AI Transparency Check: Implementing Article 50 EU AI Act in Practice
Since 2 August 2026, the transparency obligations of the EU AI Act apply. We systematically check which of your AI-supported content and assistants are affected – and implement the necessary measures in website, shop and Odoo.
Book initial consultation Start self-assessment
From bloopark – official Odoo partner focused on the DACH SME sector. Technical and procedural implementation, no legal advice.
Voluntary label: This image was generated with AI. We apply what we implement on this page.
First: Who bears which obligation?
The AI Act distinguishes between two roles. Almost every misjudgment in practice begins with confusing these two.
Provider – in the English text "provider"
Who develops or has developed an AI system and places it on the market or puts it into operation under their own name or brand (Article 3 number 3). For most companies, these are the manufacturers of the tools used – not you. Caution: an assistant operating under your own brand can place you in the provider role.
Operator – in the English text "deployer"
Who uses an AI system professionally under their own responsibility (Article 3 number 4). That is you, as soon as your marketing, your editorial team, your shop or your service uses AI. The English term is “deployer”, not “user” – the obligations for published content depend on this role.
What applies from 2 August 2026
Article 50 of the EU AI Act regulates transparency, not prohibitions. Six points are relevant for companies that use AI in marketing, websites, shops, or customer communication.
Article 50 Paragraph 1 · Provider
Recognisable AI interaction
Systems that interact directly with people must make it clear that they are AI. The obligation is directed at the providers of the system. In the case of an assistant operating under your own brand, this role may also apply to your company.
Article 50 Paragraph 2 · Provider
Machine-readable marking
Providers of generative AI systems must mark their outputs in a machine-readable way. For systems that were placed on the market before 2 August 2026, Article 111 paragraph 4 sets a deadline until the 2 December 2026. As an operator, you do not bear this obligation – but you do have a practical interest in not losing existing provenance data in your systems.
Article 50 Paragraph 3 · Operator
Emotion recognition and biometric categorisation
Anyone using a system for emotion recognition or biometric categorisation must inform the affected individuals about the operation of the system and process personal data in accordance with GDPR. This is practically relevant as soon as a voice agent or service tool evaluates mood or tonality.
Article 50 Paragraph 4 · Operator
Image, audio and video
Anyone publishing AI-generated or AI-edited image, audio, or video content that could appear real must disclose this. This obligation falls on the operators, namely your company or agency. An intention to deceive is not required for this.
Article 50 Paragraph 4 · Operator
AI texts: narrower scope than often assumed
The obligation to disclose AI-generated texts applies to publications that inform the public about matters of public interest. It does not apply when there is substantive human review with clear editorial responsibility. Normal product texts and newsletters generally do not fall under this.
Article 50 Paragraph 5 · Form
Clear, distinguishable and accessible
The information must be clear and distinguishable at the latest by the time of the first interaction or exposure – and meet the applicable accessibility requirements. The article does not specify any regulatory framework; for companies within the scope of the Accessibility Strengthening Act, these requirements arise from the BFSG. This is the point that is almost always forgotten in implementation.
Even if Article 50 does not concern you: AI Competence
Article 4 has been in effect since 2 February 2025 and is directed at providers and operators. It requires measures that promote the development of AI competence of their own staff and those acting on their behalf – scaled according to prior knowledge and context of use.
Important for classification: According to the version in effect since 27 July 2026, Article 4 explicitly requires no guaranteed competence level of individual persons, and it is not listed in the penalty catalogue of Article 99. Therefore, there is no training obligation with a certification requirement – but an obligation that, unlike Article 50, practically reaches every operator, including those who, after the check, find: there is nothing to label in our content.
In implementation, these are components that we will deliver anyway: an internal guideline for the use of AI in client content and a documented briefing for marketing and editorial.
Who is subject to which obligation?
