1. The Three Regulatory Tiers: GPAI, Deployer, Provider
The EU AI Act does not treat generative AI as a monolith. It creates a layered compliance architecture with distinct obligations depending on where an organization sits in the AI value chain. Understanding your position in this architecture is the first step toward compliance.
Tier 1: GPAI Model Provider
Organizations that train and release general-purpose AI models — including foundation models, large language models, and multi-modal models. This includes companies like OpenAI (GPT-4), Google (Gemini), Mistral, and open-source publishers. Obligations: technical documentation (Annex XI), copyright compliance summary, transparency to downstream providers, systemic risk assessment if applicable.
Tier 2: Deployer
Organizations that integrate GPAI models into their own products, services, or workflows and make those systems available to end users or employees. A European marketing agency using the GPT-4 API to power its content platform is a deployer. Obligations: Article 50 disclosure, incident reporting, human oversight, post-market monitoring, high-risk compliance if the application falls under Annex III.
Tier 3: End User / Operator
Individuals or business units using deployed AI systems. While end users face few direct regulatory obligations, they must be informed about AI interaction (Article 50) and retain the right to human review for consequential decisions. Organizations deploying AI to their own employees are both deployer and operator simultaneously.
Critical Point: Most European organizations using generative AI are deployers, not GPAI providers. This means they inherit responsibilities that cannot be fully delegated to the model provider through contract. The deployer is responsible to end users and regulators for the AI-generated outputs their system produces.
2. ChatGPT-Class vs. Systemic Risk Models (1025 FLOPs Threshold)
The EU AI Act distinguishes between standard GPAI models and systemic risk GPAI models based on training compute. This distinction triggers a substantially different compliance regime and has major implications for organizations evaluating which foundation models to use.
| Obligation | Standard GPAI Model | Systemic Risk GPAI Model (>1025 FLOPs) |
|---|---|---|
| Technical documentation | Required (Annex XI) | Required + enhanced (Annex XII) |
| Copyright compliance summary | Required | Required (higher scrutiny) |
| Downstream provider transparency | Required | Required + EU AI Office notification |
| Adversarial testing (red-teaming) | Not required | Required before market release |
| Incident reporting | Not required | Required — report to EU AI Office |
| Cybersecurity measures | Best practice | Mandatory, documented |
| Energy consumption reporting | Not required | Required annually |
Implications for Deployers
Deployers using systemic-risk GPAI models cannot simply rely on the provider's compliance. They must contractually verify that the provider has completed adversarial testing, that incident reporting channels are in place, and that the provider's systemic risk classification status is clearly documented. TraceGov.ai automates this contractual verification and maintains the evidence trail required for regulatory audits.
Which models may qualify as systemic risk? While providers are not always transparent about training compute, models likely approaching or exceeding the threshold include GPT-4 and successors, Gemini Ultra, Claude 3 Opus-class models, and large open-source releases like LLaMA 3 (405B parameters). Open-source providers have specific obligations even when releasing weights freely — the compute threshold applies at training time, not deployment time.
3. Copyright Compliance for Training Data Under the TDMA Directive
One of the most consequential and least understood dimensions of generative AI compliance in Europe is training data copyright. The intersection of the EU AI Act with the Text and Data Mining (TDM) exception in the Copyright Directive (2019/790) creates a complex compliance landscape for GPAI providers — and significant due diligence obligations for deployers.
The TDM Exception Framework
Research organisations and cultural heritage institutions may mine any lawfully accessible content for non-commercial scientific research. Rights holders cannot opt out. This benefits academic model training but not commercial GPAI development.
Commercial TDM is permitted unless rights holders have explicitly reserved their rights 'in an appropriate manner, such as machine-readable means.' This is the legal basis for commercial training data crawling — but only for content where opt-out has not been asserted.
Rights holders can opt out via robots.txt directives, metadata flags (X-Robots-Tag), or contractual terms of service. If a GPAI provider crawled opted-out content, that training is potentially infringing regardless of the model's downstream popularity or utility.
GPAI providers must publish a 'sufficiently detailed summary of the content used for training' to allow rights holders to check whether their opted-out content was used. The EU AI Office's template for this summary was published in early 2025 as part of the GPAI Code of Practice.
Deployer Due Diligence: As a deployer, you cannot fully transfer copyright liability to the GPAI provider. If you commercially deploy a model trained on infringing data, you may face secondary infringement claims. Request training data provenance documentation from providers and include indemnification clauses for copyright claims in your API agreements.
4. Transparency Obligations for Generative AI Outputs (Article 50)
Article 50 creates the most visible compliance obligations for every organization deploying generative AI in Europe. Unlike the high-risk provisions that apply to specific sectors, Article 50 applies broadly to any AI system that generates content interacting with natural persons.
