Teams building on large language models under the EU AI Act often cannot tell whether they face the general-purpose AI (GPAI) model rules, the high-risk AI system rules, both, or neither, and a wrong answer means either missed obligations or wasted conformity work. The two regimes regulate different objects: the GPAI chapter governs the model whatever its eventual use, while the high-risk chapter governs an AI system because of what it is used for. This guide is a classification router. It sets out the tests in order, shows where the regimes meet and links out to the detailed obligations. It is written for AI platform, product, legal and risk teams. Status checked on 13 November 2026 against the European Commission's AI Act Service Desk consolidated text (as at 27 July 2026, including Digital Omnibus on AI amendments) and the Commission's AI Act policy pages.
Key takeaways
- Model rules and system rules are separate layers. The Commission's GPAI questions and answers state that the GPAI provisions apply to the model itself, regardless of its ultimate use, while the system provisions depend on the context of use.
- Both can apply at once. A GPAI model provider has GPAI duties, and whoever builds a high-risk system on that model has high-risk duties; when one company does both, it carries both sets.
- Repurposing is the common trigger. Article 25(1)(c) makes anyone who changes the intended purpose of an AI system, including a general-purpose AI system, so that it becomes high-risk under Article 6 the provider of that high-risk system.
- Systemic risk is a model test, not a use test. A model is presumed to have high-impact capabilities above 10^25 floating-point operations of training compute (Article 51(2)), and the Commission can also designate models (Article 52(4)).
- The timelines differ. GPAI obligations applied from 2 August 2025, with Commission fining powers from 2 August 2026; high-risk obligations apply from 2 December 2027 (Annex III) and 2 August 2028 (Annex I), as amended by the Digital Omnibus on AI.
Short answer: which regime governs what
The Act defines a general-purpose AI model as a model that displays significant generality and can competently perform a wide range of distinct tasks, and that can be integrated into a variety of downstream systems (Article 3(63)). A general-purpose AI system is an AI system based on such a model that can serve a variety of purposes, both for direct use and for integration into other AI systems (Article 3(66)). High-risk status, by contrast, attaches to an AI system under Article 6, either through product safety legislation in Annex I or through the use cases in Annex III.
| Question | GPAI model regime (Chapter V) | High-risk system regime (Chapter III) |
|---|---|---|
| What is regulated | The model, whatever its use | The AI system, because of its intended purpose or product |
| Who is the main duty holder | Provider of the model (Article 53); systemic-risk providers also under Article 55 | Provider of the system (Article 16), plus deployer duties (Article 26) |
| Classification test | Significant generality (Article 3(63)); systemic risk by compute presumption or designation (Articles 51 and 52) | Annex I safety component with third-party assessment (Article 6(1)) or Annex III use case (Article 6(2)), subject to the Article 6(3) derogation |
| Core duties | Technical documentation, information for downstream providers, copyright policy, public training content summary (Article 53(1)(a) to (d)) | Risk management, data governance, documentation, logging, transparency, human oversight, accuracy, robustness and cybersecurity, quality management, conformity assessment, registration |
| Open-source treatment | Article 53(1)(a) and (b) do not apply to qualifying free and open-source models, except those with systemic risk (Article 53(2)) | No equivalent exemption for a high-risk system |
| Main supervisor | The Commission through the AI Office; fines up to 3% of worldwide turnover or EUR 15 million, whichever is higher (Article 101) | National market surveillance authorities, with AI Office competence for systems built on a GPAI model by the same provider, subject to exclusions (Article 75(1)) |
| Application date | 2 August 2025; models placed on the market before then comply by 2 August 2027 | 2 December 2027 (Annex III) and 2 August 2028 (Annex I) |
The duties in the right-hand column are covered in the EU AI Act High-Risk AI Systems Guide; the left-hand column is covered in the EU AI Act GPAI Obligations Guide.
The classification router
Run these tests in order for each model and each AI system you build, buy, fine-tune or release.
- Is there a GPAI model you place on the market under your name? If you train or have trained a model that meets Article 3(63) and make it available in the EU (through an API, download or product), you are a GPAI model provider. Models used only for research, development or prototyping before being placed on the market are outside the definition. If you only use another provider's model, you are not its provider unless you modify it significantly (see step 3).
