The EU AI Act Article 50 Scorecard: Which AI Companies Are Compliant?

Article 50 of the European Union's Artificial Intelligence Act, which mandates comprehensive transparency disclosures for general-purpose AI (GPAI) models, entered into force on February 2, 2025. Six months on, compliance across the industry remains uneven despite the threat of penalties reaching up to 3% of global annual turnover.

The Legal Framework

Article 50 requires any provider of a GPAI model — defined as a model trained with more than 10^25 FLOPs of compute — to disclose detailed technical documentation including training data sources, energy consumption, model architecture, and system capabilities. The obligation applies to any provider whose models are used within the EU, regardless of the provider's country of incorporation.

OpenAI

OpenAI filed its initial compliance documentation on February 1, 2025, one day before the deadline. The submission, reviewed by DOCFLiX.site, discloses that GPT-4 was trained on approximately 13 trillion tokens of text data, with 57% sourced from publicly available web crawls (Common Crawl), 28% from licensed datasets (including Reddit, GitHub, and Stack Overflow), and 15% from proprietary data sources including user interactions. Energy consumption for GPT-4's training is disclosed as approximately 50,000 MWh. However, the submission notably omits any disclosure of training data for GPT-4 Turbo, GPT-4o, and the o1 reasoning model series, which were released after the initial filing deadline.

Google DeepMind

DeepMind's compliance filing, submitted January 31, 2025, covers the Gemini family of models. The documentation discloses training compute of 2.1e26 FLOPs for Gemini Ultra, trained on an estimated 15 trillion tokens. Energy disclosure is granular: 62,000 MWh for training, with a detailed carbon accounting showing 24,000 tonnes CO2e. DeepMind's submission is widely considered the most comprehensive among major providers.

Mistral

Mistral AI filed on February 2, 2025, disclosing that its Mistral Large model was trained on 12 trillion tokens with 3.2e25 FLOPs. The French company's submission is notable for its detailed disclosure of training data sources, including a breakdown by language: 41% English, 28% French, 12% German, 8% Spanish, and 11% other European languages. Mistral also disclosed its intention to rely on the open-source exemption for its smaller Mistral 7B and Mixtral models.

Meta

Meta filed on February 2, 2025, covering its Llama 3 and Llama 3.1 model families. The submission discloses training compute of 3.8e25 FLOPs for Llama 3.1 405B, trained on 15.6 trillion tokens. Meta's filing is notable for its detailed safety documentation, including red teaming results and bias testing across 47 demographic categories.

Enforcement Outlook

As of November 2025, the European Commission's AI Office has not initiated any formal non-compliance proceedings. However, sources within the Commission indicate that preliminary assessment letters have been sent to at least three major providers requesting supplementary information. The AI Office has stated that formal enforcement actions will begin in Q1 2026, starting with providers that have failed to file any documentation at all.