Notes from the Innovation Lab is a monthly roundup from Lukasz Domanski, Director of the Freedman Innovation Lab, on the technology and process shifts that matter in global marketing. This edition features an in-depth look at the upcoming EU AI Act and an update on how far AI dubbing has come.
From 2 August, the EU AI Act's transparency obligations take full effect and it lands specifically on advertising. Penalties can run to 15 million euros or 3% of global turnover, whichever is higher.
The systems that generate synthetic image, video, audio or text have to mark their outputs in a machine-readable way, so the file itself signals that it was made or altered by AI.
Whoever publishes certain content, most clearly AI-generated or manipulated video and audio of people, has to label it so the viewer can see.
In practice the invisible layer is provenance written into the file: a cryptographically signed record of what was generated and how it was changed, combined with an imperceptible watermark in the pixels or the audio waveform. For clips too short or too heavily re-encoded to hold either, a fingerprint logged to a registry so the asset can still be traced.
The visible layer is what the audience sees, and the finalised Code of Practice points to a standardised EU label paired with a short, plain disclosure, something close to “this content was created or manipulated by AI”, shown at the point of first exposure. Picture it across one global campaign:
A TV and online spot where a real actor is re-voiced into German by a cloned voice: a brief on-screen line at the head of the film, plus signed provenance baked into the master.
A social cutdown where the talent's lips are reanimated to match the dubbed track: the disclosure carried in the caption and the platform's own AI tag, with marking built to survive a re-upload that strips metadata.
A fully generated product scene in a display banner: the credential rides in the file, and where the ad server discards it, the fingerprint fallback does the work.
A digital out-of-home screen with a generated backdrop behaves like video and can carry all three layers; the printed poster beside it can carry none, so the only record of how it was made lives in your production system, not the artefact.
That is the real exposure for international marketing. The same master can need a visible label in one market and not another, machine-readable marking in every EU territory, and a different judgement about what even counts as “AI-generated” at each step. Made ad hoc, market by market, those decisions produce not one clean failure but thirty slightly different answers, and inconsistency erodes brand coherence more quietly than any single breach.
On recent figures nearly four in five organisations now use generative AI somewhere in their work, yet only a minority have any formal policy on how those tools are used or how their outputs are disclosed. Adoption has sprinted well ahead of the discipline for handling it.
The organisations that take 2 August 2026 in their stride will be the ones treating disclosure as a workflow discipline: deciding before generation which assets are in scope, confirming the tools actually emit machine-readable marking, building the label logic into the master so it cascades through every version, and recording provenance as the asset moves rather than reconstructing it under deadline.
Because provenance is now a property of the asset, in the same way a colour profile or a safe area is. It travels with the file, it has to be right in every version, and it cannot be added at the end. The technology can mark the file. It cannot decide what was meaningfully AI-generated, what a given market expects, or whether thirty versions tell the same story about how they were made. That judgement, and the system built around it, is the part the deadline cannot automate. The window to build it is open now, but it's narrowing!
AI voiceover means generating spoken audio from text or rebuilding a performance in another language without a recording booth. What is interesting is how fast it has travelled. For thirty years, a synthetic voice gave itself away in the first syllable but today - it can pass for a person. This is a shift changing the economics of reaching every market you operate in.
In the past this technology stitched voices together from pre-recorded fragments: functional, unmistakably mechanical, the sound of an old screen reader. That changed around 2016, when models began generating raw audio directly and the level of realism jumped, closing a big part of the gap to human speech. However, it still stayed emotionally flat and able to work with a single speaker only. By 2023, systems started to detect multiple speakers, but separating them needed manual work and the result rarely passed for professional talent.
However, in controlled blind tests today, listeners can no longer reliably tell synthetic voices from real ones, and some are even rated as more trustworthy. Costs have fallen by up to 90%, timelines have dropped from weeks to days making AI voice an amazing localisation solution for anyone running campaigns across dozens of markets.
Important to note that what it still cannot do is the part that matters. Complex emotional delivery, the cadence that marks a true native speaker, and the consistent vocal identity that makes a brand recognisable across borders all still depend on human direction.
One day the gap may close entirely: a single performance, re-voiced live into forty languages, each one sounding native. Until then, the technology supplies the speed. People still supply the voice.
Governing brand at scale is the bigger question underneath all this. We discussed it with a room of senior marketing leaders at our last Roundtable, and it produced ten themes, most still unsolved across the industry.