AI News ai-governance eu-ai-act
AI News Medium

The EU AI Act’s GPAI rules are now a procurement question

General-purpose AI obligations began applying in August 2025. If you buy or deploy a model, the documentation you can obtain is now part of your risk assessment.

KT Kusal Tharindu

The EU AI Act is not only a rule for companies building high-risk AI systems. It also creates a documentation layer around general-purpose models, and that layer reaches anyone who deploys one.

From August 2, 2025, providers of general-purpose AI models placed on the EU market are subject to obligations including technical documentation, a summary of training content, a copyright policy, and information for downstream providers. Models with systemic risk have additional evaluation and incident-reporting duties. The Act gave models already on the market a transition period, so the exact date depends on the model and the provider.

Why a deployer should care

You may not be the provider. You are still the party that chose the model, integrated it, and put it in front of users. When a regulator, customer, or insurer asks how you assessed the model, “the vendor’s website looked fine” is not an answer.

The useful artifacts to request:

  • The model card and intended-use statement.
  • A summary of training content, at the level the provider is required to publish.
  • Evaluation results relevant to your use case, including known limitations.
  • The provider’s acceptable-use policy and any use restrictions.
  • Versioning and change-notification commitments.
  • Security documentation: data retention, logging, subprocessors, and incident contact.

A procurement checklist

  • Model provider and model version are recorded, not just the product name.
  • The provider’s GPAI documentation is saved with the assessment date.
  • The intended use matches your use; prohibited or high-risk uses are flagged for legal review.
  • Data residency and retention are documented and match the contract.
  • A model change triggers a review, not a silent upgrade.
  • The vendor can explain how it handles training-data rights and opt-outs.
  • There is an escalation path for a security or safety incident.
  • The assessment is repeated when the model, provider, or use changes.

The practical shift is small but real: model selection is now a documented decision. Keep the documentation with the model version, and treat a provider’s silence as a risk, not a clean bill of health.

Sources

  1. European Commission: AI Act regulatory framework
  2. EUR-Lex: Regulation (EU) 2024/1689
  3. European Parliament: AI Act overview
3 posts

Discussion

Spotted an error? Send corrections to [email protected] .

Search Oxelan Signal

Type to search the archive.

Search by title, tag, summary or body text.