AI Interaction Disclosure
Systems designed to interact with natural persons must inform those persons that they are interacting with an AI system, in a clear and distinguishable manner, unless this is obvious from context. Exception: lawful use for criminal investigation or for authorized security testing.
Emotion and Biometric Disclosure
Deployers using AI for emotion recognition or biometric categorization must inform exposed natural persons. This applies to video analytics, HR systems with facial analysis, and customer service emotion detection tools.
Deepfake Labeling
AI-generated images, audio, or video that falsely appears authentic must be labeled as artificially generated or manipulated. The label must be machine-readable and, where displayed to users, visible. This obligation has applied since February 2025.
Public Interest AI Text Disclosure
AI systems generating text published to inform the public on matters of general interest (news, political content, scientific articles) must label this content as AI-generated. This applies regardless of whether the publisher or an underlying API is the deployer.
Article 50 applies from August 2, 2026 for most obligations (with deep fake labeling already active). Organizations must audit all customer-facing and employee-facing AI systems to identify where disclosure is required, then implement both technical labeling and user-interface disclosure mechanisms.
5. Watermarking and Detection Requirements from the 2025 Codes of Practice
The 2025 GPAI Code of Practice — developed through the EU AI Office's multi-stakeholder process and finalized in mid-2025 — operationalizes Article 50's technical requirements. For systemic-risk model providers, it mandates specific content provenance standards that deployers must also support in their implementation.
Technical Standards Required
C2PA (Coalition for Content Provenance and Authenticity)
The dominant standard for embedding cryptographically signed provenance metadata in generated content. Required for systemic-risk providers. Deployers should use C2PA-compatible content pipelines to preserve provenance through processing.
Mandatory (systemic risk)Invisible Digital Watermarks
Imperceptible signal embedded in generated images, audio, and video. Examples include SynthID (Google DeepMind). Must survive common transformations (resizing, compression, format conversion) to be compliant.
Recommended (all GPAI)AI Text Watermarking
Statistical patterns embedded in token selection that allow detection of AI-generated text. Emerging standard — the Code of Practice requires providers to disclose their text detection capability and false-positive rates.
Emerging (systemic risk)Machine-Readable Labels
For content labeled under Article 50, the label must be machine-readable (e.g., IPTC metadata field, EXIF tag, manifest file) in addition to any visible user-facing disclosure.
Mandatory (all Article 50)For deployers, the practical implication is that content pipelines must preserve, not strip, provenance metadata. Many existing content management systems, image editors, and publishing tools strip EXIF/IPTC metadata as part of optimization workflows. This must be changed before Article 50 obligations take full effect.
6. Real Case: How a European Marketing Agency Must Comply When Using GPT-4
Consider a mid-sized European marketing agency — headquartered in Amsterdam — that uses the OpenAI GPT-4 API to generate copy for client campaigns, social media posts, and email sequences. The agency has 80 employees and serves B2C clients across the EU. What does generative AI compliance look like in practice?
The agency is a Deployer (Tier 2). OpenAI is the GPAI Provider (Tier 1). The agency cannot delegate compliance to OpenAI — it owns the obligations that arise from how it uses the API.
Low risk if addressed earlyClient-facing content generated by AI must be labeled where required by Article 50(4). The agency must review each content type: social posts (potentially AI disclosure required), news articles for client PR (AI disclosure mandatory), employee-facing drafts (internal use, disclosure to employees required).
Medium effortRequest OpenAI's training data summary (per Article 53 requirements). Add API agreement clause requiring OpenAI to indemnify against copyright claims arising from training data. Document this due diligence in the agency's AI compliance file.
Contractual action requiredEstablish a procedure for AI incidents — e.g., GPT-4 generates defamatory content about a real person in a client campaign. Log the incident, assess severity, implement corrective action. If a high-risk application is involved, report to the national market surveillance authority.
Procedural implementationReview the agency's content pipeline to ensure C2PA or equivalent provenance metadata from OpenAI's generation API is preserved through editing, formatting, and publishing workflows. Most CMSs require explicit configuration to preserve metadata.
Technical implementationArticle 4 requires AI literacy for all staff using AI systems. The agency must provide training on: what GPT-4 can and cannot do, when AI disclosure is required, how to identify and escalate AI incidents, and how to review AI outputs for quality and compliance.
Operational overheadTraceGov.ai Application: TraceGov.ai automates the compliance mapping for this agency. The TAMR+ reasoning engine (74% accuracy on EU-RegQA vs. 38.5% baseline) maps the agency's specific use of GPT-4 to the applicable Article 50 obligations, generates the required disclosure templates, and maintains the evidence trail for regulatory audit — reducing the compliance burden from weeks of legal review to hours of guided configuration.