- Does that model have systemic risk? Above 10^25 FLOP of cumulative training compute, high-impact capabilities are presumed (Article 51(2)). The provider must notify the Commission without delay and in any event within two weeks of meeting the requirement or knowing it will be met (Article 52(1)), and may submit substantiated arguments that the model exceptionally does not present systemic risk (Article 52(2)). Below the threshold, the Commission can still designate a model using the Annex XIII criteria (Articles 51(1)(b) and 52(4)).
- Have you modified someone else's model? The Commission's GPAI guidelines page states that only significant modifications trigger provider obligations, and the Commission's questions and answers say that where a downstream modifier becomes a provider, the Article 53 obligations should be limited to the modification or fine-tuning. Document the scale and nature of every modification and compare it with the Commission guidelines.
- Is the AI system you build or use high-risk? Check Article 6(1): is the system a safety component of, or itself, a product under Annex I legislation that requires third-party conformity assessment? The Digital Omnibus added paragraphs 1a to 1c, which state that systems used solely for non-safety purposes such as user assistance, performance optimisation or convenience are not safety components unless their failure would endanger health and safety. Then check Article 6(2): does the intended purpose fall in an Annex III area such as employment, education, essential services or biometrics?
- Does the Article 6(3) derogation apply? An Annex III system is not high-risk if it does not pose a significant risk of harm, including by not materially influencing the outcome of decision making, and it performs a narrow procedural task, improves the result of a completed human activity, detects decision patterns without replacing human assessment, or performs a preparatory task. A system that profiles natural persons is always high-risk. A provider relying on the derogation must document the assessment before placing the system on the market and register it under Article 49(2) (Article 6(4)).
- Who are you for that system? If you develop the high-risk system and place it on the market or put it into service under your name, you are its provider. If you use it under your authority, you are a deployer. If you rebrand it, substantially modify it or change its intended purpose so that it becomes high-risk, Article 25(1) makes you the provider. The EU AI Act Provider vs Deployer Obligations Guide covers this test in depth.
Router outcomes
| Your position | GPAI duties | High-risk duties | Where to go next |
|---|---|---|---|
| You train and release a GPAI model; you build no system | Yes (Article 53; Article 55 if systemic risk) | No, unless you also place a high-risk system on the market | GPAI obligations guide |
| You integrate a third-party GPAI model into a chatbot for general internal use | No (you are a downstream provider of a system, not of the model) | No, unless the use falls in Annex I or Annex III; transparency duties may still apply | EU AI Act Article 50 Transparency Obligations Guide |
| You integrate a third-party GPAI model into a CV-screening product you sell | No, unless you significantly modify the model | Yes, as provider of an Annex III employment system | High-risk systems guide; obtain Article 53(1)(b) information from the model provider |
| You use a general-purpose assistant to rank job applicants | No | Yes, as provider under Article 25(1)(c), because you changed the intended purpose so it became high-risk | Provider vs deployer guide |
| You train a model and build your own high-risk product on it | Yes | Yes | Both guides; check AI Office competence under Article 75(1) |
| You fine-tune an open model significantly and release it | Yes, limited to the modification | Only if you also place a high-risk system on the market | GPAI obligations guide |
Where the regimes meet: a GPAI model inside a high-risk system
This is the case most organisations get wrong, because the duties sit with different parties and depend on information flowing between them.
Architecture in words. A model provider releases a GPAI model through an API. A downstream provider (Article 3(68)) builds an AI system on it, adds prompts, retrieval, tools and a user interface, and places it on the market for an Annex III purpose. A deployer uses that system on real people. The model provider's obligations travel with the model; the system provider's obligations attach to the finished system; the deployer's obligations attach to use.
- The downstream provider carries the high-risk duties. The Commission's questions and answers state that an entity incorporating a GPAI model into an AI system must comply with the AI Act requirements for AI systems, whether or not it is also deemed a provider of the model. The model provider's GPAI compliance does not discharge any high-risk requirement.
- Information must flow down. Article 53(1)(b) requires GPAI providers to give providers who intend to integrate the model information and documentation that lets them understand its capabilities and limitations and comply with their own obligations, containing at least the Annex XII elements. Your Article 9 risk management and Article 15 robustness work depends on it.
- Repurposing creates a new provider. Under Article 25(1)(c), modifying the intended purpose of an AI system, including a general-purpose AI system, so that it becomes high-risk makes you its provider. Under Article 25(2), the initial provider is no longer the provider of that specific system but must cooperate and make information, technical access and other assistance available, unless it clearly specified that its system is not to be changed into a high-risk system.
