What EU AI Act Article 50 Actually Requires
Article 50(2) of the EU Artificial Intelligence(AI) Act requires that providers of AI systems generating synthetic text, images, audio, or video ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.Future of Life Institute, “Article 50 Transparency Obligations for Providers and Deployers of Certain AI Systems EU Artificial Intelligence Act.” https://artificialintelligenceact.eu/article/50/
Most organisations have noted this, few have asked what it means technically, and fewer have assessed whether their current approach actually satisfies it. The provision is directly applicable in all EU Member States and enters into application on August 2, 2026.Prokopiev Law Group, “EU AI Act Article 50 Imposes Transparency Obligations for AI-Generated Content, August 2026,”https://www.prokopievlaw.com/post/eu-ai-act-article-50-imposes-transparency-obligations-for-ai-generated-content-european-union-augu/ That deadline is close enough to matter for compliance teams planning implementation now.
What "Machine-Readable" Actually Means
The most common assumption is that compliance means a watermark, a disclosure banner, or a detection score. None of these satisfy the requirement as written.
A watermark is a visual signal for human readers. A disclosure banner ("this content was produced with AI assistance") is a human-readable statement. An AI detection score is an inference about finished output, not a record of how it was produced.
Machine-readable means structured data that a system can parse, query, and audit without human interpretation.
Open standards like RDF, JSON-LD, or specific HTML tags are the appropriate formats, as they ensure compatibility with exisiting verification tools and moderation systems. To meet the requirement, systems must record who generated the content, when, and with what parameters, and logs must be accessible and auditable.Itequia, “What Article 50 of the AI Act Requires,” Itequia. https://itequia.com/transparency-in-the-age-of-ai-what-article-50-of-the-ai-act-requires/
In the context of AI-assisted writing in documentation, that means a structured record that answers at minimum:
- Which parts of this document were AI-generated?
- What AI system produced them?
- Was the AI output reviewed by a human before use?
- What did the human do with the AI output (i.e. accept it, modify it, or discard it)?
A detection score cannot answer any of these questions. It does not record what happened during production.
The Compliance Gap
Consider a compliance officer reviewing an AI-assisted regulatory submission.
The document was drafted using an AI writing tool. The writer reviewed each AI suggestion and substantially rewrote several sections. The final document reads as human-written (because it largely is). An "AI detector" scores the document as "likely human."
That score tells the compliance officer nothing about what Article 50 requires them to document.
The gap is provenance not detection. Article 50 does not ask you to prove AI was not involved, it asks you to label where and how it was. The European Commission's own consultation confirms that compliant techniques may include watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity of content, and logging methods or a combination of these. “AI Act - Article 50 consultation on transparency.” https://www.ddg.fr/actualite/article-50-of-the-ai-act-the-european-consultation-on-transparency-requirements-september-2025
A post-hoc detection score, is not definaely sufficient.
Defining Compliance
A compliant AI transparency record is a structured log of the production process. At minimum, it records:
1) Each AI interaction during drafting: when it occurred, what type of assistance was requested, which AI system was used
2) The outcome of each interaction: whether the AI output was accepted unchanged, partially used, substantially modified, or rejected
3) The character spans affected in the document
This record is machine-readable (structured JSON), auditable via hash chain, and ideally author-controlled. The draft Code of Practice developed by the AI Office explicitly prohibits providers from relying on a single marking technique, Fiona Ghosh and Patricia Wade, “Transparency of AI-generated content: the EU’s first draft Code of Practice,” Ashurst. Accessed: Apr. 07, 2026. https://www.ashurst.com/en/insights/transparency-of-ai-generated-content-the-eu-first-draft-code-of-practice/
which further reinforces that provenance logging sits alongside, not instead of, other measures.
Who This Applies To
Businesses and public bodies that operate AI systems or publish AI-generated content, bear direct compliance obligations from the August 2, 2026 application date and must implement disclosure mechanisms in product interfaces, content pipelines, and terms of service.Future of Life Institute, “Article 50 Transparency Obligations for Providers and Deployers of Certain AI Systems EU Artificial Intelligence Act.” https://artificialintelligenceact.eu/article/50/
High-risk documentation contexts include:
- Clinical trial documentation and medical writing
- Regulatory submissions and compliance reports
- Legal documents and contract drafting
- Financial disclosures and prospectuses
- Employment decisions and HR documentation
- Procurement and supplier assessment reports
Any organisation using AI writing tools in these contexts has an Article 50 exposure they may not have fully assessed.
