The regulatory landscape for Artificial Intelligence has shifted, and so did the compliance timeline for businesses relying on multilingual operations.

While the broader European rollout of high-risk AI rules has been delayed, transparency obligations are already live. For organisations using AI for localisation and translation, compliance is now more critical than ever.

This article analyses what has changed, what rules apply currently, and how to build a defensible, compliant translation workflow.

Recent regulatory changes

• The EU AI Act entered into force with a phased rollout, but the initial timeline proved too ambitious. Recognising that industry, standardisation bodies (CEN and CENELEC), and conformity assessment infrastructures required more time to prepare, the Digital Omnibus on AI was introduced in late 2025 and officially approved in mid-2026.

The revised timeline is disclosed as such:

• High-risk obligations (Standalone Annex III Systems): Deferred from August 2026 to 2 December 2027.
• High-risk obligations (AI Embedded in Regulated Products, Annex I): Deferred to 2 August 2028.
• Article 50 transparency rules: left out of the deferral entirely. They became active on 2 August 2026, and national market surveillance authorities are actively enforcing them.

Who does this apply to

The transparency duties under Article 50 apply to any business using generative AI to produce content; they are not restricted to systems classified as high-risk. Furthermore, the Act has extraterritorial reach: it applies globally to providers, deployers, importers, and distributors who place AI on the EU market or whose AI outputs are utilized within the European Union.

Violations carry severe financial penalties, with non-compliance triggering fines of up to €15 million or 3% of worldwide annual turnover.

What Article 50 means for AI localisation

Article 50 outlines four key situations requiring strict transparency. For translation and localisation teams, two areas demand immediate attention: 

1. AI-generated content marking: AI-generated or manipulated content must be clearly marked and technically detectable as artificial. 
2. Public-Interest disclosures: Deepfakes and AI-generated text published to inform the public on matters of public interest must be explicitly disclosed as artificially generated. 

Practical implications for localisation: 

• Public-facing material: If your organisation translates public-facing material concerning matters of public interest using AI (e.g., public health updates, financial disclosures, consumer safety notices), disclosure is legally mandatory now. 
• The grace period exception: Systems placed on the market before August 2, 2026, had a brief grace period until 2 December 2026 to meet marking and detection obligations. Content generated prior to August requires no retrospective labelling, but everything moving forward must comply. 
• The “accuracy fallacy”: A disclosure label satisfies Article 50, but it does nothing for translation quality. Marking a document as machine-generated fulfills regulatory transparency, but it does not protect you from a mistranslated indemnity clause or a critical error in a localised legal contract. 

Why act now?

While the delay to December 2027 for high-risk systems sounds like a reprieve, it is misleading. 

This is because the EU AI Act ultimately expects an evidence trail. When regulators or courts examine your localised assets, they will not just ask if you used AI- they will demand to see who reviewed what, when, against which source, and with what professional qualifications. 

This audit trail cannot be assembled retroactively in 2027; it must accumulate organically from the exact day you begin utilizing AI in your translation pipeline. 

Three steps to take before December 

To protect your organisation from regulatory exposure and operational liability, focus on three foundational steps: 

1. Audit your AI translation footprint: Map out exactly where AI translation, machine translation (MT), and generative localization are being used across your entire organization. 
2. Classify content by consequence: Do not treat all text equally. Internal notes and low-impact product descriptions carry entirely different risk profiles than courtroom bundles, compliance documents, or clinical trials. Treating them identically wastes resources on low-risk items while exposing you dangerously on high-risk ones. 
3. Vet your Language Service Providers (LSPs): Ask your translation partners two critical questions: 
   1. What happens to your data when it passes through their systems? (Data sovereignty and privacy). 
   2. Can they produce an immutable record of who verified a given translation? (Auditability and human oversight). 

Sovereign AI with verification: GAI Translate  

Regulators will not ask whether you used AI. They will ask who was accountable for what it produced. 

GAI Translate was engineered from the ground up for high-risk settings where absolute accuracy and human verification are mandatory. Eschewing the approach of retrofitting legacy infrastructure with basic AI tools, GAI Translate is built natively around: 

• Sovereign deployment: Complete control and security over organisational data. 

• Expert human assurance: Seamless integration of certified human linguists. 

• End-to-end auditability: Granular tracking of every review cycle. 

Low-risk content flows effortlessly at AI speed, while high-risk output automatically escalates to a vetted human expert for verification or certification. The verifiable record of that review stays permanently tethered to the document, ensuring your localised outputs are fully defensible, brand-safe, and compliant.

Start a free trial with us today.