What is the EU AI Act?
The AI Act is the first legal framework on AI, which lays down harmonized rules on artificial intelligence in the EU and sets out risk-based rules for AI developers and deployers regarding specific uses of AI.
How to Navigate the EU AI Act
We recommend going straight to the source for the clearest understanding of the Act and to stay up to date: AI Act Explorer | AI Act Service Desk
What is the AI Act Timeline Currently?
The original timetable has been amended, pushing out some timelines into 2027 and 2028. The legislation simplifying and delaying some parts of the Act was passed in July 2026 and is known as the Digital Omnibus)
| 2 August 2025 | GPAI model provider obligations (Articles 51-56) take effect | In force |
| 2 August 2026 | General AI Act application date; Article 50 transparency obligations take effect | In force |
| 2 December 2026 | Grace period ends for machine-readable watermarking on generative systems; new Article 5 prohibition on restricted (criminal) AI-generated imagery | Upcoming |
| 2 August 2027 | National sandboxes to be operational; further guidelines from Commission | Moved from 2 August 2026 |
| 2 December 2027 | High-risk obligations apply to stand-alone systems (Annex III) | Moved from 2 August 2026 |
| 28 January 2028 | Notified bodies must apply for AI Act designation | |
| 2 August 2028 | High-risk obligations apply to AI embedded in regulated products (Annex I: medical devices, machinery, vehicles, aviation) | Moved from 2 August 2027 |
| 2 August 2030 | Deadline for public authority high-risk systems already in use to reach compliance |
Impact of In-Force Regulations
GPAI and Transparency rules are now in force and are summarised below.
1. GPAI (General Purpose AI) model provider obligations (Articles 51-56)
GPAI models are defined as those with a general purpose and which can be integrated into other systems and where the “generality of the model could, inter alia, also be determined by a number of parameters, models with at least a billion parameters and trained with a large amount of data using self-supervision at scale.”
All GPAI model providers (which, to the best of our knowledge do not include any Machine Translation providers) now have to comply with Article 53 of the Act, which focuses on technical documentation (including testing and training, policies to comply with EU copyright law and cooperation with the authorities. Additional safety and security measures are required from “systemic risk” systems (compute exceeds 10^25 flops, or designated as high-impact by the Commission) under Article 55.
Many of the companies seeking to comply with the GPAI obligations have signed up to a Code of Practice, which can be read here: The General-Purpose AI Code of Practice | Shaping Europe’s digital future and which covers three main areas: transparency, copyright, safety and security.
Impact for language service providers and customers: seek confirmation whether model providers are classified as GPAI and the extent of their compliance with current regulations and the GPAI Code of Practice.
2. Transparency Obligations (Article 50)
These obligations do interact with language services activities and can be read directly here: Article 50: Transparency obligations for providers and deployers of certain AI systems | AI Act Service Desk or here: Quick Facts: Transparency rules for AI systems | Shaping Europe’s digital future.
The transparency rules focus on ensuring that:
- individuals are explicitly informed when they interact with an AI system directly;
- machine-readable watermarks are added to AI-generated content or manipulated content. The content that needs to be marked includes text, images/video and audio;
- individuals are informed when they are exposed to emotion recognition and biometric categorization tools, deepfakes and text publications on matters of public interest without human review or editorial control.
In addition to Article 50, the Commission has provided further Guidelines which clarify when these rules apply or not to AI translated content.
- Article 50(2) Marking and Detection of AI Generated or Manipulated Content does not apply if “1) the AI system performs an assistive function, 2) the AI system does not substantially alter the input data provided by the deployer or the semantics thereof.”
- AI automated translation or transcription tools not directly interacting with natural persons fall outside the scope of Article 50(1).
- Regarding texts published with the purpose of informing the public on matters of public interest, AI-supported translation of a human-written article is exempt if it has undergone human review.
“Public interest” is defined in the Guidelines and is not, on the surface, referring to general commercial communications.
Impact for language service providers and customers. There are some specific cases where transparency and watermarking rules will apply – such as a multilingual chatbot, or text and audio that is more than just editing a source (or could be deceptive). Content types and use-cases will need careful review. The exemptions and clarifications provide an indication of future developments: that non-material adaptation of human-reviewed source material, or human review in the loop, will be considered mitigations.
Conclusion
Two major parts of the EU AI Act are now in force: GPAI obligations since August 2025 and Article 50 transparency rules since August 2026. These are unaffected by the delay to the high-risk rules introduced through the Digital Omnibus. Language service customers should focus now on understanding what models are being used in their workflows and, secondly, work with their providers to review transparency requirements and workflow mitigations.
