AI disclosure rules: provider or deployer?
Put Claude in your own product and the AI Act may reclassify you from deployer to provider. The three triggers in Article 25, and what Article 50 asks you to mark.
TL;DR
Using Claude internally makes you a deployer. Shipping Claude inside a product you sell can make you the provider, and the provider carries almost every obligation in the Act. Three things flip the switch: your trademark on the system, a substantial modification, or a change of intended purpose that makes the system high-risk.
Provider or deployer: the definition that decides everything
The AI Act allocates duties by role, not by company size or by who wrote the model. Two definitions in Article 3 do the allocating, and the difference between them is the single most expensive fact in the Act.
A provider is a person or body that develops an AI system or a general-purpose AI model, or has one developed, and places it on the market or puts it into service under its own name or trademark, whether for payment or free of charge. A deployer is a person or body using an AI system under its authority, except where the use is a personal non-professional activity.
Read those twice. “Under its own name or trademark” is the operative phrase, and “free of charge” closes the obvious loophole. A free beta of your AI feature is placed on the market just as firmly as a paid one.
There is a third role that catches most companies building on Claude. Article 3(68) defines a downstream provider as a provider of an AI system, including a general-purpose AI system, which integrates an AI model, whether that model is vertically integrated or provided by another entity based on contractual relations. If you wrap Claude in your own interface and sell it, that is you.
| What you are doing | Your role | Who the provider is |
|---|---|---|
| Staff use Claude in the browser or in Claude Code | Deployer | Anthropic |
| You build an internal tool on the API for your own employees | Deployer | Anthropic |
| You ship a customer-facing feature powered by Claude under your brand | Downstream provider of that system | You, for the system. Anthropic remains provider of the model |
| You resell Claude unchanged under your own name | Provider | You |
The middle two rows are where companies get it wrong. Our AI Act deployer guide covers the duties that attach when you stay on the deployer side of that line.
The three triggers that turn you into a provider
Article 25(1) is the reclassification rule, and it applies to high-risk AI systems. A distributor, importer, deployer or other third party becomes the provider, with the full set of Article 16 obligations, in three cases:
- Branding. You put your name or trademark on a high-risk AI system that is already on the market or in service. Contractual arrangements may still allocate obligations differently between the parties.
- Substantial modification. You substantially modify a high-risk AI system that is already on the market or in service, in a way that leaves it high-risk under Article 6. Article 3(23) defines a substantial modification as a change not foreseen in the provider’s initial conformity assessment that affects compliance with Chapter III Section 2, or that modifies the intended purpose.
- Change of intended purpose. You change the intended purpose of an AI system that was not high-risk, including a general-purpose AI system, so that it becomes high-risk under Article 6.
Trigger three is the one that matters for Claude. Claude is a general-purpose AI system, and nothing about it is inherently high-risk. Point it at CV screening, credit decisions, or access to public benefits and you have given it a new intended purpose that lands in Annex III. At that moment you are the provider, and the obligation set changes completely.
When Article 25(1) applies, the initial provider stops being the provider of that specific system. It still has to cooperate closely: supply technical documentation sufficient to assess compliance with Article 16, inform you of known limitations and failure modes, and give targeted technical access including for testing and validation. That cooperation duty does not apply if the initial provider clearly stated that the system must not be changed into a high-risk system.
One more clause worth reading before you sign anything. Article 25(4) requires providers of high-risk systems and the third parties supplying them with models, tools, services, components or processes to agree in writing on the information, capabilities, technical access and assistance needed for full compliance, based on the generally acknowledged state of the art. Free and open-source tools made publicly available are exempt, other than general-purpose AI models.
What changes when you become the provider
The gap between the two obligation sets is not a matter of degree.
| Obligation | Deployer, Article 26 | Provider, Article 16 |
|---|---|---|
| Use per instructions, human oversight by competent staff | Yes | Beyond scope |
| Input data relevance where you control it | Yes | Via Article 10 |
| Monitoring and serious-incident reporting | Yes, to the provider and authorities | Yes, plus corrective actions under Article 20 |
| Log retention | At least six months, longer or shorter where the purpose or other law requires | Keep logs under your control per Article 19 |
| Inform workers’ representatives and affected workers before use at work | Yes | Beyond scope |
| Quality management system | No | Yes, per Article 17 |
| Technical documentation | No | Yes, per Article 18 |
| Conformity assessment before placing on the market | No | Yes, per Article 43 |
| EU declaration of conformity and CE marking | No | Yes, per Articles 47 and 48 |
| Registration in the EU database | Public authorities only, per Article 49 | Yes, per Article 49(1) |
| Name, trade name and contact address on the system | No | Yes |
| Accessibility under Directives 2016/2102 and 2019/882 | No | Yes |
A deployer runs a system carefully. A provider runs a conformity programme. Quality management, technical documentation, conformity assessment, CE marking and registration are the four line items that turn a product decision into a programme with a budget. Our governance framework guide maps those onto the policy documents you probably already have.
