At a glance:
- What’s changed: Article 50 of the EU AI Act came into force on 2 August 2026, introducing transparency obligations for publishers of AI-generated and AI-modified text.
- Who it applies to: Publishers (‘deployers’) of AI-generated text on matters of public interest in the EU, regardless of where they’re based.
- The obligation: Text generated or modified in AI-only pipelines must carry an AI label.
- The exemption: If a human with relevant subject matter expertise reviews the substance of the text and a named person or organisation assumes editorial responsibility, no AI label is required.
- The penalty: Fines of up to €15 million or 3% of global annual turnover, whichever is higher.
- How Sandberg can help: We can identify your in-scope texts and carry out human reviews, which when paired with your editorial responsibility qualifies for the AI label exemption.
Introduction
Do your texts need an AI label? Since 2 August 2026, some might.
The transparency obligations under Article 50 of the EU AI Act are now in force. They apply to publishers of AI-generated or AI-modified texts, not just the companies building the AI systems, and they cover any texts published in the EU on matters of public interest.
In this post, we explain the new AI labelling obligations for publishers, when they apply and the exemption criteria.
What happens if you don’t comply with the AI Act transparency obligations?
From 2 August 2026, texts that meet the AI Act’s transparency obligations must carry an AI label. Texts published prior to this date do not need to be labelled — in other words, the Act is not retroactive.
If you fail to label texts that fall under the obligations, you may be liable for a fine issued by the competent market surveillance authority in the relevant member state. The fine may be up to €15 million or 3% of worldwide annual turnover for the preceding financial year, whichever is higher, though proportionality applies and regulators typically offer a grace period for correction. We don’t think this is a risk worth taking.
What are the transparency obligations for AI text?
The transparency obligations for publishers of AI-generated or AI-modified texts are laid out in Article 50(4) (our emphasis):
‘Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated. This obligation shall not apply where […] the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.’
The obligations apply only if the subject of your text is of public interest.
If it is, there are two scenarios:
- The text is generated or modified with AI and has no human input: you need to add the AI label.
- The text is generated or modified with AI and reviewed by a human: you do not need to add the AI label. In this case, a named person or organisation must assume editorial responsibility for the text.
If you’re operating an AI-only pipeline, your text must carry an AI label. If there’s a human in the loop, it’s not necessary.
That’s the top line. Now let’s look more closely at what ‘public interest’, ‘human review’, ‘editorial control’ and ‘editorial responsibility’ mean.
What does ‘public interest’ mean?
The European Commission’s FAQs on Article 50 define ‘public interest’ as politics and democratic processes, public administration and services, administration of justice and law enforcement, fundamental rights, public security, public health, environmental protection, consumer safety or any economic, financial, political, scientific or cultural developments that may be the relevant subject of public debate.
In practice, this may include:
- Public notices about scams, phishing campaigns or warnings about high-risk products
- Financial prospectuses and annual reports
- Terms of service for financial products
- Product recalls
- News articles covering corporate misconduct
- Medical device instructions for use (IFU) and patient information leaflets
- Scientific research summaries
- Corporate ESG and sustainability disclosures
There are, of course, many other text types considered to be of public interest and that therefore must meet the transparency obligations.
What do ‘human review’ and ‘editorial control’ mean?
The same European Commission FAQs define ‘human review’ as (our emphasis):
‘[…] the deliberate examination of the substance of the content by one or more natural persons possessing relevant knowledge and professional judgement pertaining to the subject matter under scrutiny […].’
The review must concern the subject matter of the text and be carried out by a person with relevant expertise. It is not a quick linguistic proofread, checking the spelling and adding the odd comma here and there.
Editorial control, on the other hand (our emphasis):
‘[…] refers to the control exercised in practice by a responsible editorial entity (e.g. an editor-in-chief) over the content having the authority to approve, alter or reject the substance of the text based on substantive grounds (including factchecking of information and ensuring the trustworthiness of sources).’
This is about the decision of what to publish — how it is framed, what is cut, what is corrected.
You need either human review or editorial control. Whichever you select, it must be overseen by editorial responsibility.
What does ‘editorial responsibility’ mean?
According to the same FAQs, editorial responsibility means (our emphasis):
‘[…] that a person must hold the ultimate legal responsibility over the publication of the content […].’
An AI-generated or AI-modified text of public interest, which has been reviewed or edited by a human, must also have a named human or organisation responsible for it.
Glossary: public interest, human review, editorial control, editorial responsibility
| Term | Definition |
| Public interest | Concerning politics, public services, justice, rights, security, health, the environment, consumer safety and matters of public debate |
| Human review | A review of the substance of the text by a person with relevant subject matter knowledge |
| Editorial control | The decision of what is published |
| Editorial responsibility | A named person or organisation that holds legal responsibility for the publication of the text |
Do the transparency obligations apply to AI-generated translations?
Yes. Under the Commission’s guidelines, an AI-generated translation of a human-written original is treated as AI-generated or AI-modified text and falls under Article 50.
To be exempt from the AI label, your AI-generated translation of a public interest text must undergo human review or editorial control, with a named person holding editorial responsibility.
