On 31 July 2026, the European Commission announced that the AI Act would expand its scope from 2 August, applying the transparency rules and the rules for providers of general-purpose AI models. The regulation, which entered into force on 1 August 2024 with a staged application timetable, has reached its broadest stage of application — most provisions apply from 2 August.
For Chinese AI companies offering products or services to the EU market, 2 August is not a date in a headline but a substantive expansion of the list of obligations. Which systems must disclose, which content must be labelled and which model providers must carry systemic-risk duties now have an enforceable basis.
A staged timetable reaching a key milestone
The AI Act has applied in stages since entering into force on 1 August 2024: first prohibitions and definitions, then the rules on general-purpose AI models and the governance architecture, with most provisions deferred to 2 August 2026. The current expansion is the fulfilment of that timetable.
The regulation is risk-based, dividing AI systems into unacceptable risk, high risk, transparency risk and minimal or no risk. After 2 August, the transparency-risk tier is the first to apply in full — and it covers not a handful of high-risk scenarios but a large number of general-purpose AI applications aimed at the public. For an earlier walk-through of compliance pathways, see our EU AI Act service page.
Transparency rules: two hard requirements
The transparency rules centre on two obligations.
The first is interaction disclosure. Interactive AI systems such as chatbots must inform users that they are interacting with AI, not a human. This targets the experience design of “disguising AI as a person” — users are entitled to know that the other end of the conversation is not human.
The second is content labelling. Image, video and audio content generated or modified by AI — “deepfakes” — must be labelled, and AI-generated or AI-modified content must carry machine-readable markings so that it is easy to identify. Labelling is not only a visible notice for people; it must also reach the machine-readable layer, which means marking must be embedded at the content-production stage rather than added as a statement afterwards.
The Commission states that these rules aim to reduce deception and manipulation. For product teams, this means the prompt at the interaction entry point, the output flow for generated content and the metadata marking all need to be part of compliance design, not optional extras.
General-purpose AI model providers: from transparency to systemic risk
In parallel with the transparency rules, from 2 August the European AI Office can enforce the AI Act’s rules against providers of general-purpose AI models.
For advanced general-purpose AI models that may pose systemic risk, providers must also meet additional obligations to guard against large-scale harm — for example cyberattacks, loss of model control, harmful manipulation and infringements of fundamental rights. This goes beyond transparency disclosure: providers must build risk-management, evaluation and reporting mechanisms commensurate with model capability.
For products built on foundation-model or large-model capabilities, this signals that whether an upstream model provider complies will flow down to downstream applications. When procuring or integrating a general-purpose model, the upstream compliance status should be part of supplier assessment.
High-risk system rules deferred: two new dates
One deferral is worth noting. Through the AI-related part of its “omnibus simplification” package, the EU deferred the rules for high-risk AI systems used in recruitment, education, law enforcement and migration management to 2 December 2027, and for high-risk AI systems embedded in regulated products such as medical devices and machinery to 2 August 2028.
Deferral is not exemption; it grants a longer preparation period. Companies in these scenarios can slow the pace but cannot strike compliance from the roadmap.
What it means for Chinese companies going global
First, classify the system. Whether the transparency rules apply depends on whether the product constitutes an interactive AI system and whether it generates or modifies content. Classification is the starting point for obligations and the basis for all subsequent work.
Second, build labelling into the product flow. Visible notices and machine-readable markings must be completed at the production stage; retrofitting later often means changing the architecture.
Third, bring the upstream model into the assessment. Applications integrating general-purpose models need to understand the upstream provider’s compliance status to prevent risk flowing down the supply chain.
The EU’s transparency requirements are being read alongside regulatory developments in the UK and the US. Companies operating across multiple jurisdictions should place the EU labelling duties and other jurisdictions’ requirements on a single compliance map; see our EU AI Act and Multi-Jurisdiction Compliance services.
This article is based on the European Commission communication of 31 July 2026 and press reporting. It is for reference only and does not constitute legal advice. Specific compliance judgements should be based on the official text of the AI Act and official EU guidance.