In this article8 sections
  1. Your Tool Vendor Has Legal Obligations, Just Not All of Them to You
  2. What the AI Act Actually Requires From GPAI Providers
  3. What You Can Ask For as a Deployer
  4. Red Flags in Current Tool Terms of Service
  5. What to Keep in Your Vendor File
  6. Who Gets Fined, and for What
  7. One Action to Take This Week
  8. References

GPAI vendor obligations under the EU AI Act are not optional extras buried in enterprise pricing tiers. They are legal requirements for every provider placing a general-purpose AI model on the EU market, whether that provider is based in San Francisco or Seoul. But they are owed to specific parties: the AI Office and national authorities, and companies that build their own AI systems on top of the model. If your company simply uses ChatGPT, Claude, or Gemini in its daily work, you are a deployer of an AI system, and your own obligations, starting with Article 4, are what matter most. Knowing who owes what to whom stops you from chasing documents nobody has to give you, and from missing the ones you can get.

The Act entered into force on 1 August 2024. The provisions governing general-purpose AI (GPAI) models, found in Articles 51 through 56, have applied since 2 August 2025. Models already on the market before that date have until 2 August 2027 to comply. Since 2 August 2026 the Commission can fine GPAI model providers.

What the AI Act Actually Requires From GPAI Providers

A GPAI model is an AI model trained on broad data that can competently perform a wide range of tasks: text generation, summarisation, code, translation. The models behind ChatGPT, Claude and Gemini are typical examples. Under Article 53, providers of these models must:

  • Draw up and keep up to date technical documentation of the model, including its training and testing, to provide to the AI Office and national competent authorities on request.

  • Make information and documentation available to providers of AI systems who intend to integrate the model into their own AI systems, so those downstream providers understand its capabilities and limitations and can meet their own obligations.

  • Put in place a policy to comply with EU copyright law.

  • Publish a sufficiently detailed summary of the content used for training.

There is no duty to register a GPAI model in the EU database. That database is for high-risk AI systems.

For models with systemic risk (presumed when training compute exceeds 10^25 floating-point operations, or designated by the Commission), Article 55 adds further obligations: model evaluation including adversarial testing, assessing and mitigating systemic risks, tracking and reporting serious incidents to the AI Office, and adequate cybersecurity. Which specific models are classified this way is a matter for the providers and the Commission; we do not list them here.

Providers can show compliance by following the General-Purpose AI Code of Practice, published in July 2025, which the Commission and the AI Board have confirmed as adequate. Most large model providers have signed it, in whole or in part.

What You Can Ask For as a Deployer

This is the practical question: what can you actually expect?

It depends on how you use the model.

If you use a finished tool (ChatGPT, Microsoft Copilot, Gemini in Workspace), you are a deployer of that AI system. Article 53(1)(b) does not give you a right to the model provider's integration documentation. What you can get is what providers publish or put in their business terms: usage policies, model or system cards, the public training-content summary, the copyright policy, and a data processing agreement for the GDPR.

If you build your own AI system on a model API (for example a customer chatbot or a document classifier on top of Claude or GPT), you are the provider of that AI system. Then Article 53(1)(b) is written for you: the model provider must make information and documentation available that lets you understand the model's capabilities and limitations and comply with your own obligations, such as the Article 50 transparency duties.

Either way, useful things to obtain are:

Usage instructions. Guidance on intended use cases, known failure modes, and contexts where the model should not be used without additional safeguards.

Capability and limitation disclosure. As specific as you can get: known performance gaps on certain languages, behaviour on domain-specific tasks, limitations with dates and real-time information. This is also the raw material for your Article 4 training: the Commission's AI literacy Q&A says staff using tools like ChatGPT should be informed about specific risks such as hallucination.

Information for higher-risk use cases. If you plan to use a general-purpose model for something listed in Annex III, such as screening candidates or assessing creditworthiness, stop and get advice first. Whether that makes your setup a high-risk AI system, and whether that makes you its provider under Article 25, depends on how you build and use it. High-risk deployers face Article 26 duties from 2 December 2027, and providers of high-risk systems far more.

Change and incident channels. How you will hear about material model changes, and where to report problems.

Red Flags in Current Tool Terms of Service

Watch for these specific red flags:

Broad liability disclaimers. A clause stating the provider bears "no responsibility for your use of outputs" is common and not automatically unlawful. It cannot remove the provider's own statutory obligations under Articles 53 and 55, but it also does not create duties the Act puts on you. Read it to understand which risks stay with you.

No documentation at all. If a vendor publishes no model or system documentation, no usage policy and no training-content summary, ask directly. A general-purpose model provider that publishes no training summary is not meeting Article 53(1)(d).

Incident reporting opacity. Systemic-risk model providers must report serious incidents to the AI Office. That duty does not run to you, so if you depend on the model for something important, ask in your contract how and when you will be told about incidents or material changes.

Unilateral model update clauses. Several providers reserve the right to update the underlying model without notice. If your workflow relies on a specific model version and the provider silently swaps it, your earlier testing and your staff guidance may no longer fit. Treat version changes as a trigger to re-check.

What to Keep in Your Vendor File

Negotiating with OpenAI or Google on contract terms is not realistic for most SMEs. But enterprise agreements, API terms, and data processing addenda often have more flexibility than the standard consumer terms suggest. Even where you cannot negotiate, documenting what the vendor provides (and what it does not) is good practice.

At a minimum, your vendor file for any GPAI tool should contain:

  1. A copy of the provider's current documentation or model card, with a retrieval date. Links go stale; save the document.

  2. The model version your integration uses, so that changes are detectable.

  3. The provider's incident and change notification policy, including contact details.

  4. The provider's public copyright policy and training-content summary, where published.

  5. The data processing agreement you rely on under the GDPR.

Who Gets Fined, and for What

Under Article 101, the Commission can fine providers of GPAI models up to 3% of global annual turnover or €15 million, whichever is higher, for breaching their GPAI obligations. Those powers apply from 2 August 2026. The AI Office, not national authorities, supervises GPAI model providers.

Those fines fall on model providers, not on you. Your exposure as a deployer comes from your own obligations: Article 4 AI literacy since 2 February 2025, Article 50 if you deploy deepfakes or publish AI-generated text on matters of public interest, Article 26 if you use a high-risk system, and the GDPR throughout. Your national market surveillance authority supervises those. In the Netherlands, the Autoriteit Persoonsgegevens is the intended supervisor for most AI topics, but it only gets AI Act powers once the Dutch implementing law is in force.

Start the paper trail today. Save the vendor documentation. Note what is missing. It feeds both your staff training and any future question about how you chose your tools.

One Action to Take This Week

Pull the terms of service and any available model cards for every GPAI tool currently in use at your company. Check whether you have usage instructions, a capability and limitation description, the copyright policy and training summary, and a data processing agreement. Note the gaps. Then use what you found to brief the staff who use those tools: that is a concrete Article 4 measure.

If you want a structured way to run that exercise across your full tool stack, Khairos Comply offers a free 2-minute compliance check that maps your current tools against your EU AI Act obligations as a deployer. Just a clear picture of where you stand.

References

  1. Articles 51 through 56

  2. Article 53

  3. Article 55

  4. Article 101

  5. Article 4

  6. Article 26

  7. European Commission, General-Purpose AI Code of Practice

  8. Khairos Comply

Not sure where your company stands?

Answer 15 questions about how your team uses AI. You get a compliance score and the gaps to close first, as a PDF.

Am I compliant? Check in 2 minHow Khairos Comply works