Google and Meta reject EU draft code for advanced AI models

The European Union's attempt to translate its landmark AI Act into practical rules for the most powerful models is faltering before it has begun. Google and Meta, two of the handful of companies that actually build frontier systems, have declared the draft code of practice unworkable and signalled they may refuse to sign. The intervention lands days before the AI Action Summit in Paris, where the Commission hopes to demonstrate that Europe remains open for business even as the new US administration brands EU digital regulation a disguised tariff.

What the code of practice is meant to do

The AI Act, adopted in summer 2024, sets out obligations for providers of general-purpose AI models, the category that includes OpenAI's GPT series, Anthropic's Claude, Google's Gemini and Meta's Llama. Because the legislation lays down principles rather than technical specifications, the Commission convened a drafting process to produce a voluntary code of practice. Companies that sign up would gain a presumption of compliance; those that stay out face the full weight of regulatory scrutiny without a safe harbour.

The draft covers four main areas: transparency about training data, copyright compliance, systemic risk assessment and mitigation, and governance including third-party testing. The first version circulated among stakeholders in late 2024; a revised text is expected in March with finalisation targeted for April. The Commission has insisted the code must not create new obligations beyond the Act, a point that is now the centre of the dispute.

Google argues the draft exceeds the law

Kent Walker, Google's president of global affairs, told Politico that the code is a "step in the wrong direction" at a moment when Europe is trying to restore competitiveness. He argued that several requirements, particularly on copyright and third-party model testing, go beyond the AI Act's text, duplicate existing frameworks or impose disproportionate burdens. Walker stopped short of ruling out signature, saying it was "too soon to tell" and hinting that the Paris summit could be a pivotal moment for negotiations.

Google's position reflects a broader industry calculation. The company has invested heavily in Gemini and its cloud AI infrastructure; a code that mandates detailed disclosure of training data or opens models to external auditors could expose proprietary techniques and create liability vectors that the Act itself does not require. Walker's language, "too soon to tell", is deliberately ambiguous, leaving room for a last-minute compromise while signalling that the current text is unacceptable.

Meta draws a harder line

Meta's Joel Kaplan was more direct. In a video interview with a Brussels audience he said the code established "unworkable and technically unfeasible requirements" and went "beyond the requirements" of the AI Act. He indicated the company would not sign the code in its current form. Meta's Llama models are released with open weights, a distribution model that complicates several code provisions, particularly those around downstream control and third-party testing of deployed systems.

Kaplan's refusal to commit is significant because Meta has positioned itself as the leading advocate for open-source AI in Europe. If the company that most benefits from a permissive regulatory environment walks away, the code's claim to represent industry consensus collapses. The Commission would then face a choice: weaken the text to secure signatures, or proceed with a code that only smaller or more risk-averse providers endorse.

The Paris summit as leverage

The AI Action Summit on 10-11 February brings together heads of state, tech executives and civil society. Ursula von der Leyen and Henna Virkkunen, the executive vice-president for technological sovereignty, want to use the event to showcase European AI ambition and counter the narrative that the EU is a regulatory graveyard. The presence of Google and Meta at the summit gives both companies direct access to the political principals who ultimately own the code process.

Walker's remark that Paris could be a "pivotal moment" suggests Google expects the political level to intervene if the technical drafting does not yield concessions. Virkkunen's cabinet has been closely involved in the code negotiations; a political signal from the summit that the EU values investment over procedural purity would strengthen the companies' hand. Conversely, a united front from member states backing the current draft would call the bluff.

Transatlantic pressure compounds the problem

The criticism arrives as the Trump administration escalates its campaign against EU digital rules. The US president has described fines and enforcement actions against American platforms as equivalent to tariffs, and his trade team has signalled that digital regulation will feature in any future trade talks. That rhetoric emboldens companies to resist European rules publicly, calculating that Washington will treat regulatory retaliation as a trade dispute.

European officials privately acknowledge the dynamic but insist the code process is independent of geopolitics. The Commission's argument is that the AI Act was co-legislated by the European Parliament and the Council, representing member states, and that the code merely operationalises democratic law. Yet the timing, two US giants coordinating pushback days before a summit designed to attract US investment, makes the political context impossible to ignore.

What the code actually requires

The disputed provisions centre on three areas. First, training data transparency: the draft asks for a "sufficiently detailed summary" of datasets used, which companies argue could reveal competitive secrets and expose them to copyright litigation. Second, systemic risk: providers of models deemed to pose systemic risk, a threshold likely met by GPT-4, Gemini Ultra and Llama 3.5, must conduct adversarial testing, incident reporting and mitigation planning. Third, third-party evaluation: the code envisions independent auditors assessing compliance, a requirement that does not exist in the Act's text but appears in the draft as a governance measure.

Copyright is the most legally fraught. The AI Act requires providers to put in place a policy to respect EU copyright law and to publish a summary of training content. The draft code elaborates this into specific disclosure templates and opt-out mechanisms for rightsholders. Google and Meta contend that the Act's wording leaves the method to the provider, and that the code's prescriptive templates constitute new law by stealth.

The open-source complication

Meta's Llama models are released with open weights, meaning anyone can download, fine-tune and deploy them. The code of practice assumes a provider-deployer relationship where the model maker controls access and can monitor use. Open weights break that assumption: once released, the provider cannot enforce downstream compliance, cannot run third-party tests on every deployment, and cannot reliably report incidents it does not see. Kaplan has argued that the code fails to account for this distribution model, effectively penalising openness.

This is not a theoretical concern. The EU has funded open-source AI initiatives and publicly praised the transparency benefits of open weights. If the code makes open release legally hazardous, European researchers and startups that rely on Llama or similar models lose a critical resource. The Commission's drafters are aware of the tension but have not yet produced a carve-out that satisfies both safety advocates and open-source proponents.

Why the code matters even if voluntary

How we got here

What happens next