Non-compliance is subject to administrative fines of up to €15 million or up to 3% of the operator's total worldwide annual turnover for the preceding financial year, whichever is higher.“Limited-Risk AI—A Deep Dive Into Article 50 of the European Union’s AI Act,” WilmerHale. https://www.wilmerhale.com/en/insights/blogs/wilmerhale-privacy-and-cybersecurity-law/20240528-limited-risk-ai-a-deep-dive-into-article-50-of-the-european-unions-ai-act
GDPR as a Pattern
Why Article 50 Is the Beginning
In 2018 when GDPR was introduced it was widely treated as a European compliance problem. Within a few years, Brazil had passed the LGPD, California had enacted CCPA, India was developing its own data protection framework, and the GDPR had become the de facto template for data protection law.
Companies that built GDPR-compliant infrastructure early found themselves already positioned when other jurisdictions followed. The same pattern is already visible with AI transparency obligations, and it is moving faster than GDPR did.
China has required all network information service providers to label AI-generated content with both explicit and implicit markers since September 1, 2025, under its Measures for Identifying Artificial Intelligence-Generated Synthetic Content. Explicit identifiers include visible watermarks, while implicit identifiers are technical in nature.A. Hilliard, A. Gulley, E. Kazim, and A. S. Koshiyama, “Artificial intelligence policy worldwide a comparative analysis,” Royal Society Open Science, vol. 13, no. 2, p. 242234, Feb. 2026, doi: 10.1098/rsos.242234.
Providers operating in China must implement both explicit and metadata labels for generative AI outputs, update service agreements, and maintain event logs. T. Lawrence, “AI Regulation Updates H2 2025,” FairNow. https://fairnow.ai/ai-regulations-updates-h2-2025/
That last requirement (i.e maintaining event logs) is structurally identical to what Article 50 demands in Europe.
South Korea's Basic AI Act (from January 2026), applies extraterritorially where systems affect Korean users, and introduces requirements for transparency, risk assessment, human oversight, and documentation.
Vietnam's Law on Digital Technology, also effective in 2026, includes AI labeling and transparency provisions. “Where AI Regulation is Heading in 2026: A Global Outlook | Blog | OneTrust.” https://www.onetrust.com/blog/where-ai-regulation-is-heading-in-2026-a-global-outlook/
Canada's Artificial Intelligence and Data Act focuses on high-impact AI systems and includes obligations for transparency, human oversight, safety, and accountability, aligned with international AI governance norms. Anecdotes A.I Ltd, “AI Regulations in 2025: US, EU, UK, Japan, China & More.” https://www.anecdotes.ai/learn/ai-regulations-in-2025-us-eu-uk-japan-china-and-more
The practical implication for organisations investing in Article 50 compliance infrastructure now is the same as it was for GDPR early adopters.
A provenance logging system built to satisfy the EU requirement is already structurally compatible with what China, South Korea, Vietnam, and Canada are independently converging on. The specific technical standards differ, but the underlying requirement (i.e that you can demonstrate what your AI system did and when) is consistent across jurisdictions.
A detection score is jurisdiction-specific and probabilistic. A structured event log is portable, deterministic, and already points in the direction regulators worldwide are heading.
TWFF as a Technical Implementation
TWFF Tracked Writing File Format, https://github.com/Functional-Intelligence-Research-Lab/twff is an open standard, designed to produce exactly this kind of machine-readable provenance record.
It records AI writing interactions as a deterministic event log attached to the document. The author controls the record and discloses it voluntarily alongside the document.
The W3C Data Privacy Vocabularies and Controls Community Group (DPVCG) develops and maintains the Data Privacy Vocabulary (DPV), which enables the creation of machine-readable, interoperable, and standards-based representations for describing the processing of personal data and technologies, and already includes an EU AI Act extension providing concepts relevant to AI system documentation and compliance. H. J. Pandit, B. Esteves, G. P. Krog, P. Ryan, D. Golpayegani, and J. Flake, “Data Privacy Vocabulary (DPV) – Version 2.0,” in The Semantic Web – ISWC 2024, G. Demartini, K. Hose, M. Acosta, M. Palmonari, G. Cheng, H. Skaf-Molli, N. Ferranti, D. Hernández, and A. Hogan, Eds., Cham: Springer Nature Switzerland, 2025, pp. 171–193. doi: 10.1007/978-3-031-77847-6_10.
TWFF is being developed as a compatible vocabulary profile within this ecosystem.
Caveat
TWFF is in active development. The current reference implementation covers the core logging protocol. Integrations with enterprise writing environments including Microsoft Word, Google Docs, and specialist documentation tools are on the roadmap for 2026 to 2027.
Early adopter engagements are structured as research partnerships, not turnkey deployments.
We (Functional Intelligence Research lab) are forming an early adopter programme for organisations that want to evaluate TWFF as an Article 50 compliance mechanism in their documentation workflows.
Early adopters receive direct access to the implementation team, input into the specification, and a named position in the published research.
If your organisation is assessing Article 50 compliance and wants to understand what a process-log based approach looks like in practice, get in touch.