact wording is free.

Swapping faces, cloning voices, inventing statements: where is the line? The episode separates three scenarios: the clear-cut face swap or voice clone putting words in someone's mouth, the grey zone of an AI character with a recognisable likeness to a real person, and the clean case of an AI character without any likeness.

The Bruce Willis and RBB case shows that a voice is part of personality rights, and that an AI clone which merely "sounds like" someone can amount to identity theft.

Three routes are sketched for practice: prompt fictional personas, use an AI casting agency, or work with real people who have expressly agreed to AI use.

Protection is possible on three levels: as a person, as a brand, as a production company.

A client contract asks for "all copyright" in the delivered material, but the material is AI-generated. Can you transfer what you do not legally own? This episode works through practical questions: may client material such as logos or protected designs be uploaded as prompt input at all? What applies when your own work goes in, such as retouching or compositing, where the finished piece can hold its own copyright even though the purely AI-made parts remain unprotected?

And what do you do when the client insists on a transfer of copyright anyway?

The answer: make the situation transparent, offer usage rights, and refer to a lawyer where needed.

The golden rule: you cannot transfer what you do not own. Put into the contract what you can actually deliver.

An advertising campaign, produced entirely with AI, and three days after release somebody gets in touch: "that is me in there." Who is liable? For this episode Marc Soiné, an insurance specialist for the creative and media industries at Schäfer & Soine, was the guest.

At the centre: whether your own policy covers AI damage at all, what a cyber policy does when AI tools have been fed client material, and whether solo freelancers without their own servers need cyber cover.

The episode makes clear that insurance cover for AI productions cannot be assumed and has to be checked actively.

The follow-up to episode 3, this time on putting it into practice before the 2 August 2026 deadline.

Not sufficient: telling people verbally up front, a brief on-screen note, or a mention only in the video description.

Two examples make the difference tangible. The fully AI-generated commercial without real actors has to be labelled, and so does the hybrid production with real footage as soon as an AI background, an AI voice or AI text shapes the overall impression.

Platform labels do not replace your own duty.

The golden rule: could someone believe the AI output is real? Then label it, and check regularly that the label stays clearly recognisable at every point.

Art. 50 of the EU AI Act applies uniformly from 2 August 2026; realistic AI content such as deepfakes and unedited AI text has to be labelled.

The standard is the free official EU icons, published by the European Commission in June 2026 as part of the voluntary Code of Practice. Signing that code brings a kind of safe harbour; anyone using their own solution may have to prove it is equivalent.

What the law does not yet settle is who inside a supply chain is liable when the label is missing. Hence the golden rule: agree in writing before the project starts who does the labelling, and put the use of AI into the project confirmation.

A solo episode without Karolin, with a stronger disclaimer in the housekeeping.

The starting point is again the standard contract demanding "all copyright", although AI output in Germany is generally not automatically protectable for lack of human creative input.

What does exist are usage rights from each tool's terms, not copyright, neither with you nor with the provider.

The episode explains why this is no reason to panic: most clients want usage rights rather than legal constructs, and in practice that is usually enough.

Cross-reference to episode 6 for anyone who wants to go deeper into liability.

Another solo episode without Karo and without a legal part, opening with the community: what is annoying about AI text? The expected answers are sameness, interchangeability, three-word sentences, em dashes, boredom, in short: AI slop.

The episode asks what detection tools actually achieve and what they do not, and when AI genuinely helps with writing: think for yourself first, then open the tool; the idea comes from you; only on subjects you understand; with clear context about audience and purpose; write it yourself first, then have it taken apart.

The golden rule: is there a sentence in it that only you could have written?

Three weeks after a client project, GEMA gets in touch; the music came from Suno. Karolin puts the Munich Regional Court ruling in context: Suno used six well-known songs, among them "Atemlos" and "Rasputin", for training without a licence, in part by circumventing protective measures, and the generated music came confusingly close to the originals.

Under the ruling the provider is responsible, not the user, though the judgment is not yet final.

So far only Warner Music and Bertelsmann have licensing deals with Suno; Universal and Sony are still litigating in the US.

For client work the episode recommends a similarity check, prompt documentation, a check of the usage licence and contractual cover for liability risks, because technical output never protects against copyright infringement by itself.

The costing is done, the production is running, and then 4K renders, risen credit prices and the fifth round of feedback hit in the middle of the project. "That is covered by the day rate?!" is the fallacy of this episode: higher render quality and pricier credits eat the budget in minutes, and the production pays in the end.

At the centre are the token and credit models of the common AI tools and how mid-project changes can blow up a costing.

The episode shows how to protect yourself, among other things by knowing the pricing models realistically and by building a buffer for feedback rounds into the quote.