US tech giants sign EU AI code as Meta holds out

Brussels has secured a quiet but significant victory in its effort to shape the rules governing artificial intelligence. By the close of business on Friday, the developers of the world's most powerful general-purpose AI models, Google, Microsoft and OpenAI among them, had all signed the European Commission's voluntary General-Purpose AI Code of Practice. The deadline was August 2, 2025, the first major milestone in the phased rollout of the EU AI Act, which becomes fully applicable on the same date next year.

A voluntary tool with mandatory consequences

The code is not legally binding. It is a compliance framework designed to help companies demonstrate they meet the AI Act's requirements for general-purpose models, systems such as GPT-4, Gemini and Llama that can generate text, images, video and code. But the Commission made clear that companies choosing not to sign would face heightened scrutiny from the newly created AI Office. They would still have to comply with the Act's baseline obligations, but without the procedural clarity and dialogue the code provides.

That pressure appears to have worked. Of the companies in scope, only Meta declined to participate. Elon Musk's xAI signed the main text but refused the chapter dealing with copyright, calling it a clear overreach. The result is a near-uniform front among US tech leaders, something few in Brussels expected when the drafting process began nearly a year ago.

The drafting process and the transatlantic shift

The AI Act, adopted in 2024, left the technical details for general-purpose models to a group of 13 academics. Their task was to translate broad legal principles into practical measures: how companies should disclose training data, comply with EU copyright law, and mitigate systemic risks such as loss-of-control scenarios. The drafting took place against a backdrop of intense lobbying from industry, civil society and lawmakers, each watching to ensure the guidance did not drift beyond the legislation.

Midway through, the political environment shifted. The new US administration publicly urged European governments to adopt light-touch AI regulation. That emboldened companies to criticise early drafts. In February, Google's Kent Walker described the voluntary rules as a step in the wrong direction, while Meta's Joel Kaplan warned they went beyond what the law required. For a moment, it looked as though the code might collapse under the weight of transatlantic friction.

Google's pivot and the art of the possible

By this week, the tone had changed. Walker confirmed Google would sign, stating that the final version comes closer to supporting Europe's innovation and economic goals than when it began. The shift reflects a calculation: the code is now the only structured channel for influencing how the AI Office interprets the Act's vague provisions on transparency, copyright and systemic risk. Walker said Google wants to work with the Commission to ensure the code remains proportionate and responsive to the rapid and dynamic evolution of AI.

Microsoft's Nanna-Louise Linde struck a similar note. She said the company signed to demonstrate long-standing compliance with EU law and to contribute ideas for the code's improvement. OpenAI, which does not maintain a large Brussels lobbying operation, issued a shorter statement saying the decision reflected its commitment to providing capable, accessible and secure AI models to Europeans. All three companies gain a seat at the table for the code's first review cycle.

Meta's calculation: confrontation over cooperation

Meta took a different view. Kaplan, writing on LinkedIn, slammed the EU's inconsistent, restrictive and self-defeating approach to AI. The company argues the code imposes obligations, particularly on training data disclosure, that exceed the AI Act's text. Meta has also signalled it will not participate in the voluntary labelling scheme for AI-generated content, a separate but related Commission initiative. The refusal is a gamble: Meta bets that its open-source Llama models give it enough developer goodwill and regulatory leverage to negotiate bilaterally rather than through the code's collective framework.

The Commission's warning on scrutiny is not empty. The AI Office, led by Lucilla Sioli, has the power to request documentation, conduct evaluations and, ultimately, impose fines of up to 3% of global annual turnover for non-compliance with the Act's general-purpose model provisions. Meta will now face that apparatus without the procedural safe harbour the code provides.

Copyright: the faultline that remains

Even among signatories, the copyright chapter is contested. The code requires companies to disclose a sufficiently detailed summary of the content used to train their models. Rights-holders, publishers, authors, collecting societies, see this as the first step toward remuneration claims. Walker warned against departures from EU copyright law and requirements that expose trade secrets. xAI's refusal to sign that chapter signals a broader industry anxiety: that transparency obligations will become a discovery tool for litigation.

The Commission has acknowledged the tension. It plans a review of the EU copyright directive by mid-2026, explicitly to address the intersection with generative AI. That review will reopen the legislative fight, with tech companies arguing that text-and-data-mining exceptions already cover training, while creative industries push for a new remuneration right. The code's copyright provisions are, in effect, a holding position until that battle is resolved.

Why this matters

How we got here

What happens next