Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

Using ChatGPT, Copilot or Gemini at Work: What the GPAI Rules Mean for Your Company

Chapter V puts the heavy duties on model providers, not on you. This guide maps Articles 53–55 to what your company must still do when staff use an LLM at work.

Renomira Team · 10 min read
Rows of server racks and cabling inside a data centre
Photo: Carl Lender from Sunrise, USA / CC BY 2.0 via Wikimedia Commons

ChatGPT drafts contract summaries, Copilot writes formulas, Gemini condenses call transcripts. The EU AI Act does not ignore any of it — but the rules do not land where most teams assume. Chapter V (Articles 53 to 55) writes a detailed rulebook for the companies that build general-purpose AI models, and a far shorter one for everyone downstream. Knowing where the handover sits — what OpenAI, Google and Anthropic owe you, and what still lands on your company — is the difference between an afternoon’s questionnaire and an audit-time gap.

What counts as a GPAI model — and what does not

A model, not an app. Article 3(63) defines a general-purpose AI model as one displaying significant generality, capable of competently performing a wide range of distinct tasks, typically trained on large amounts of data using self-supervision at scale, and integrable into a variety of downstream systems. Models used purely for research, development or prototyping before market placement are excluded.

What that covers. The foundation models behind GPT, Gemini, Claude, Llama, Mistral and DeepSeek — including those inside Copilot — are GPAI models. The chat window your staff use is a GPAI system built on one; Chapter V regulates the model underneath it, not your browser tab.

What it does not. A narrow, task-specific model — invoice OCR, churn prediction, a CV ranker — is an AI system under Chapter III. Trap: fine-tune an open-weights foundation model and both layers can apply.

Who Chapter V actually obliges

The costly misunderstanding: assuming that using a model imports the provider’s duties, or that the provider’s compliance is your compliance.

Role Typical example Rules that apply
GPAI model provider OpenAI, Google, Anthropic, Mistral Chapter V — Arts 53–55; enforcement (Arts 91–93), fines under Art 101
Downstream system provider An ISV wrapping a model into a product Chapter III if high-risk; Art 25 triggers; Art 53(1)(b) information rights
Deployer Your company using a licensed assistant Art 50 transparency, Art 4 AI literacy, Annex III duties from 2 December 2027 where high-risk; no Chapter V duties
Provider of a model you developed A company that trains and releases its own model Full Chapter V duties, including the training-content summary

Downstream users are not model providers by default. Chapter V attaches to whoever places a model on the Union market; Recital 106 confirms that adapting, integrating or distributing someone else’s model does not make you its provider. “We use GPT-4o” does not pull Article 53 onto your books.

Four things still attach to you. Article 50 transparency, Article 4 AI-literacy records, the Article 25 provider-status triggers below, and the duty not to mislead a supervisory authority under Article 99.

The Article 53 checklist: what your vendor owes you

The four core obligations. Article 53(1) is the workhorse of GPAI regulation. Every GPAI provider must:

What the provider must do Article What you should ask for
Technical documentation of training, testing and evaluation (Annex XI) Art 53(1)(a) Compliance summary or audit attestation
Capability and limitation information for integrating providers (Annex XII) Art 53(1)(b) Model card, integration guidance, version-change notice
Copyright policy honouring machine-readable rights reservations Art 53(1)(c) The policy and its copyright contact point
Detailed public summary of training content on the AI Office template Art 53(1)(d) The public summary — read it before relying on it

The training-data summary is what you can read. The AI Office template arrived in July 2025, covering data sources, content volume and how rights reservations were handled. For a publishing or market-research client whose value is proprietary archives, it answers procurement’s first question: was our material in the training set?

Where the open-source exemption stops. Article 53(2) frees genuinely open-weights models from points (a) and (b) only; the copyright policy and training summary still apply, and the exemption never covers a systemic-risk model. Open source is a lighter burden, not an absent one.

Compliance through the code of practice. Under Articles 53(4) and 56, providers may rely on an approved code of practice until a harmonised standard is published. The GPAI Code of Practice, finalised in July 2025 and signed by most major providers, is your fastest way to check a vendor’s posture.

Systemic-risk models: Article 55 and the heavier tier

The threshold. Article 51(2) presumes systemic-risk status when cumulative training compute exceeds 10²⁵ floating-point operations; the Commission may also designate models ex officio against the Annex XIII criteria.

What Article 55 adds. Model evaluation and documented adversarial testing, assessment and mitigation of systemic risks, serious-incident reporting to the AI Office without undue delay, and adequate cybersecurity for the model and its infrastructure.

Why it matters to a deployer. You will not meet these duties directly, but you will be asked about them. The Code of Practice’s Safety and Security chapter expects a risk framework, adversarial testing and incident reporting; its Transparency chapter expects a Model Documentation Form retained for ten years. Providers who did not sign must show alternative adequate means — ask which route they took.

The dates that matter from October 2026

  • 2 August 2025 — Chapter V applied. GPAI provider obligations became live under Article 113(b).
  • 2 August 2026 — enforcement and transparency. The general applicability date brought Article 99 penalties and Article 50 transparency duties into force.
  • 2 December 2026 — synthetic-content marking. Systems placed on the market before 2 August 2026 must comply with Article 50(2) by this date (Article 111(4)).
  • 2 August 2027 — legacy models. GPAI models placed on the market before 2 August 2025 must be brought into compliance (Article 111(3)).
  • 2 December 2027 — Annex III high-risk duties, moved from 2 August 2026 by the Digital Omnibus.