What Article 50 asks of you, and what Claude already does
Article 50 applies whether or not the system is high-risk, which is why it catches far more products than the high-risk chapter does. The duties split by role, and that split is the detail most summaries get wrong.
Providers, Article 50(1). Design systems that interact directly with people so those people know they are dealing with an AI system. Not required where that is obvious to a reasonably well-informed person.
Providers, Article 50(2). Where the system generates synthetic audio, image, video or text, mark the outputs in a machine-readable format so they can be detected as artificially generated or manipulated. Solutions have to be effective, interoperable, robust and reliable as far as technically feasible. Assistive editing functions and systems that do not substantially alter the input data are out of scope.
Deployers, Article 50(4). Disclose deepfake image, audio or video content as artificially generated or manipulated. For evidently artistic, creative, satirical or fictional work, disclosure is limited to noting that such content exists, in a way that does not hamper enjoyment of the work. Text published to inform the public on matters of public interest also has to be disclosed as AI-generated, and this is where the useful exemption sits: the duty does not apply where the text has undergone human review or editorial control and a person or entity holds editorial responsibility for the publication.
Timing, Article 50(5). Clear and distinguishable information, meeting applicable accessibility requirements, at the latest at the time of the first interaction or exposure. Not in a settings page, not in the terms.
On the marking duty, the ground moved in 2026. Anthropic has signed the EU AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content, and Claude now marks its own output two ways:
| Marking method | What it covers | Where it applies |
|---|---|---|
| Embedded text watermark | Generated text, persisting through copying, pasting and some editing | Supported models on Claude Platform, Claude, Claude Code, Cowork and Claude Tag, and through AWS, Google Cloud and Microsoft Foundry |
| Signed C2PA Content Credentials | Provenance metadata on supported generated files, with PNG and JPEG given as examples | Only where the platform offers file generation, which is the Claude apps and Claude Platform including Claude Platform on AWS and Claude in Microsoft Foundry |
Models launched in the EU on or after 2 August 2026 support marking at launch, and earlier models are being added. Fable 5.1, Fable 5, Mythos 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5 and Haiku 5.5 are listed with the full set of marks.
Three caveats decide whether this actually discharges your obligation. Watermark detection is in private preview for eligible organizations rather than generally available, C2PA credentials can be checked with a free content checker, and a detected mark is explicitly not conclusive proof of provenance. Marks also degrade: heavily edited, paraphrased or translated text, very short passages, and file metadata stripped by format conversion, re-saving or screenshots all lose the signal. The absence of a mark does not mean content was not AI-generated.
Anthropic’s own position is the one to write into your risk register: businesses building on Claude should independently assess their own Article 50 obligations. Upstream marking is a building block, not a transfer of liability.
The disclosure duty Anthropic adds on top
Anthropic’s Usage Policy imposes a contractual disclosure duty that is narrower than Article 50 in scope and stricter in form. Consumer-facing chatbots and AI agents have to disclose to users that they are interacting with AI rather than a human, at least at the start of each chat session or in the product interface. You do not have to name Anthropic, Claude or the model.
For the eleven high-risk recommendation areas in the policy, which run from legal and medical through credit, insurance, housing, employment and access to public benefits, two further requirements apply. A qualified person has to meaningfully review the recommendation before it reaches the individual or is used to implement a decision, with the authority to change it and any licence the underlying activity requires. And the individual has to be clearly told that AI was used to produce it. Wholly favourable decisions are exempt where law permits; partial approvals, reduced amounts and conditional approvals are not.
Note what that means commercially. For a credit or insurance product built on Claude, the human-in-the-loop is a contract term, not only a compliance choice. A fully automated decision path breaches the Usage Policy before the AI Act is even reached. Our proof of concept guide covers how to find that kind of blocker in week one rather than month six.
