Only last year, a video with an AI avatar was a guarantee of bad advertising. Today Coca-Cola, Samsung and Nestlé use it as a standard campaign format. What changed, and why creators are now rewriting their contracts.
In February 2026, Kuaishou released the Kling 3.0 model. Five days before that, its ARR had passed $300 million, for January alone. By March it had grown to an estimated $500 million. The whole of 2023 was needed before the first generation of AI video tools convinced marketing agencies that the output was worth producing. Today it takes them one month.
Kling 3.0 brought native 4K output, multi-shot clips up to 15 seconds and multilingual audio with lip sync. A week later, on 12 February, ByteDance released Seedance 2.0, a model able to take up to 9 images, 3 videos and 3 audio files at once, and to generate 1080p video with native dialogue and synchronised sound effects in a single pass. Meanwhile, Google updated Veo 3.1 in January to true 4K (3840×2160 px, up to 60fps) and 48kHz synchronised dialogue, as the only model in the field that generates soundtrack and speech together with the video, without a separate audio pipeline.
A new brand deal format
The real change is not technical. It is commercial.
Agencies such as BridgeGlobal now offer clients a process that was logistically impossible before. The voice and face of a chief executive are cloned once, and then the same video is generated in Spanish, French, German and Portuguese, each with natural lip sync and each for a different market. Gan.AI, one of the leading tools in this segment, lists among its clients Coca-Cola, Samsung, Amazon and Nestlé. Platforms such as Synthesia or HeyGen deliver multilingual output in 70 or more languages.
The average cost of producing a video fell from roughly $4,500 (traditional studio) to roughly $400 (AI), a drop of about 91%. A flat monthly fee for a whole multilingual campaign replaces a project budget. Brands are buying a format, not a specific shoot.
Creators and the contract loop
The boilerplate “likeness clause”, a standard contract provision on a person's image, is being rewritten during 2026 to include perpetual rights for AI training. In other words, a brand pays once and may use a creator's likeness to build and run AI models, potentially forever. Face, voice and gestures in a package for one flat fee. Creators who read this in their contracts are now demanding time-limited licences, approval rights over every AI-generated output and revenue participation clauses tied to the volume of use. Anyone who does not do this has a good chance of finding that a face they signed for one campaign is still running in seven markets five years later.
The open-source layer of this market works as a pressure gauge. LongCat-Video-Avatar 1.5, an MIT-licensed framework from Meituan released in May 2026, can generate a lip-synced clip with a consistent identity from one photograph and one audio file. Running it requires a 40GB GPU and roughly 44 seconds of computation for every second of resulting video. The barrier is not the law but the hardware. Anyone who has it, and agencies do, can replicate a likeness practically without limits.
Transparency as a rule
On 2 August 2026, Article 50 of the EU AI Act came into force. It requires providers of AI systems that output synthetic videos, images or audio to ensure machine-readable labelling of the content. This is an invisible watermark embedded directly in the file. Google implemented this mechanism in Veo 3.1 through SynthID, which survives compression, transcoding and cropping. By May 2026, more than 100 billion items had been marked with SynthID. OpenAI, ElevenLabs and Kakao have also adopted the standard. Penalties for failing to label reach up to €15 million or 3% of worldwide turnover, whichever is higher.
Who profits
The ability to generate multilingual video content cheaply does not automatically mean creators lose work. It means what they are paid for changes. An agency today does not need a creator to shoot a video in German. It needs them to provide their likeness and approve the output. Value shifts from production to identity. Whoever owns and controls their identity profits. Whoever sells it once for a perpetual flat fee finds out only later, in the contract terms.
These are the rules of a format that did not exist last year.
Comments