Following the EU AI Code will not keep a model available

The voluntary GPAI framework structures compliance evidence, while classification, testing, contracts and…

The short version

The EU AI Act Code of Practice standardises compliance evidence; it does not certify models or classify finished applications.
Signatories use shared documentation, while unsigned providers may demonstrate compliance through a persuasive alternative evidence framework.
Buyers still need application-level testing, clear responsibility mapping, contractual update processes and a tested migration path.

A signed EU AI code buys organised evidence. Anyone selling it as a safety certificate is handing procurement a very expensive sticker.

The EU AI Act Code of Practice gives general-purpose AI model providers a voluntary path to compliance. Signing helps regulators and buyers inspect their work. It cannot classify my application, certify the model or complete my company’s AI Act assessment. The Code standardises evidence; it is not a green badge that ends the meeting.

One warning before Brussels turns the acronyms into minestrone: this is the General-Purpose AI Code of Practice. The separate transparency code covers marking, detection and labelling of AI-generated content. Similar name. Different legal job.

Your product still needs its own EU AI Act classification

A general-purpose model can power wildly different products, so I draw the legal boundary before reading any vendor badge. When I call Claude, GPT or Mistral through an API for customer support, I usually operate at the AI-system layer while another company supplies the model. That changes if my company substantially modifies it or places it on the European market under its own control. The contract needs the model version, the responsible entity at each layer and a process for material updates. “Enterprise AI partner” has no magical legal meaning, despite the tasteful navy-blue sales deck. Leave responsibilities vague and everyone discovers they own the problem simultaneously.

People searching for EU AI Act risk levels often assume the underlying model determines every product’s classification. It doesn’t. GPAI rules govern the model and provider; a hiring tool, credit scorer or medical application is classified by the finished system and its use. A signatory cannot classify a workflow it has never seen. My team may still need to assess prohibited practices, high-risk requirements and application-level transparency duties. The customer deploying it may have separate obligations too.

The General-Purpose AI Code has three chapters. Transparency and Copyright provide a shared way to demonstrate Article 53 compliance, including technical documentation, downstream information, a copyright policy and published training-material information. Safety and Security addresses Article 55 through evaluations, adversarial testing, incident reporting and cybersecurity controls for systemic-risk models. The Commission and AI Board confirmed that the Code is an adequate voluntary compliance tool, letting signatories use adherence as evidence. Signatories then formed an AI Office-chaired taskforce to apply it more coherently. An unsigned provider can build an alternative compliance case. Declining the prix fixe menu does not cancel the bill.

The Commission published the Code on 10 July 2025. Enforcement powers for the AI Office and national authorities began applying on 2 August 2026; those specified powers did not apply earlier.

For relevant GPAI breaches, the Commission’s framework allows fines up to €15 million or 3% of worldwide annual turnover, whichever is higher. The amount depends on the infringement’s nature, gravity and duration. That exposure belongs to the provider that breached its obligations, not automatically every startup making API calls.

A signature makes AI due diligence less chaotic

Once a provider signs the EU AI Act Code of Practice, I can request evidence against a known structure instead of another bespoke legal novella. The Transparency chapter includes a Model Documentation Form containing information regulators and downstream builders need. The Copyright chapter turns “trust us” into an expected policy backed by processes. Systemic-risk model providers use Safety and Security to document evaluations, mitigations, serious-incident handling and security practices. The AI Office-chaired taskforce helps signatories resolve inconsistent interpretations. Reusable evidence can shorten buyer diligence and providers’ repetitive work, but only when they perform the commitments instead of wearing the signature as website jewellery.

“Adequate” describes the compliance route. It does not mean the Commission approved an individual model. A completed form cannot prove outputs are accurate or my deployment remains secure after connecting the model to customer data. Nor can it tell me whether my logic makes an HR ranking workflow discriminatory.

The strongest objection deserves a fair hearing: common documentation can make buyers mistake a clean file for safety. Anthropic’s Code Review documentation offers a useful parallel. It warns that AI review may miss bugs and security vulnerabilities, so no findings do not prove secure code. Compliance evidence has the same epistemic limit. It shows what process a provider claims to follow and gives regulators structured evidence to inspect. I still need application-level testing, a human decision owner and a contract covering what happens when the evidence changes.

Anthropic described the continuing review bottleneck:

Coding agents keep increasing the pace at which teams write code, and the review bottleneck hasn't eased.

Claude Code’s useful lesson is not that AI replaces oversight, but that oversight can follow risk. The team assigns internal changes zero, one or two human approvers; some code paths always require a person, while unresolved AI-review or security findings block merging. Anthropic says Claude writes about 90% of Claude Code’s own code, so review must scale without pretending the machine is infallible. From June to September, merged pull requests more than doubled while the share requiring reverts fell by over 40%, although Anthropic does not isolate any single review control’s effect. That is process evidence, not a universal guarantee.

As the Claude Code team put it:

We didn't remove review. But we did move most of the reading to machines, optimized for human time and attention, and invested heavily in making it fast and easy to undo mistakes.

I once preferred a signed-versus-unsigned filter. Efficient, yes. Also wrong: it confused willingness to use one compliance framework with model quality and product resilience.

The Commission’s maintained list includes European providers Mistral AI and Aleph Alpha alongside OpenAI, Anthropic, Google, Microsoft and Amazon. Meta declined to sign. Because the law permits alternative evidence, I treat that as a diligence trigger, not proof of breach. Signatories must deliver what they promised; unsigned providers need a persuasive alternative. Vendors that become mysteriously poetic when I mention their obligations enter a third category: expensive future meeting.