What a breach costs
| Breach | Cap, whichever is higher | SMEs and start-ups |
|---|---|---|
| Prohibited practices under Article 5 | 35 million EUR or 7 percent of worldwide annual turnover | Lower of the two |
| Provider, deployer, importer, distributor and notified body obligations, including Article 50 transparency | 15 million EUR or 3 percent of worldwide annual turnover | Lower of the two |
| Incorrect, incomplete or misleading information to notified bodies or authorities | 7.5 million EUR or 1 percent of worldwide annual turnover | Lower of the two |
The transparency duties sit in the 15 million EUR band. An undisclosed chatbot is not a paperwork offence.
The transparency rules took effect in August 2026, alongside the Act becoming broadly applicable on 2 August 2026. High-risk obligations under Annex III follow on 2 December 2027 and Annex I on 2 August 2028, both as amended by the AI Omnibus that entered into force on 27 July 2026.
Run it yourself or outsource?
| Task | Keep in-house | Sensible to outsource |
|---|---|---|
| Role classification per product | Yes, with legal. It depends on facts only you hold | A first-pass review and the decision template |
| Intended-purpose documentation | Yes. It is a product artefact, and Article 3(12) ties it to your own sales material | No |
| Article 50 disclosure design | Yes. It is interface work | A pattern review against the timing rule |
| Conformity assessment, if you become a provider | Partly. Scope and evidence are yours | The gap analysis and the technical file structure |
| Contract terms with upstream suppliers under Article 25(4) | Yes, with legal | Model clauses as a starting point |
One warning on this split. The classification question is tempting to outsource and the worst candidate for it, because Article 3(12) defines intended purpose by reference to your own instructions for use, promotional material, sales statements and technical documentation. Your marketing copy is evidence. An external adviser cannot know what your sales team has been promising.
FAQ
Does using Claude make us a provider under the AI Act?
Not by itself. Internal use makes you a deployer, with Anthropic as the provider. You become the provider of a system when you place it on the market or put it into service under your own name or trademark, or when one of the three Article 25(1) triggers applies: your branding on a high-risk system, a substantial modification, or a change of intended purpose that makes the system high-risk.
Provider vs deployer: what does the difference cost?
A deployer’s duties are operational: use per instructions, competent human oversight, monitoring, incident reporting, at least six months of log retention, and informing workers’ representatives before use at work. A provider’s duties are a programme: a quality management system, technical documentation, conformity assessment before market, an EU declaration of conformity, CE marking and registration in the EU database. Budget the first in person-days and the second in person-years.
Do we have to label AI-generated text if Claude already watermarks it?
Probably yes, and for a different reason. The Article 50(2) machine-readable marking duty falls on the provider of the generating system, which Anthropic now addresses through text watermarks and C2PA Content Credentials. The Article 50(4) disclosure duty for deepfakes and for text published on matters of public interest falls on the deployer, and that one is a visible disclosure to people rather than metadata for machines. Anthropic states directly that businesses building on Claude should assess their own Article 50 obligations.
Is AI-generated content for our company blog covered?
Only if it informs the public on matters of public interest, and even then the exemption usually applies. Where the text has undergone human review or editorial control and a named person or entity holds editorial responsibility for publication, the Article 50(4) disclosure duty does not apply. Product marketing is not a matter of public interest in the first place. An unreviewed, automatically published news feed is the case the rule was written for.
Where must the AI disclosure appear in our product?
Clear, distinguishable, meeting applicable accessibility requirements, and at the latest at the time of the first interaction or exposure. Anthropic’s Usage Policy is more specific for consumer-facing chatbots and agents: at least at the start of each chat session, or in the product interface. A line in the terms of service satisfies neither.
What happens to Anthropic’s obligations when we become the provider?
Anthropic stops being the provider of that specific system and keeps its obligations as the provider of the general-purpose AI model. Under Article 25(2) the initial provider has to cooperate: technical documentation sufficient to assess Article 16 compliance, disclosure of known limitations and failure modes, and targeted technical access. That cooperation duty falls away if the initial provider clearly stated the system must not be turned into a high-risk one, so read the terms before you build a high-risk use case.
Sources
- European Commission AI Act Service Desk: Article 3, Definitions (28 October 2026)
- European Commission AI Act Service Desk: Article 25, Responsibilities along the AI value chain (28 October 2026)
- European Commission AI Act Service Desk: Article 50, Transparency obligations (28 October 2026)
- European Commission AI Act Service Desk: Article 16, Obligations of providers of high-risk AI systems (28 October 2026)
- European Commission AI Act Service Desk: Article 99, Penalties (28 October 2026)
- Anthropic: How Claude marks AI-generated content (28 October 2026)
- Anthropic: Usage Policy (28 October 2026)
- European Commission: Regulatory framework for AI (28 October 2026)