Public Transparency for Frontier AI
Europe’s opportunity to raise the global floor before it’s too late
Authors: Aimen Taimur and Jimmy Farrell
Executive Summary
Public scrutiny is one of the few mechanisms capable of holding the most advanced AI models to account. As general-purpose AI models are deployed at a pace that outstrips both regulatory cycles and company internal review, public transparency becomes a matter of democratic accountability. Public Transparency allows harms to be identified early, strengthens trus, and reduces burden falling on the AI Office alone. Realising these benefits, however, depends on the level of public transparency built into the EU regulation, in particular the AI Act’s General Purpose-AI Code of Practice.
Examining the current text of the Code of Practice, public transparency is treated as a discretionary practice rather than a default. Under Measure 1.2 of the Transparency Chapter, publication of model cards is only encouraged, while Measure 10.2 of the Safety and Security Chapter only requires summarised documentation.
Set against the wider legal and regulatory environment, this level of public transparency is weak. The openness principles of the AI Act and the EU Treaties point towards broader disclosure. Adjacent EU regimes already publish comparable documentation, from clinical trial data at the European Medicines Agency to the general public right of access to EU documents. Across the Atlantic, California, New York and Illinois are now introducing mandatory transparency regimes for frontier AI.
Public transparency is not all bad for leading frontier AI providers. Legitimate concerns exist for frontier AI providers, such as the protection of trade secrets, prevention of harmful misuse and reputational risks. However, numerous benefits include improved positioning in liability proceedings, better insurance underwriting, reduced compliance friction, earlier identification of risks, and increased public trust.
Public transparency of frontier AI risks is increasingly threatened, especially in the US. This has been made clear by CAISI’s June 2026 suspension of model assessment publication and the December 2025 court filing from Elon Musk’s xAI against California’s AB 2013 training data transparency law. As securitisation continues, with frontier AI models increasingly only deployed internally, the public transparency gap will only widen further. A high floor in the EU is therefore essential.
Key recommendations:
Establishing public transparency of frontier model documentation as the default position under the Code. As an immediate step requiring no amendment, the AI Office should clarify that whatever is removed from public documentation under Measure 10.2 must have item-level justification, meeting the standard already established in EU precedent, in particular under EMA Policy 0070.
In Code of Practice updates, removing the conditional qualifier 'if and insofar as necessary' in Measure 10.2, setting minimum content requirements, and replacing the encouragement in Measure 1.2 with a positive obligation to publish Annex XII items by default.
Issuing standardised AI Office templates for the public versions of Frameworks and Reports, comparable to the training-data summary template under Article 53(1)(d), which the AI Office can pursue under its existing authority to keep documentation up to date with market and technological developments.
Setting out a three-tier access architecture, so that full documentation reaches the AI Office and the Scientific Panel, intermediate detail reaches downstream providers on request, and a public version with a justification log reaches the wider public.
Clarifying that public transparency forms part of adequate systemic-risk management under Article 55 of the AI Act, engaging the AI Office's supervisory powers under Articles 88 to 94 and the penalty regime under Article 101 if necessary.
Adding a public-disclosure sub-paragraph to Article 55 over the longer term, through the Act’s review under Article 112(3) due by 2nd August 2029.
Strengthening the public transparency of frontier AI models under the Code of Practice and the AI Act is an essential part of the EU’s ongoing endeavour to steer this technological revolution towards societal benefit.
Read the full paper:
