EU AI Act Article 50: AI Labelling Rules Explained
The EU AI Act's Article 50 transparency rules became binding on 2 August 2026, requiring AI content marking and deepfake labelling, with fines up to €15m.
The transparency obligations in Article 50 of the EU AI Act became applicable on 2 August 2026 [1]. They are the first provisions of the Act to reach directly into everyday generative-AI use — chatbots, AI-written articles, synthetic images and deepfakes — rather than applying only to high-risk systems or the largest model developers. Ahead of the deadline, the European Commission published a voluntary Code of Practice on Transparency of AI-generated Content on 10 June 2026, which around 190 organisations had signed by late July 2026 [1].
Non-compliance with Article 50 falls under Article 99(4) of the Act, carrying administrative fines of up to €15,000,000 or 3% of total worldwide annual turnover for the preceding financial year, whichever is higher, with a lower cap applied to SMEs and startups [3].
Key facts at a glance
| Detail | Position |
|---|---|
| Obligations applicable from | 2 August 2026 |
| Code of Practice published | 10 June 2026 (voluntary) |
| Signatories by late July 2026 | Approximately 190 organisations |
| Who is bound | Providers (developers) and deployers (users) of AI systems |
| Maximum fine | €15,000,000 or 3% of worldwide annual turnover, whichever is higher |
| Legal basis for penalties | Article 99(4), EU AI Act |
What Article 50 actually requires
The obligations split between providers — those developing and supplying AI systems — and deployers, meaning organisations using them. Article 50 imposes four core duties [1][2].
1. Tell people they are talking to an AI
Providers must ensure people know they are interacting with an AI system, "unless this is obvious from the point of view of a natural person who is reasonably well-informed" [2]. In practice this makes an unlabelled customer-service chatbot a compliance risk, though the "reasonably well-informed" carve-out gives some latitude where the context makes it plain.
2. Mark synthetic output in machine-readable form
Providers must mark synthetic audio, image, video and text outputs "in a machine-readable format and detectable as artificially generated or manipulated" [2]. The Act requires solutions to be "effective, interoperable, robust and reliable as far as this is technically feasible" [2] — a standard that acknowledges the technology is imperfect while still requiring a genuine attempt. This is the obligation driving watermarking and content-credential work across the industry.
3. Label deepfakes
Deployers must disclose image, audio or video content that has been artificially generated or manipulated to resemble real people, objects or events [2]. Artistic, creative, satirical or fictional works are treated more leniently: they need only disclose the existence of manipulation "in an appropriate manner" rather than carrying a prominent warning [2].
4. Label AI-written text on matters of public interest
Deployers publishing AI-generated or manipulated text "with the purpose of informing the public on matters of public interest" must disclose it — unless the content has undergone human review and a person or organisation holds editorial responsibility for it [2].
That exemption is the single most important detail in Article 50 for publishers, and it is widely misread. It does not exempt AI-assisted content generally. It exempts content where a human has genuinely reviewed the material and accepts editorial responsibility for what it says.
All disclosures must be made "at the latest at the time of the first interaction or exposure," in a clear and distinguishable format [2]. Exemptions exist for assistive editing functions that do not substantially alter input data, and for certain law-enforcement uses [2].
Why this matters
Most AI regulation so far has landed on model developers. Article 50 lands on ordinary businesses. A UK agency running an AI chatbot on a client's site serving EU customers, a marketing team publishing AI-assisted articles about public-interest topics, a brand producing synthetic video — all now sit inside scope, regardless of where the business itself is established, because the Act applies based on where the output is used.
The connection to product releases is direct and already visible. Anthropic published a technical explanation of Claude's text watermark on 14 August 2026, and stated that models launched before 2 August 2026 would receive watermarking over the following months — a cutoff that maps precisely onto this deadline. Article 50 is now shaping what frontier labs ship.
Who should care
Any business serving EU users with a customer-facing chatbot needs to confirm it discloses its nature at first interaction. Publishers and marketing teams using AI in content production need a documented human review process, since that is what the public-interest exemption turns on. Agencies deploying AI on behalf of clients should establish contractually who counts as the deployer and therefore carries the labelling duty. UK and US businesses should not assume they are outside scope — the test is where the output is encountered, not where the company sits.
Practical implications for buyers and users
Audit your customer-facing AI first: chatbot disclosure is the cheapest obligation to meet and the most visible to regulators and users alike. Document your editorial review process for AI-assisted content, because the exemption depends on being able to demonstrate genuine human responsibility rather than merely asserting it. When selecting AI vendors, ask directly whether their outputs carry machine-readable marking, since the provider-side obligation only helps you if the tools you use actually implement it. The Commission has published an optional icon set for labelling AI-generated content [1], which is worth adopting for consistency. Signing the Code of Practice is voluntary but allows an organisation to demonstrate compliance through its measures rather than building an individual assessment from scratch [1].