- Written agreements are expected. Article 25(4) requires the provider of a high-risk system and third parties that supply components, tools or models used in it to specify the necessary information, capabilities, technical access and assistance by written agreement. The exemption for free and open-source tools does not extend to general-purpose AI models.
- Supervision can shift. Where the model and the system are developed by the same provider or within the same undertaking, Article 75(1)(a) as amended gives the AI Office exclusive competence, except for systems linked to Annex I legislation, Annex III point 2, law enforcement, border management, certain financial institution systems and Annex III point 8. Under Article 75(2), national authorities that consider a general-purpose AI system usable directly by deployers for at least one high-risk purpose to be non-compliant cooperate with the AI Office.
Threat model and failure modes for each regime
The regimes protect against different failures, and compliance programmes fail in different ways under each. The EU AI Act Article 15 Guide covers the high-risk cybersecurity requirement in detail.
| Aspect | GPAI model regime | High-risk system regime |
|---|---|---|
| Harm it targets | Opaque models propagating into many products; for systemic-risk models, Union-level risks to public health, safety, security, fundamental rights or society (Article 3(65)) | Harm to individuals from specific decisions: hiring, credit, education, access to services, safety of products |
| Security-relevant duty | Systemic-risk providers must evaluate the model including adversarial testing, assess and mitigate systemic risks, report serious incidents to the AI Office and ensure adequate cybersecurity for the model and its physical infrastructure (Article 55(1)(a) to (d)) | Accuracy, robustness and cybersecurity of the system, logging and human oversight |
| Typical failure | Downstream information pack missing the limitations integrators need; fine-tuned releases not assessed as possible new models; compute records not kept, so the systemic-risk notification deadline is missed | Annex III use not recognised because the system was bought as a general tool; derogation claimed without documented assessment; logs not retained; no written agreement with the model supplier |
| Controls that address it | Model register with training compute, release and modification records; downstream documentation pack mapped to Annex XII; adversarial testing and incident process for systemic-risk models | Use-case register checked against Annex III; documented Article 6(3) assessments; supplier agreements under Article 25(4); logging and oversight built into the system |
| Route to show compliance | Codes of practice under Article 56 until a harmonised standard is published, then harmonised standards (Article 55(2)) | Harmonised standards and conformity assessment; see the EU AI Act Conformity Assessment and Harmonised Standards Guide |
Timelines side by side
| Date | GPAI regime | High-risk regime | Source |
|---|---|---|---|
| 2 August 2025 | Chapter V obligations apply to GPAI providers | Notified body provisions (Chapter III Section 4) apply | Article 113(b) |
| 2 August 2026 | Commission enforcement powers, including fines, apply | General application of the Act | Commission GPAI guidelines page; Article 113 |
| 2 August 2027 | Deadline for models placed on the market before 2 August 2025 | None | Commission GPAI guidelines page |
| 2 December 2027 | None | Chapter III Sections 1 to 3 apply to Annex III systems | Article 113(c)(i), as amended |
| 2 August 2028 | None | Chapter III Sections 1 to 3 apply to Annex I systems | Article 113(c)(ii), as amended |
The Digital Omnibus on AI entered into force on 27 July 2026 and moved the high-risk dates; the Commission's policy page does not describe a change to the GPAI dates. Dates can be amended again, so check the EU AI Act Implementation Timeline Tracker and the official text before relying on them.
Worked scenarios
- HR software vendor adding an LLM: a vendor builds an applicant-ranking feature on a commercial GPAI model. The feature is an Annex III employment use, so the vendor is the provider of a high-risk system from 2 December 2027. It is not a GPAI provider. Its actions: request the model provider's Article 53(1)(b) documentation, put a written agreement in place under Article 25(4), and build risk management, logging and human oversight into the feature. Its customers are deployers under Article 26.
- Bank using a general assistant: a credit team starts using the bank's licensed general-purpose assistant to score small-business loan applicants. The assistant was not intended for this. Changing its purpose to an Annex III credit use makes the bank the provider of a high-risk system under Article 25(1)(c). The fix is governance: block the use until a properly classified system is in place, or run it through the full provider programme.
- Model developer with its own product: a company trains a model above the 10^25 FLOP presumption and sells an education assessment product on it. It notifies the Commission within two weeks under Article 52(1), meets Articles 53 and 55 for the model, and meets the high-risk requirements for the product. Article 75(1)(a) gives the AI Office competence for the system, because model and system come from the same provider and the system is outside the listed exclusions.
