Get weekly AI video rankings, evaluation updates, and industry news delivered to your inbox.
The European Commission's AI Office begins enforcing rules on AI-generated fakes, chatbots, and training data disclosure
On August 2, 2026, the European Commission's AI Office began enforcing the transparency provisions of the EU AI Act. The rules require companies to tell users when they are talking to an AI, label AI-generated content, and disclose what data was used to train general-purpose models.
The European Commission will oversee general-purpose AI models directly, mandating that providers publish summaries of training content and maintain copyright policies. Fines for non-compliance apply, though their scale is not specified.
What companies must do now
The rules cover a broad range of outputs and interactions. Publishers must label AI-generated fakes and other AI-manipulated media. Providers of AI systems that produce synthetic audio, images, video, or text must embed machine-readable markers in those outputs. Any system that interacts with users must disclose that the user is communicating with an AI.
The European Commission released implementation guidelines in late July 2026 to help providers and deployers interpret these obligations. Those guidelines specify that the machine-readable markers are meant to enable detection by third parties, not just disclosure to end users, a distinction that matters for platforms trying to filter synthetic content at scale.
The watermarking gap
The requirement to watermark AI-produced content in a machine-readable format does not take effect until December 2026, giving providers roughly four months to build or integrate compliant systems.
That gap reflects a practical constraint. Standardized watermarking for AI-produced content remains an unsolved technical problem across the industry. The Act sets the deadline but does not specify the technical standard. How the AI Office will evaluate compliance in the absence of a mandated format is an open engineering and regulatory problem that December will force into view.
Training data and copyright
The disclosure requirements for general-purpose AI models go beyond labeling. Providers must publish summaries of the content used to train their models and implement copyright policies governing that training data.
These requirements sit at the intersection of transparency and intellectual property law, where litigation over AI training practices is already active across multiple jurisdictions. The Commission's rules do not resolve those underlying disputes, but they create a paper trail. Companies will now have to state, on the record, what their training data policies are.
The next checkpoint arrives in December 2026, when the watermarking deadline closes the remaining gap in the Act's transparency framework.
Get weekly AI video rankings, evaluation updates, and industry news delivered to your inbox.