It is worth clarifying that the human review must be conducted in each of the languages the original is being translated into – not just on the original text. Editorial clearance of your English original doesn’t carry over to the Swedish, French or German versions. Each translation needs its own qualified review.
Do the transparency obligations apply to my text?
There are three possible scenarios:
- Your text isn’t of public interest. The AI Act doesn’t apply. No action needed.
- Your text is of public interest and has been reviewed or edited by a human with subject matter expertise before publication. No AI label required, but a named person or organisation must hold legal responsibility for the text.
- Your text is of public interest and hasn’t been reviewed or edited by a human before publication. You must add the AI label.
AI Act compliance cheat sheet: do my texts comply with the AI Act transparency obligations?
Check whether your AI-generated or AI-modified texts require an AI label at a glance. Download the guide and save it for later reference.
How can Sandberg help you comply with AI transparency obligations?
Once you’ve shared your content with us, we can offer consultancy to identify which of your texts fall under Article 50 transparency obligations. Public interest is in scope, internal materials and non-public content typically isn’t.
For the texts that would otherwise require an AI label, we conduct a human review. Paired with your editorial responsibility for publication, our review exempts you from the labelling obligation.
The substance of your texts is reviewed by linguists with expertise in your industry – financial services, life sciences, manufacturing, marketing or software development. Medical device specialists review the instructions for use for a new health product. Financial specialists review warnings about high-risk products. The right expertise is matched to the right text.
The reviewers come from our team of around 50 in-house linguists with an average of 10 years’ experience or from our carefully vetted pool of ISO 17100-compliant freelance reviewers. The same named linguists work on your account across texts to ensure consistency and accuracy. If you have any questions for them, we’ll put you in touch.
Your dedicated project manager is on hand throughout, keeping the process on track and communication clear.
Once the review is complete, we issue a certificate listing the texts reviewed and the linguists who reviewed them. This type of documentation is what regulators expect: evidence that qualified human review took place, by whom and when. Keep it filed alongside your published content and you have a clear paper trail for every AI-generated or AI-modified text you put out.
We’ve fine-tuned this process over 30 years, backed by ISO 17100 and ISO 18587 certification for translation project management and post-editing, held since 2015 and 2018 respectively.
Example: a Nordic fintech across four EU markets
A Swedish fintech operates in Sweden, Denmark, France and Germany. They’ve used AI to generate customer-facing content – risk warnings, terms of service, marketing copy – directly in all four languages and need to make sure the texts comply with Article 50 before being published.
Here’s how the process works:
- Consultancy. We start by identifying which of the texts fall under Article 50. Financial disclosures, risk warnings and terms of service almost certainly do. Internal training materials and marketing copy probably don’t.
- Human review of in-scope content. Our Swedish, Danish, French and German linguists – all with financial services expertise – review the substance of the texts. Each reviewer is named and their qualifications meet the knowledge requirements laid out in Article 50.
- Editorial responsibility. The fintech retains editorial responsibility for what gets published. Our review provides the qualified human input that lets the fintech claim the AI label exemption.
- Documentation. At the end of the review, we issue a certificate listing the texts reviewed and the linguists who reviewed them. That gives the fintech a clear audit trail if a regulator ever asks.
- Ongoing. The same linguists handle the fintech’s content going forward. Terminology stays consistent across texts and markets and the fintech has a stable, contactable team who knows their products.
Stay compliant with AI transparency obligations with Sandberg
The AI Act’s transparency obligations are now in force and the AI label exemption route is clear — qualified human review, editorial responsibility and a documented audit trail. If you’re publishing AI-generated or AI-modified text, now is the time to see where you stand.
Schedule a complimentary Compliance Review with one of our experts. We’ll walk you through your content, identify what falls under Article 50 and set out the practical steps to compliance.
Disclaimer: this article reflects our reading of the AI Act and is not legal advice. For compliance decisions, refer to the text of the AI Act and consult qualified legal counsel.
While we’re on the topic of transparency, Sandberg uses AI across a range of processes, including content production. This blog post is no exception. In line with Article 50, every piece of public interest content we publish is reviewed by a named human with relevant subject matter knowledge and a named person on our team holds editorial responsibility for it.
Frequently Asked Questions
Article 50 came into force on 2 August 2026. Texts published before this date do not need to carry an AI label. The Act is not retroactive.
It applies to publishers (‘deployers’) of AI-generated or AI-modified text on matters of public interest published in the EU, regardless of where the publisher is based. It’s not limited to the companies that build the AI systems.
AI-generated translations of public interest human-written texts are considered AI-generated texts under the AI Act. To be exempt from AI labelling, publishers must conduct a human review of the translated text. Reviewing the original is not enough.
Sandberg identifies which of your texts fall under Article 50, carries out human review with subject matter experts in your industry and issues a certificate listing the texts reviewed and the linguists who reviewed them. Paired with your editorial responsibility, our review exempts you from the AI labelling obligation and gives you the audit trail regulators expect.
Fines can reach up to €15 million or 3% of worldwide annual turnover for the preceding financial year, whichever is higher. Proportionality applies and regulators typically offer a grace period for correction, but the reputational and financial exposure is real.
Artificial intelligence, Clinical trials, Financial translation, Fintech, Machine translation, Manufacturing, Medical devices, Pharmaceuticals