- Internal coding assistant: an engineering team fine-tunes an open model lightly for internal code completion and does not release it. The model is not placed on the market, and code completion is not an Annex III use. Neither regime's core duties apply, but AI literacy and any transparency duties should still be checked with the EU AI Act Compliance Guide.
Decision checklist
- Model register: every model you train, fine-tune or release, with training compute, release channel and modification record.
- Use-case register: every AI system with its intended purpose mapped against Annex I and Annex III, and the name of the business owner.
- Derogation file: a dated Article 6(3) assessment for every Annex III system you treat as not high-risk, with the registration under Article 49(2).
- Purpose-change control: a gate that stops a team repurposing a general tool for an Annex III decision without classification review.
- Supplier evidence: Article 53(1)(b) documentation and an Article 25(4) agreement for each GPAI model used in a high-risk system.
- Supervisor map: for each system, whether the AI Office or a national market surveillance authority is competent.
Detection and measurement
- Classification coverage: percentage of AI systems in the inventory with a recorded classification decision and evidence.
- Unreviewed purpose changes: number of AI uses discovered (through procurement, access reviews or usage monitoring) in Annex III areas without a classification record; the target is zero.
- Supplier documentation: percentage of GPAI models in high-risk systems with downstream documentation received and reviewed.
- Derogation quality: percentage of Article 6(3) claims with a documented assessment and registration before go-live.
- Compute tracking: for model developers, training compute recorded per run, with an alert as cumulative compute approaches 10^25 FLOP.
Discovery of unclassified uses is largely an inventory and identity problem: AI systems run under service accounts, API keys and agent identities that can be found in access reviews. The AI Usage Monitoring Guide explains how to find them.
Framework mapping
| Framework | Reference | How the router applies it |
|---|---|---|
| NIST CSF 2.0 | GV.OC-03 | Legal and regulatory requirements for each model and system understood and managed |
| NIST CSF 2.0 | ID.AM-02 | Inventory of models and AI systems as the basis for classification |
| NIST CSF 2.0 | GV.SC-05 | Article 25(4) and Article 53(1)(b) requirements written into supplier agreements |
| NIST SP 800-53 Rev 5 | PM-9, SA-9, SR-3 | Risk strategy, external services and supply chain controls for third-party models |
| NIST SP 800-53 Rev 5 | CA-2 | Assessment evidence for classification and derogation decisions |
Common failure patterns
- Assuming the model provider is responsible: integrators believe the model vendor's GPAI compliance covers their high-risk product. It does not.
- Silent repurposing: a business unit uses a licensed assistant for hiring or credit decisions, making the organisation a provider under Article 25(1)(c) without anyone knowing.
- Undocumented derogation: Article 6(3) is relied on without the written assessment and registration Article 6(4) requires, or for a system that profiles people.
- Relying on the open-source label: assuming the Article 53(2) exemption covers systemic-risk models or high-risk systems built on open models.
- Planning to the old dates: programmes still working to the pre-Omnibus high-risk dates, or assuming the GPAI dates moved with them.
Practitioner checklist
- Build one register that links each model to every AI system built on it.
- Run the six router steps for each entry and store the evidence.
- Record training compute for in-house models and set the Article 52 notification process.
- Check each system's intended purpose against Annex I and Annex III, and apply Article 6(1a) to (1c) for product safety cases.
- Document any Article 6(3) derogation before go-live and register it.
- Obtain Article 53(1)(b) documentation and sign Article 25(4) agreements with model suppliers.
- Put a purpose-change gate into AI procurement and onboarding.
Standards and references
- European Commission AI Act Service Desk: Article 6, Classification rules for high-risk AI systems
- European Commission AI Act Service Desk: Article 25, Responsibilities along the AI value chain
- European Commission AI Act Service Desk: Article 51, Classification of GPAI models with systemic risk
- European Commission AI Act Service Desk: Article 53, Obligations for providers of GPAI models
- European Commission AI Act Service Desk: Article 75, Mutual assistance, market surveillance and control of general-purpose AI systems
- European Commission AI Act Service Desk: Article 113, Entry into force and application
- European Commission: General-Purpose AI Models in the AI Act, Questions and Answers
- European Commission: AI Act regulatory framework
Related NHI Mgmt Group resources: EU AI Act GPAI Obligations Guide, EU AI Act High-Risk AI Systems Guide, EU AI Act Provider vs Deployer Obligations Guide, EU Digital Omnibus Guide. This guide summarises the classification rules of Regulation (EU) 2024/1689 as amended and is not legal advice.