- Article 50: only for affected content or systems
- Article 4: for every professional use of AI
- Both addressable with guidelines and briefing
What this means in practice
Not every AI-generated asset requires the same labelling. The following classification follows the wording of Article 50 and the guidelines of the EU Commission from July 2026. The final legal assessment in individual cases remains reserved for your legal advice.
Photorealistic AI images in marketing and website
A photorealistic image of a person, product, or place may fall under the disclosure obligation, even if nothing real is depicted. The argument "purely synthetic, therefore not affected" is not tenable according to the Commission's interpretation.
Product photo with AI-generated background
A real product in an AI-generated environment is generally not subject to disclosure according to the Commission's examples, as long as the product itself is not altered or misleadingly presented. If the product itself is enhanced, the assessment changes.
Abstract and recognisably unrealistic motifs
Illustrations, fantasy scenes, and motifs that are recognisably not real are generally outside the definition according to the guidelines. Also, common image editing such as colour correction or cropping does not fall under this.
Product texts, category texts, newsletters
For purely commercial content, the text obligation usually does not apply. It may look different when statements are made regarding health, safety, environment and sustainability, or finance. A documented editorial approval process is the most effective lever here.
Chatbots, assistants and voice agents
Customers should recognise at the latest during the first contact that they are speaking with an AI system. A notice that is only in the imprint or in the terms and conditions is usually not sufficient for this.
Advertising and campaign motives
The exemption for recognisably artistic, satirical or fictional works is to be interpreted narrowly according to the guidelines and usually does not apply to predominantly commercial content.
Self-assessment in six questions
The same logic we use in the review – condensed into six questions. The classification on the right is a first professional tendency, not a legal assessment of your individual case.
Action required likely
1. Do you publish photorealistic images, videos or audio content that are wholly or partially generated by AI?
Then you are in the core area of Article 50 Paragraph 4. The question is not whether, but which assets and how they are labelled.
Probably not affected
3. Do you exclusively use abstract or recognisably unrealistic AI motifs?
Such motives are generally outside the definition according to the guidelines. A brief documentation of this classification remains sensible.
Action required likely
5. Does any of your systems evaluate emotions or categorise individuals based on biometric data?
Then Article 50 paragraph 3 applies as an operator obligation, along with the data protection requirements. Often overlooked in voice and service tools.
Probably not affected
2. Do you only change the background or environment in real product photos without altering the product itself?
According to the Commission's examples, generally uncritical. The crucial factor is product fidelity, not the synthetic origin of the background.
Action required likely
4. Is there a chatbot, assistant or voice agent running on your website, in the shop or in telephony?
Then it is about recognisability at first contact – and the question of whether you slip into the provider role through your own branding.
Discretionary matter
6. Are AI texts being created with statements about health, safety, sustainability or finance – without documented editorial approval?
Here the subject decides, not the channel. A named approval step transforms uncertainty into a clear exception.
Answered more than one question with yes? Then the review framework is worthwhile – and a 45-minute conversation.
What you should check now
These ten points form the review framework that we work with in the AI transparency check. It is deliberately practical and ultimately provides a list that you can prioritise and work through.
Request review framework as PDF
Ten points on two pages, without sales text.
AI inventory
What tools today generate or change customer-facing content – in marketing, editorial, shop, support and sales?
Image stock
Which assets are AI-generated or AI-edited, how photorealistic are they, and is this documented anywhere?
Website and shop
Where do these contents specifically appear: product images, category images, backgrounds, banners, landing pages, blog, social preview images?
Assistants and automation
Is there a recognisable indication at first contact – in the chat widget, in the voice agent, in automated email responses?
Editorial approval process
Who reviews AI texts in terms of content, with what expertise, and who bears the documented editorial responsibility?
Labelling logic
How and where is it labelled? A maintainable system is crucial instead of manual individual cases that get lost in the next release.
Accessibility of the notice
Is the transparency notice also perceivable for screen readers and when using the keyboard? Article 50 paragraph 5 requires compliance with applicable accessibility requirements.
