The AI Act is entering its operational phase, humanoid robots are making headlines for the wrong reasons, and generative models are facing new safeguards. It has been a busy week for the tech sector, with direct implications for software architectures and compliance strategies for companies in Europe.
Transparency Obligations of the AI Act: What August 2, 2026 Changes for Deployed Systems
Article 50 of the AI Act is now fully applicable as of August 2, 2026. Any AI system that interacts directly with a user, such as a chatbot, voice assistant, or virtual agent, must clearly inform the user that they are interacting with an AI. The obligation is not limited to a simple disclaimer: the synthetic content generated (text, image, audio, video) must be marked in a machine-readable format.
The penalties can go up to 15 million euros or 3% of global revenue for general-purpose models. Generative systems already on the market before this date benefit from a grace period but must integrate automatic marking by December 2, 2026, at the latest.
We observe that many SaaS publishers have not yet implemented machine-readable watermarking on their text outputs. The compliance window is shrinking to less than three months. For those following tech news on Les News, this topic deserves close monitoring until the December deadline.
Digital Omnibus Regulation (EU 2026/1744): A Targeted Stay, Not a Delay of the AI Act
The regulation (EU) 2026/1744, published in the Official Journal on July 24, 2026, and effective from July 27, 2026, has generated confusion in the sector. Some actors have interpreted it as a blanket delay of the obligations of the AI Act. This is incorrect.
The Digital Omnibus grants an additional delay only for certain categories of systems, without affecting the transparency obligations of Article 50 or the prohibitions already in effect. SMEs and research projects benefit from procedural adjustments, but large deployers remain subject to the original timeline.
The distinction is technical but strategic: a company developing a B2C conversational agent cannot invoke the Digital Omnibus to postpone the marking obligation. We recommend checking precisely which category each deployed system falls into before planning a compliance timeline.
Points of Caution for Product Teams
- Machine-readable marking is not just a metadata tag: it must withstand common transformations (resizing, format conversion, copy-pasting) to comply with the technical specifications being finalized by the European AI Office.
- Hybrid systems that combine human output and AI output pose a threshold problem: at what percentage of generated content does marking become mandatory? The text of the regulation does not resolve this point, and the CNIL guidelines do not yet provide a clear answer.
- Third-party APIs used in the backend (summary generation, machine translation, text-to-speech) engage the responsibility of the final deployer, not the API provider. The chain of responsibility is upward.
Humanoid Robots and Cybersecurity: This Week’s Incidents
A Unitree humanoid robot struck its engineer during a test in China, reigniting the debate on physical safety protocols in bipedal robotics. The incident, captured on video, shows a martial arts-trained robot delivering a kick during a calibration phase.
The problem is not the embedded AI but the motor control layer. On current humanoid robots, safety systems rely on joint torque limits and virtual exclusion zones. When the movement model is trained on combat sequences, these classic safeguards become insufficient.
On the cybersecurity front, Anthropic has released a report detailing detected and blocked malicious uses on Claude: espionage attempts, content generation related to weapon manufacturing, automated scams. This type of proactive disclosure remains rare in the sector and raises questions about the expected transparency standard from other model providers.
Regulation of Generative AI: OpenAI and Anthropic Converge on Frameworks
OpenAI has announced it is considering slowing down the pace of development of its models. Sam Altman reportedly confirmed this direction to his teams. Almost simultaneously, both OpenAI and Anthropic are calling for the establishment of rules governing the development of AI, a position that contrasts with the historical stance of the sector.
This convergence is not selfless. Both companies anticipate that the European regulatory framework (AI Act, Digital Omnibus) will become a global reference. Participating in drafting technical standards gives them a structural advantage over Chinese actors and open-source projects that lack the resources to sit on standardization committees.
What This Implies for the Cloud and Enterprise Ecosystem
Companies integrating generative AI models via cloud services must monitor two variables: the compliance of the model itself and the compliance of their use of the model. A compliant cloud provider does not guarantee that the final deployer is compliant.
The service level agreements (SLAs) of major cloud providers do not yet cover the liability related to AI Act marking. This contractual gap exposes client companies to direct legal risk in the event of an audit after December 2026.
This week confirms a fundamental trend: tech is no longer just about product innovation. The European regulatory layer is becoming a full-fledged design parameter, alongside performance and security. Teams that do not integrate AI Act compliance into their product roadmap now are accumulating technical debt that will be costly to address.
