On July 30, 2026, the literary world was rocked by a scandal: the agent of Nigerian writer Jerry Falade terminated his contract and withdrew his novel amid suspicions of artificial intelligence use. The manuscript, fiercely contested by 14 publishing houses and potentially worth over $2 million to the author, became the center of a dispute over the text's origin. This incident became a perfect illustration of why the incoming labeling rules for synthetic content create new risks for "living" authors.
On August 2, key provisions of Article 50 of the EU AI Act officially came into force in the European Union. For the first time in global practice, labeling AI-generated content became legally mandatory. The regulation requires machine-readable tags and digital watermarks for any synthetic audio, video, photos, and text, as well as mandatory disclosure of interaction with AI systems and clear labeling of deepfakes depicting real people or significant events.
Europe is not alone in this endeavor. China, since autumn 2025, has already implemented its own standards, requiring WeChat, Douyin, and other platforms to automatically tag neural content. In the US, New York has been fining the use of AI-synthesized actors in advertising since June 9, 2026—from $1,000 for a first offense to $5,000 for repeat violations. California, in turn, launched the AI Transparency Act (SB 942), synchronizing it with the European deadline.
Technical solutions and their vulnerabilities
The industry relies on three main approaches: the C2PA standard (a coalition of Adobe, Microsoft, Intel, and others), invisible watermarks like SynthID from Google DeepMind or AudioSeal from Meta, and statistical text labeling implemented in Gemini. However, none of them are bulletproof. C2PA tags are easily removed when a file is re-saved in editors that do not support the standard. Midjourney, for example, categorically refuses C2PA, limiting itself to easily erasable metadata.
The problem is especially acute with text. The Falade incident showed that even when publishers suspect AI, there is no technical confirmation. Decisions are made based on probabilistic assessments from detectors and reputational risks. This creates a dangerous precedent—a presumption of guilt. Authors using AI for editing or fact-gathering risk finding themselves in a situation where the absence of a tag is perceived as proof of authenticity, while its presence is a verdict.
The blind spot of Open Source and market realities
Open models like Llama, Qwen, or Stable Diffusion remain beyond the reach of regulators: they can be downloaded, modified, and run locally, disabling built-in labeling. Even reliable watermarks, like those in Claude from Anthropic, do not survive deep paraphrasing or translation. This means spammers and bad actors will find workarounds, while conscientious users will not.
Platforms are already reacting: YouTube automatically tags synthetic videos, and Spotify is introducing an AI Persona badge for artists starting in September, removing them from recommendations. Music streaming service Deezer reports that 44% of new tracks are AI-generated, and 97% of listeners cannot distinguish them from human ones. This pressure will inevitably intensify filtering and corporate paranoia.
My conclusion: The new laws create an illusion of control, but in reality, they shift the burden of proof onto humans. For images and videos, labeling works; for text, it does not. Authors, editors, and programmers using AI in their work will find themselves at risk not because of fines, but because of reputational accusations that cannot be technically defended against. Until reliable methods of text verification emerge, the market will live in a state of permanent distrust, where honest professionals suffer first, not spammers.