Metadata and origin
Do the embedded source information from the provider remain during upload, image conversion, and in the CDN, or are they removed?
Roles, policy and traceability
Is there an internal policy, and can it be traced for each asset which tool it was created with, when, and who approved it?
Agencies and suppliers
Is it regulated what agencies, photographers, and content service providers must disclose about AI contributions in delivered content?
How realistic is the risk?
The circulating millions are upper limits and the least likely scenario for SMEs. A more useful approach is a sober classification into three levels.
The legal framework
Article 99 paragraph 4 states that for violations of Article 50, fines of up to 15 million euros or 3 percent of global annual turnover apply – whichever amount is higher. This is a maximum limit, not a standard penalty, and the national implementation is up to the member states.
What actually applies to SMEs
For small and medium-sized enterprises including start-ups, Article 99 paragraph 6 reverses the rule: the lower of the two values applies. Anyone who unfiltered applies the “15 million or 3 percent” to SMEs is calculating with the wrong figure.
The more likely scenario
In the German market, pressure typically comes first from the side: a competition law warning from competitors or associations, triggered by a publicly visible website. This is quicker, cheaper to initiate, and targets exactly the content in question. Whether a specific case is suitable is a legal question – reducing vulnerability is a technical one.
How bloopark supports
We work on the technical and procedural side: first clarity about the status, then a viable marking, then a process that also holds for the next release and the next campaign.
Review and inventory
We record which AI tools are in use and which customer-side content they generate – across website, shop, Odoo, newsletter, and assistants.
Technical classification and gap analysis
We clearly separate what is affected by Article 50, what is not, and what has discretion – as a prioritized basis for your legal assessment, not as a substitute.
Implementation in Odoo and website
Marking on images and content, transparency notice in the chatbot or voice agent, adjustment of snippets, product data, and shop templates – maintainable instead of hand-crafted, and perceivable in an accessible manner.
Process and evidence
Approval and documentation steps in Odoo, clear roles, an internal policy, and an evidential situation that also holds up in an audit or supplier questionnaire.
This is what the result looks like
You will receive an audit report – no certificate and no seal. This is intentional.
- Checked scope, audit date, and applied criteria are named in the document
- For each finding: affected, not affected, or discretion – with justification
- Prioritised gap list as a basis for the legal assessment
- Contains the note: this is not a certification by an accredited body
The AI transparency check
Three levels, clearly defined. The entry point is always a free initial consultation, in which we jointly determine which level suits your AI application.
Basic check
890 € plus VAT.
Clarity about the inventory and a prioritised action list – without implementation.
Scope: one domain (website including shop) and one Odoo system. Additional brands, shops or systems on request.
- Review of website, shop and customer-side AI touchpoints
- AI inventory of the tools and content types used
- Technical classification according to Article 50 with prioritised gap list as a basis for the legal assessment
- Audit report as a basis for decision-making
- Results discussion, 60 minutes
Check and implementation recommended
2,490 € plus VAT.
The standard case: check and implement the measures directly in the website, shop and Odoo.
Scope as basic check, plus the implementation mentioned below.
- Everything from the basic check
- Implementation of the labelling in up to 5 page and snippet types (e.g. product image, category image, banner, blog, landing page)
- Transparency note for an assistant (chatbot, assistant or voice agent)
- Maintainable tagging logic instead of manual individual cases
- Internal AI content policy as a template
- Briefing for marketing and editorial, 60 minutes
Governance and operations
On request
For companies that use AI across multiple systems and brands.
Scope and effort depend on system landscape and number of brands.
- Review of website, shop, PIM, CRM and assistants
- Approval and documentation process in Odoo
- Roles, responsibilities and documentation
- Guidelines for agencies and suppliers
- Ongoing review with new tools and releases
Ongoing review and documentation
New tools, new campaigns, new releases: tagging rarely fails on the first day, but rather on the fourth update. Upon request, we will regularly check and keep the documentation up to date – as monthly support, regardless of the chosen level.