The signature starts my questions. Which models does it cover? What documentation do downstream builders receive, and how quickly does it change after a new release? For systemic-risk models, I also want the provider’s approach to evaluations and serious incidents. If an AI Office evaluation triggers restrictions or mitigations, the contract should say what happens to my service.

An orange rugged portable solid-state drive with an integrated connector stands against a deep cobalt background.

Digital sovereignty needs a working exit

Before reviewing any badge, I draw four boxes: model provider, cloud platform, my finished application and the organisation deploying it. One company may fill several, but responsibility follows what each actor supplies and controls. If my product reaches Claude through AWS, Anthropic retains model-layer obligations while AWS controls parts of the serving environment and contract chain. My company may supply the finished system; my customer may deploy it. Model-provider documentation should travel downstream so I can integrate responsibly. The file still does not complete my classification work.

This is where digital sovereignty stops being conference-stage poetry. I want the exact model version, supplying legal entity and named owner for evidence updates. Can I export prompts and logs? Can my application logic move to another model without a six-month rewrite? Then I test the migration, because an untested exit clause is procurement fan fiction. “Powered by Claude” has roughly the legal precision of “contains tomatoes.”

European cloud providers can improve jurisdictional control and buyer leverage, especially when workloads and support stay in Europe. But hosting cannot create European model capacity. An American model running in Frankfurt may meet a data-location requirement while keeping critical technology under external control. European infrastructure and credible European model alternatives belong in the same procurement conversation. Europe needs Mistral AI, Aleph Alpha and their successors competing across the Union. A fragmented home market will not create continental champions; coordinated European demand might.

Open models need the same role mapping. A genuinely free and open licence can change parts of the documentation framework when weights, architecture and usage information are public. Systemic-risk obligations still apply where relevant. “Open” on a landing page means little until I inspect the licence and identify who maintains the release.

A provider can follow the Code yet change an API, withdraw a model or alter commercial terms on Friday evening, naturally, five minutes after engineering leaves for aperitivo. My fallback model, cloud path and data export matter on that Friday.

Sovereignty begins when the switch works.

Europe can make compliance infrastructure useful

A common evidence format can make Europe’s fragmented procurement environment act more like one market. A Torino startup, Tallinn ministry and Amsterdam bank can request recognisable model documentation instead of inventing national rituals. Providers can reuse one structure across tenders and contracts. Regulators know where to find and compare evidence. The AI Office-chaired taskforce helps signatories resolve inconsistent application, while national authorities retain system-level roles. This is where stronger EU institutions work: one shared framework has more leverage than incompatible national checklists. My federalist bias is showing. Bene.

The separate transparency code has its own machinery. More than 90 signatories formed two taskforces (Marking and Detection, and Labelling) according to the Commission’s announcement on 30 September 2026. Chaired by the AI Office, they convene ad hoc technical working groups and seek expert evidence. The process aims to improve implementation and advance the state of the art in marking, detection and labels. By contrast, the GPAI Code’s Transparency and Copyright chapters help model providers meet their obligations. One structures model-provider compliance; the other coordinates how synthetic content is marked and communicated. Brussels has made us a linguistic lasagna, but the distinction matters.

The evidence gap remains enormous. We do not know whether either code will materially improve detection, interoperability or robustness in ordinary content-sharing environments. No comparative false-positive and false-negative rates exist for signatories’ marking methods after routine copying, translation, editing or re-encoding. Nor is it clear how often the AI Office will rely on Code adherence instead of alternative provider evidence. Watermark detectors may stay private or restricted, preventing downstream providers, researchers and the public from independently verifying markings. Until outsiders test these systems under messy real-world conditions, implementation claims remain claims.

Europe must also stop compliance costs becoming a hyperscaler moat. Common forms reduce bespoke legal work for smaller buyers, but smaller providers still need compute, compliance tooling and customers willing to buy European technology. Coordinated EU investment matters. Public procurement should reward portability, credible evidence and working exits, giving European companies a home market large enough to scale.

In the European Commission’s AI Continent Action Plan press release on 9 April 2025, Executive Vice-President Henna Virkkunen argued that AI sits at the centre of Europe’s competitiveness, security and technological sovereignty. She is right. Regulation is not Europe’s finish line; it is infrastructure for a real continental market.

By 2028, I expect standardised Code evidence in major European tenders. Those tenders should demand a live migration test too.

Any vendor promising digital sovereignty can bring a stopwatch.

Frequently asked questions

What does the EU AI Act Code of Practice actually do?

The General-Purpose AI Code of Practice gives model providers a voluntary route for demonstrating compliance through standardised transparency, copyright, safety and security evidence. It helps regulators and downstream buyers inspect provider processes, but it does not certify a model, classify a finished AI system or complete an application-level assessment.

Does a general-purpose AI model determine an application’s EU AI Act risk level?

EU AI Act risk levels are determined at the finished-system and use level, not automatically by the underlying general-purpose model. A hiring tool, credit scorer or medical application therefore requires its own classification. Providers address model-layer GPAI duties, while suppliers and deployers may retain separate application-level obligations.

Does signing the EU AI Act Code guarantee digital sovereignty?

No. Code adherence can improve due diligence by providing recognisable documentation, but digital sovereignty also requires exact model and legal-entity identification, exportable prompts and logs, portable application logic, fallback models, viable cloud paths and a tested migration. European hosting can improve jurisdictional control without creating European model capacity.