New op-ed on 'Too much Silicon Valley, not enough Europe? The new GPAI Code shows light and shadow'

Yesterday, the European Commission published the long-awaited Code of Practice for General Purpose AI (GPAI) models – a key element for implementing the AI Act. Having followed this process closely from the legislative negotiations until now, I concluded in my piece for Süddeutsche Zeitung Dossier that the Code is not a breakthrough but that it is a necessary step forward – though one with significant design flaws.

Yesterday, the European Commission published the long-awaited Code of Practice for General Purpose AI (GPAI) models – a key element for implementing the AI Act. Having followed this process closely from the legislative negotiations until now, I concluded in my piece for Süddeutsche Zeitung Dossier that the Code is not a breakthrough but that it is a necessary step forward – though one with significant design flaws. Read it here.

👉 The drafting process was imbalanced: many EU companies and civil society actors had limited opportunity to shape the outcome. Instead, a small group of largely US-based GPAI developers dominated the room – and their influence is clearly visible in the final text.

👉 On the positive side: The Code mandates independent testing for systemic GPAI models, introduces greater transparency through published model and safety summaries, and – importantly – stays within the legal framework of the AI Act. These are meaningful wins, made possible by sustained pressure from a broad alliance of MEPs, researchers, civil society, and parts of industry.

👉 But challenges remain: key concepts like fine-tuning, systemic risk, and model lifecycle are still vaguely defined. Downstream providers may not receive enough information to meet their own AI Act obligations. And there is still no clarity on what happens when major GPAI providers simply refuse to sign the Code or do not follow its guidance.

What needs to happen next: The AI Office must be properly staffed and empowered, starting to monitor the Code's uptake. It must also finalise the GPAI guidelines without delay. For the Codes' next versions (⚠️ yes, there need to be regular updates + you should think of it as 3 separate codes: on transparency, copyright, and safety & security) EU AI developers and deployers need a real seat at the table – not just a symbolic one. Overall, this version of the GPAI Code must be just a beginning, not the final word.

🇺🇸 Check also my quote for POLITICO's US newsletter: "Kai Zenner, digital advisor to European Parliament member Axel Voss of Germany, told POLITICO’s Daniella Cheslow that the guidance “isn’t looking that bad.” He said civil society groups and lawmakers pushed the European commission to resist pressure from tech companies, but that those advocates are likely to have more fights on their hands in the coming years." (LINK)