All prices net plus statutory VAT.
This is how the check proceeds
Initial conversation, free of charge
45 minutes in which we assess your AI usage and determine which level is sensible. No preparation required on your part.
Review
We will look at the website, shop, Odoo, assistants as well as image and text inventory and document the tools used.
Classification
You receive a clear distinction: affected, not affected, discretion – each with justification and priority.
Implementation
Labelling, notes, processes and documentation are integrated into your existing Odoo and website structure.
Operation
Upon request, we will regularly check when new tools, campaigns or releases are added.
Frequently Asked Questions
The questions that existing customers most frequently ask us about Article 50.
No. The disclosure obligation is not solely linked to the origin of an image, but to its effect: relevant are contents that could appear genuine. Abstract or recognisably unrealistic motifs generally do not fall under the guidelines of the EU Commission, while photorealistic representations can very well do so. It is precisely this distinction that we make in the review.
This obligation is directed at the providers of generative AI systems, not at you as the operator. For systems that were placed on the market before 2 August 2026, Article 111 paragraph 4 sets a deadline until 2 December 2026. Practically relevant for you is mainly that your own systems do not remove existing origin information unnoticed.
Article 50 paragraph 5 requires clear and distinguishable information at the latest by the time of the first interaction or exposure, and this information must comply with the applicable accessibility requirements. A notice that only appears in documents outside of the actual use generally does not meet this.
Article 4 addresses providers and operators of AI systems and has been in effect since 2 February 2025. It requires measures that promote the development of AI competence among staff, scaled according to prior knowledge and context of use. According to the version in effect since 27 July 2026, there is explicitly no requirement to guarantee a specific competence level for individual persons, and Article 4 is not listed in the penalty catalogue of Article 99. An internal policy and documented briefing are the pragmatic response here.
No. We work on the technical and procedural side: inventory, assessment framework, implementation, documentation. The result is a technical classification with a prioritised gap list as a basis for the legal assessment – not as a substitute for it. The legal assessment of your individual case remains the responsibility of your legal counsel.
The basic check usually runs within two weeks from approval. The implementation depends on the scope and release cycle of your website and will be coordinated in advance.
Article 99 paragraph 4 specifies fines of up to 15 million euros or 3 percent of global annual turnover for violations of Article 50, depending on which amount is higher. For small and medium-sized enterprises including start-ups, the lower value applies according to Article 99 paragraph 6. These are maximum limits and not standard penalties; the national implementation is up to the member states. In the German market, a competition law warning from competitors or associations is typically closer than a fine procedure.
Then the effort is low and the timing is good: Once AI content is distributed across multiple channels and systems, subsequent labelling becomes significantly more expensive than a one-time clean structure. And Article 4 on AI competence will reach you even if there is nothing to label in your content.
Legal bases for further reading
All statements on this page can be verified at the source. These are the versions we refer to.
- Regulation (EU) 2024/1689 (EU AI Act), consolidated version of 27 July 2026 – EUR-Lex
- Article 50 in the wording – AI Act Service Desk of the European Commission
- Guidelines from the EU Commission on transparency obligations (July 2026)
- FAQ from the EU Commission on Article 50 – including the deadline of 2 December 2026
- Regulation (EU) 2026/1744 (Digital Omnibus on AI) – amendments to Articles 4, 99 and 111
Where does your AI usage stand today?
45 minutes is enough to assess your starting position and decide whether there is any need for action at all. After that, you will know what to do – and what not to do.
Legal notice: The content of this page serves as professional guidance and does not constitute legal advice. It is based on the consolidated version of Regulation (EU) 2024/1689 of 27 July 2026, the amendments made by Regulation (EU) 2026/1744, the guidelines of the European Commission on Article 50 from July 2026, and the voluntary Code of Practice on Transparency of AI-generated Content. Questions of interpretation are partially open; the binding interpretation is the responsibility of the courts. Please have your individual case legally examined. Status: 2 September 2026.