Limitations, availability and unresolved questions
The Code of Practice is voluntary, so signing it is not a legal safe harbour in itself — the Article 50 obligations bind regardless. Enforcement practice is untested: the obligations only became applicable on 2 August 2026, and there is no body of decisions yet indicating how strictly national authorities will interpret terms like "reasonably well-informed" or "matters of public interest." The technical standard of "as far as this is technically feasible" is inherently movable and will likely tighten as watermarking improves. This article is general information about a regulation, not legal advice, and organisations should take qualified advice on their specific circumstances.
Frequently asked questions
When did EU AI Act Article 50 come into force?
The Article 50 transparency obligations became applicable on 2 August 2026 [1].
What are the fines for breaching Article 50?
Up to €15,000,000 or 3% of total worldwide annual turnover for the preceding financial year, whichever is higher, under Article 99(4). SMEs and startups face the lower of the two figures [3].
Does Article 50 apply to UK or US companies?
The Act applies based on where AI output is used rather than where the company is established, so businesses outside the EU serving EU users can fall within scope. Take qualified legal advice on your specific position.
Do I have to label every AI-assisted article?
No. The obligation covers text published to inform the public on matters of public interest, and exempts content that has had human review with a person or organisation taking editorial responsibility [2].
Is the Code of Practice mandatory?
No. The Code published on 10 June 2026 is voluntary, but signing it allows organisations to demonstrate compliance using its measures [1]. The underlying Article 50 obligations are binding either way.
Verdict
Article 50 is the point at which AI regulation stops being a concern for model developers and becomes an operational requirement for ordinary businesses. The obligations are, on the whole, reasonable and cheap to meet: disclose your chatbot, label synthetic media, document who takes editorial responsibility for AI-assisted content. The uncertainty is in interpretation rather than intent, and with no enforcement precedent yet, the sensible posture is to comply visibly rather than argue the edges. For most organisations this is a short compliance project, not a strategic problem — but it is no longer optional.
The current AI Marketing & Automation shortlist
Where this sits in the wider market: our current shortlist for AI Marketing & Automation, what each tool is best at and the main caution to check before committing.
| Tool | Best for | Current position | Important caution |
|---|---|---|---|
| HubSpot Connected growth suite | CRM-centred marketing, AEO and lifecycle operations | HubSpot now connects marketing automation, customer context, AI agents and dedicated answer-engine visibility tooling. | Value depends on data quality and disciplined CRM use, not merely enabling AI features. |
| Jasper Brand content | Governed campaign content across teams | Jasper remains focused on marketing teams that need brand context, repeatable workflows and approvals. | Plans, limits and model availability change frequently; confirm the current vendor page before purchasing. |
| Writesonic AI-search workflow | Content production plus search and AI visibility | Writesonic is relevant to teams combining content operations with monitoring for newer answer-engine channels. | Visibility scores are directional; connect them to qualified traffic and revenue. |
| Semrush Search intelligence | SEO research, competitive visibility and content planning | Semrush remains a broad search and competitive-intelligence platform as teams add AI visibility to established SEO work. | Plans, limits and model availability change frequently; confirm the current vendor page before purchasing. |
| Surfer On-page workflow | Search-aware briefs and page optimisation | Surfer fits teams that want structured on-page guidance inside a repeatable content process. | Optimisation scores do not replace original evidence, expertise or good writing. |
| Copy.ai GTM automation | Repeatable sales and marketing workflows | Copy.ai is aimed at automating go-to-market processes rather than simply generating isolated pieces of copy. | Plans, limits and model availability change frequently; confirm the current vendor page before purchasing. |
| Klaviyo Lifecycle commerce | Ecommerce email, messaging and customer segmentation | Klaviyo combines commerce data, lifecycle automation and assisted campaign work. | Revenue attribution and deliverability need independent monitoring. |
| Canva Campaign creative | Fast delivery of on-brand marketing assets | Canva gives non-design teams a practical layer for adapting AI-assisted creative into channel-ready formats. | Plans, limits and model availability change frequently; confirm the current vendor page before purchasing. |
| AdCreative.ai Paid creative | Rapid ad variations and testing inputs | AdCreative.ai focuses on producing and iterating paid-media creative rather than managing the whole marketing stack. | Measure incrementality and creative fatigue instead of trusting predicted scores alone. |
| Buffer Social operations | Small-team scheduling and social workflow | Buffer remains a straightforward social publishing layer for teams that value simplicity. | Plans, limits and model availability change frequently; confirm the current vendor page before purchasing. |
| n8n Flexible automation | Technical teams building owned AI workflows and agents | n8n combines workflow automation with reusable agents, tools, memory and self-hosting options. | Flexible automation also creates operational responsibility for credentials, logs and failures. |
| Zapier Accessible automation | Connecting common SaaS tools without heavy engineering | Zapier remains the approachable choice when speed of integration matters more than deep custom control. | Costs can rise with task volume and complex multi-step automations. |
Related reading
Sources and verification notes
Primary product documentation checked for this update: