EU launches voluntary AI code as Meta refuses to sign and Google warns on innovation

The European Commission's voluntary Code of Practice for general-purpose AI models came into force on Saturday, marking the first concrete step in the bloc's attempt to regulate the most powerful artificial intelligence systems before the full AI Act takes effect in 2026. The code sets out expectations for companies building models such as those behind ChatGPT and Google's Gemini: explain how the systems work, disclose training data sources, assess risks including bias and misinformation, and help users understand the technology's operation. Compliance is not legally binding, but the Commission intends the code to shape industry behaviour ahead of mandatory rules.

What the code requires

The Code of Practice, published by the Commission in late July after months of drafting with industry and civil society, translates the AI Act's high-level obligations for general-purpose AI into operational commitments. Signatories agree to publish summaries of training data, document testing procedures, implement risk mitigation for systemic risks such as large-scale disinformation or cybersecurity threats, and establish governance structures for ongoing compliance. The code also asks model providers to support downstream users, companies that fine-tune or integrate the models, with sufficient technical documentation to meet their own regulatory obligations. The Commission has said it will monitor adoption and may adjust the code before the AI Act's GPAI provisions become enforceable.

The AI Act itself, adopted by the European Parliament and Council in March 2024, entered into force in August 2024 with a phased implementation schedule. Rules on prohibited practices applied from February 2025. Obligations for high-risk AI systems in sectors such as medical devices, recruitment and critical infrastructure take effect in August 2026. The general-purpose AI provisions, including requirements for model evaluation and systemic risk assessment, are scheduled for August 2027. The voluntary code is designed to bridge the gap.

Google signs but sounds a warning

Google announced on Tuesday that it would sign the code, joining OpenAI and the French startup Mistral AI, which had already committed. Kent Walker, Google's president of global affairs, said the company would "join several other companies, including US model providers, in signing." But he coupled the commitment with a pointed critique: "Over-regulation risks slowing Europe's development and deployment of AI." Walker argued that departures from existing EU copyright law, steps that slow approvals, or requirements that expose trade secrets "could chill European model development and deployment." The company has been lobbying for the code to align closely with the AI Act text and to avoid creating parallel compliance tracks.

Google's position reflects a broader calculation. The company operates Gemini, one of the world's most advanced general-purpose models, and has significant research presence in Europe, including DeepMind's London headquarters. Compliance costs are manageable for a firm of Google's scale; the greater risk is regulatory fragmentation that could complicate global product rollouts. By signing early, Google positions itself as a constructive partner while preserving leverage to shape the code's evolution before it hardens into law.

Meta refuses, citing legal uncertainty

Meta has taken a different path. Joel Kaplan, the company's chief global affairs officer, wrote in a LinkedIn post that Meta would not sign. "This code introduces a number of legal uncertainties for model developers, as well as measures which go far beyond the scope of the AI Act," he said. Meta argues that the code's training data transparency requirements, risk assessment methodologies and downstream documentation obligations exceed what the legislation mandates. The company also objects to what it sees as an open-ended commitment to mitigate systemic risks that are difficult to define or measure.

The refusal continues a pattern of confrontation between Meta and Brussels. In recent months the company pulled all political advertising from its European platforms rather than comply with new rules on online campaigning transparency under the Regulation on the transparency and targeting of political advertising. Meta has also challenged aspects of the Digital Markets Act and Digital Services Act. The company's Llama models are open-weight, not fully open source, a distinction that complicates its positioning under the AI Act's provisions for free and open-source licences. By refusing the voluntary code, Meta signals it will litigate or lobby rather than pre-comply.

European industry pushes back

The split among US tech giants is mirrored by unease among Europe's own industrial champions. Earlier this month, a group including Airbus, Lufthansa, Siemens and SAP wrote to the Commission urging a pause in AI regulatory implementation. They warned that the bloc risks falling behind in the global race for technology leadership, particularly against the United States and China, where state-backed investment and lighter regulatory touch are accelerating deployment. The letter argued that uncertainty over compliance requirements is already delaying investment decisions in AI-driven manufacturing, logistics and maintenance.

Airbus, for instance, uses AI for predictive maintenance across its commercial fleet and for design optimisation in new aircraft programmes. Lufthansa applies machine learning to crew scheduling, fuel efficiency and customer service automation. Both companies say they support the principle of trustworthy AI but contend that the current regulatory trajectory, multiple overlapping acts, evolving codes, national transposition differences, creates a compliance burden disproportionate to the risk profile of their use cases. They have asked for a "regulatory pause" until the full implementation framework is settled.

The Commission holds firm

Despite the pushback, the Commission has shown no sign of slowing. A spokesperson said the voluntary compliance period beginning this weekend demonstrates that "Brussels plans to press ahead with its approach to AI governance." The strategy is deliberate: by establishing a de facto standard now, the Commission shapes the market before the legally binding rules arrive. Companies that engage early influence the interpretative guidance that national regulators will follow. Those that stay out risk being measured against a benchmark they did not help write.

This approach mirrors the Commission's playbook with the General Data Protection Regulation, where early adopters set the compliance tone and late movers faced enforcement actions. It also reflects a geopolitical calculation. The EU wants its AI Act to become a global reference standard, the so-called Brussels effect. A credible, functioning regime before the United States enacts federal legislation, currently stalled in Congress, strengthens that position. The Commission is also aware that the incoming US administration may pressure Europe to soften its stance on US tech firms.

Why this matters

How we got here

What happens next