EuroAIGuard shows what voluntary transparency looks like before the rules arrive

What EuroAIGuard disclosed

EuroAIGuard Conseil, a French organisation developing advanced AI systems, has published a transparency report through the OECD's Hiroshima AI Reporting Framework.^[1] The report documents the organisation's AI systems, their capabilities, governance measures and risk management practices.

The Hiroshima framework, agreed by G7 members in 2023 and maintained by the OECD, provides a voluntary structure for organisations to disclose information about their AI systems.^[2] It covers system design, capabilities, limitations, testing procedures, safety measures and governance arrangements. Organisations can choose which sections to complete and how much detail to provide.

EuroAIGuard's report includes information about system architecture, data sources, intended uses, known limitations and the organisation's approach to monitoring and updating its systems. The level of detail varies across sections, reflecting the voluntary nature of the framework.

The difference between voluntary and mandatory

Voluntary transparency, like EuroAIGuard's report, allows organisations to demonstrate good practice and build trust with users and stakeholders. It can also serve as preparation for incoming mandatory requirements.

Under the EU AI Act, certain transparency obligations are not voluntary. From 2 August 2026, providers must ensure that people are informed when they interact with an AI system, and that AI-generated content is marked in a machine-readable format.^[3] Deployers must disclose when images, audio or video are deepfakes, and when AI-generated text published to inform the public has not been reviewed and endorsed by a natural person.

For high-risk AI systems under Annex III, broader obligations apply from 2 December 2027, including technical documentation, automatic logging and conformity assessment.^[4] The transparency requirements under Article 50 are separate and apply regardless of whether a system is classified as high-risk.

The Hiroshima framework does not create legal obligations, but it demonstrates what structured disclosure can look like. Organisations using the framework to prepare may find it easier to meet mandatory requirements when they arrive.

What the framework covers

The Hiroshima AI Reporting Framework is organised into sections covering system design, capabilities, limitations, testing, safety measures and governance.^[2] Organisations can select which sections apply to their systems and how much information to provide in each.

Key areas include:

• System description: the AI model, its architecture, training data and intended purpose
• Capabilities and limitations: what the system can do and where it is known to fall short
• Testing and evaluation: how the system was tested and what results were observed
• Safety and security: measures to prevent misuse, errors or harmful outputs
• Governance: who oversees the system, how decisions are made and how risks are managed

The framework is designed for flexibility. A small organisation deploying a third-party model will complete different sections to a large provider training its own foundation model. The OECD publishes reports on its AI policy observatory platform, where they can be searched and compared.^[1]

The evidence question

Transparency reports document what an organisation says about its AI systems at a point in time. They do not, by themselves, prove what the system actually did in a given instance.

When a decision, output or interaction is later questioned, the evidence that matters is the record of what happened. A transparency report can describe governance processes and testing results. A log sealed at the time of the event shows what the system did.

The two serve different purposes. Transparency builds understanding and trust in general terms. Evidence answers specific questions: what decision was made, what data was used, which version of the model was running.

For organisations subject to the AI Act's logging requirements from December 2027, the question is not only what to log, but how to show that a log has not been altered since the event it describes. Article 12 specifies what must be logged over a high-risk system's lifetime.^[4] It does not specify how to protect those logs, but their evidential value depends on being able to show they have not changed.

What this means for you

If you are preparing for the AI Act's transparency requirements, voluntary frameworks like Hiroshima can help you structure your documentation early. They are not a substitute for compliance, but they can make the work easier when mandatory deadlines arrive.

Questions to consider:

• Have you identified which of your AI systems will need to meet Article 50's transparency obligations from August 2026?
• For systems that communicate with people or generate content, do you have a process to inform users they are interacting with AI?
• If you deploy high-risk systems under Annex III, have you mapped the technical documentation and logging requirements that apply from December 2027?
• Can you show, after the fact, what your system did on a given day, and that the record has not been altered?

The last question applies whether or not you publish a transparency report. When a decision is challenged, the organisation with its own unchanged record starts in a stronger position than the one relying on memory or editable files.

What to watch

2 August 2026: Article 50 transparency requirements for AI that communicates with people or generates content apply across the EU.

2 December 2027: High-risk requirements under Annex III, including logging and technical documentation, apply to new systems placed on the market.