Skip to content
Rafael Patron

Note Notes 2026

GPAI Code of Practice: what OpenAI, Anthropic and Google signed up to

Translated from the Italian original.

In July 2025 the EU AI Office published the final version of the Code of Practice for general-purpose AI models, after twelve months of consultation with providers, academic experts and civil society. Thirteen companies signed up immediately, two took time, and one opted out in a conspicuously public way. Meta’s missing signature triggered a chain of assessments in the compliance teams of European companies using Llama in production. The operational point, for those who follow more than the controversy, is to understand what changes for downstream deployers.

The document matters. But the real implications are less reported than the press releases suggested.

What the Code of Practice is and why it exists

To frame the competitive positioning of providers, the analysis of automatic routing between models is useful. Article 56 of the AI Act provides that the Commission encourage codes of practice to facilitate the concrete application of GPAI obligations. The rationale is simple: Articles 53 and 55 set general principles (transparency on training data, systemic risk assessment, incident reporting), but concrete technical implementation requires operational detail that the law cannot anticipate.

The Code of Practice is that operational detail. It defines standard formats for technical documentation, templates for training data summaries, methodologies for assessing systemic risks, incident reporting processes. For providers that adhere to it there is a presumption of conformity: those who comply with the Code are considered compliant with the corresponding AI Act obligations, unless proven otherwise.

For providers that do not adhere, the framework does not formally change. They must still comply with Articles 53 and 55, but they must demonstrate conformity through alternative means. The documentary and evidentiary burden rises, regulatory risk exposure grows.

Who signed and under which chapter

The Code is structured in three chapters corresponding to three families of obligations. Transparency and copyright, applicable to all GPAI. Safety and security for models with systemic risk. Governance and internal processes.

The evolution of the AI market under pressure from new players further complicates the choice of supplier. Signed up to all three chapters: OpenAI, Anthropic, Google DeepMind, Microsoft (for proprietary models), Amazon, Cohere, IBM, Mistral, Adobe (for Firefly as a substantial part). Aleph Alpha and Cohere adhered to all chapters with interpretative clauses on specific technical points. Apple signed up in 2026 after initially taking time.

Meta publicly refused to join. Elon Musk’s xAI did not join. Meta’s refusal was justified by Mark Zuckerberg as “regulatory overreach incompatible with open-source innovation.” xAI’s refusal had no formal public justification, but it is consistent with the company’s general line of moving more freely on safety and governance issues.

What changes for those using Llama in production

The concrete question for Italian companies is this. Continuing to use Llama (Meta’s open-weight model) has compliance implications. Let’s look at them.

First, the obligations of those deploying Llama do not formally change. The Italian company remains a deployer and has its obligations regardless of the Code. Second, the basic documentation provided by Meta is considered insufficient by part of the compliance industry: it does not follow the Code of Practice templates and lacks some sections the Code requires of signatory providers. Third, in case of an inspection, the Italian company using Llama will have to document more, on its own, to cover the gaps.

In operational terms: using a model from a Code signatory reduces the deployer’s compliance cost. Using a non-signatory model increases it. It is not a legal bar, it is a preference for documentary efficiency. For an Italian SME with limited compliance resources, choosing a signatory model may make economic sense even if the non-signatory model’s performance is comparable.

Transparency on training data: the most interesting section

The transparency chapter of the Code requires providers to publish sufficiently detailed summaries of the data used for training. The template defined at the end of 2025 includes categories of sources (web scraping, open datasets, licensed content, synthetic data, human feedback data), approximate quantities per category, main languages covered, presence of applied filters (removal of illegal content, personal data, copyrighted content).

The first summaries published in early 2026 by OpenAI, Anthropic and Google are documents of 30-80 pages each. They provide useful information without revealing trade secrets. For Italian publishers they have become working material: SIAE, FIEG and publishing groups are using this information to set up structured opt-out or licensing requests.

For those who want to check whether their site was used for training, the main tool is the robots.txt file updated with specific directives for AI crawlers (OpenAI’s GPTBot, Anthropic’s ClaudeBot, Google’s Google-Extended). The Code requires providers to respect these directives at least for crawls after the opt-out, but the historical picture (what has already been scraped) remains complex.

Safety and systemic risks: the most technical chapter

The second chapter applies only to GPAI with systemic risk (above 10^25 cumulative FLOPS). It defines categories of systemic risks to be assessed: misuse capabilities (CBRN, cyber, manipulation), loss-of-control risks, market concentration risks. For each category it defines assessment methodologies, severity thresholds, mitigation processes.

These are the systems providers call frontier models. GPT-5 and later, Claude 3 Opus and later, Gemini Ultra, Llama 4 (if Meta were to recognize it) are within the perimeter. Signatory providers commit to independent red team testing, stress scenarios on dangerous capabilities, structured dialogue with the AI Office on results and mitigations.

For Italian companies using these models, the implications are indirect but substantial. A model that has passed independent red team tests is documentarily more solid. A model on which these tests have not been done, or not publicized, leaves more gray areas in the deployer’s compliance documentation. In early 2026 the AIPIA community published a technical note for members operating in regulated sectors on how to integrate this information into their own risk assessment.

The Italian blind spot: a Code Application Plan is needed

Among the European policy topics little discussed in Italy is the application dimension of the Code. Adhering to the Code is one thing. Applying it in practice with documented processes, internal controls and independent audits is another. The Commission is working on a system for monitoring application that will become operational in 2027.

For Italian deployer companies the smart move is to start asking their suppliers not only “have you joined the Code” but “can you show us your Code Application documentation”. Serious providers already have materials ready. Less serious ones rely on formal adherence without investing in operational application. Only those who ask the right question will see the difference.

The cascade effect on enterprise contracts

A recurring clause in contracts circulating since 2026 asks the AI system supplier to declare whether the models used come from providers adhering to the Code of Practice and to commit to notifying the customer of any change of supplier. These are clauses that the more structured companies are putting into standard tender specifications, no longer just as an option.

For those developing AI products in Italy, the practical consequence is that transparency about one’s own technology stack becomes a commercial element. Changing the underlying model is no longer just a technical decision, it is a change that affects enterprise contracts. This creates stability in the market but also rigidity for software houses.

What to do by the end of 2026

Four operational actions for those operating in Italy with products that integrate GPAI. Map which models are in use and whether the providers are Code signatories. For models from non-signatory providers, build more solid autonomous compliance documentation. Update contracts with customers and suppliers to reflect the new transparency requirements. Prepare standard answers to the compliance questions that will inevitably come from enterprise customers.

The Code of Practice is not a law. It does, however, have the capacity to standardize market expectations faster than primary legislation does. Italian companies that make it a working tool have a competitive advantage; those that treat it as external PR remain exposed to the next waves of compliance requests.

In the same section

The whole section

Advisory, talks and contact

Contact Rafael

Advisory, speaking and contact requests are handled on rafaelpatron.it.