In practice, the only GPAI deadline ahead belongs to your vendors’ oldest models.

Where your own duties start: Article 50

Chatbot disclosure (Art 50(1)). Offer a customer-facing assistant and you are the provider of that AI system: tell users they are interacting with an AI system, unless that is obvious.

Synthetic-content marking (Art 50(2)). Providers must mark generated audio, image, video and text in a machine-readable, detectable way. Deployers must not strip watermarks or metadata when republishing generated assets.

Deepfake and public-interest text disclosure (Art 50(4)–(5)). Deployers must disclose deepfakes — AI-generated campaign imagery, for instance — and AI-generated text published to inform the public on matters of public interest, such as AI-written press releases on safety or environmental claims.

Emotion recognition (Art 50(3)). Deployers must inform exposed persons — but emotion inference in the workplace is separately prohibited under Article 5, so disclosure cannot rescue it.

When customisation makes you a provider: Article 25

The three triggers. Article 25(1) deems you the provider of a high-risk AI system — with Article 16 obligations — if you (a) put your own name or trademark on a system already on the market, (b) make a substantial modification, or (c) modify its intended purpose so it becomes high-risk.

Mid-market examples. A logistics firm rebrands a vendor’s screening assistant as its own: trigger (a). An engineering consultancy fine-tunes an open-weights model on incident reports and sells the resulting hazard-prediction tool: substantial modification plus purpose change, making it potentially a GPAI provider in its own right.

Fine-tuning alone is not the line. The Commission’s July 2025 GPAI guidelines set compute-based thresholds for substantial modification. Prompting, retrieval-augmented generation and modest fine-tuning keep you downstream; a significant change in generality, capability or systemic-risk profile may not. Document the assessment either way.

What to ask for in procurement and contract reviews

  • Ask for posture, not marketing. Which articles apply, did the vendor sign the code of practice, and which chapters?
  • Request the training-data summary and copyright policy — both are Article 53(1) products, so their absence is itself a finding.
  • Insist on version-change notification. A model update can move your classification; silent upgrades cause drift.
  • Write Art 53(1)(b) information rights into the contract, with limitation disclosures and a named contact point.
  • Allocate Article 50 duties. Who marks synthetic output, who discloses deepfakes, who updates labelling when the tool changes.
  • Keep answers accurate. Misleading information to an authority is a Tier 3 penalty under Article 99 — questionnaire answers count too.

How to check your own position

  1. List every LLM touchpoint — model and version, interface, business owner. Include shadow AI; discovery usually surfaces 20 to 60 tools.
  2. Write your role in one word per line: deployer, system provider, or model provider. If you cannot decide, it is an Art 25 question — resolve it before the register freezes.
  3. Run the Article 50 test: does anything customer-facing interact with people, publish generated media or deepfakes, or issue AI-generated text on matters of public interest?
  4. Run the Article 25 test: our name on it, substantial modification, or changed intended purpose?
  5. Collect vendor evidence: training-data summary, copyright policy, code-of-practice signature, systemic-risk status, version history.
  6. Flag Annex III uses — recruitment, credit, education — for 2 December 2027 rather than treating today’s silence as compliance.
  7. Check Article 4 literacy records for everyone in steps 1 to 6; that duty has applied since 2 February 2025.

Frequently asked questions

Does using ChatGPT at work make our company a GPAI provider?

No. Chapter V obligations follow whoever places a model on the Union market under their own name or trademark. Using or integrating a licensed model keeps you a deployer. Provider status shifts only under Article 25 — your brand on the system, substantial modification, or changed purpose — or when you train and release a model yourself.

What exactly is a GPAI model under the AI Act?

Article 3(63) defines it as an AI model displaying significant generality, capable of competently performing a wide range of distinct tasks, and integrable into a variety of downstream systems — typically trained at scale with self-supervision. GPT, Gemini, Claude and Llama qualify; narrow single-task models, and models used only for research or prototyping, do not.

If our vendor is non-compliant, are we liable?

The Article 53 duties belong to the model provider, whom the Commission enforces directly with fines up to €15 million or 3% of worldwide turnover under Article 101. Your exposure differs: contract and audit risk, misleading answers to authorities, and your own duties under Articles 50 and 4. Vendor non-compliance is a procurement problem to document, not a fine you inherit.

When must we label AI-generated content?

Article 50(2) marking binds providers; systems placed on the market before 2 August 2026 must comply by 2 December 2026 (Article 111(4)). As a deployer, your deepfake and public-interest disclosure duties under Article 50(4)–(5) have applied since 2 August 2026, and you must never strip a provider’s machine-readable marks.

Do we have to sign the GPAI code of practice?

No. It exists under Article 56 as a voluntary route for providers to demonstrate compliance with Articles 53 and 55 until harmonised standards arrive; providers who skip it must show alternative adequate means. Record, however, whether your vendors signed it and which chapters — transparency, copyright, and safety and security — they committed to.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

gpai obligations downstream usersgeneral purpose ai model act article 53what is a gpai model ai actai act llm provider obligationsusing llm at work ai act compliancegpai code of practice complianceai act training data summary gpai

Keep reading