Will the EU AI Office formally classify any frontier model as "systemic risk" during 2026?

The EU has increasingly focused on regulating AI, particularly with the AI Act set to be fully implemented by 2026. While the European Commission has indicated a cautionary approach towards labeling AI systems as "systemic risk," recent statements suggest a growing acknowledgment of the potential societal impacts of frontier models, evidenced by their classification frameworks and the emphasis on high-risk categories. However, actual classifications may depend heavily on ongoing evaluations and consensus among EU member states regarding what constitutes systemic risk.

The EU AI Act establishes a clear framework for identifying and regulating high-risk AI systems, with a specific category for "systemic risk" models. Given the rapid advancement and widespread potential impact of frontier models, it's highly probable that the EU AI Office will identify at least one such model as systemic by 2026 to enforce the Act's stringent requirements. Current discussions and the preparatory work by the AI Office indicate a proactive approach to managing these powerful AI systems.

The EU AI Act's Article 4(4) establishes "systemic risk" classification mechanisms, but formal designations require sustained evidence of real-world harm and regulatory consensus. As of late 2024, no frontier model has been formally classified despite significant regulatory scrutiny of models like GPT-4 and Claude—suggesting high evidentiary thresholds exist. The EU AI Office, operational since 2024, has focused on compliance monitoring rather than systemic designations. For 2026 classification to occur, a frontier model would need to demonstrate documented systemic harms (e.g., widespread infrastructure failures, election interference, or severe financial system disruption) substantial enough to trigger coordinated EU action within approximately 12 months. Current frontier models show capability risks rather than demonstrated systemic impact, and regulatory bodies historically move cautiously on novel designations.

The EU AI Office must first finalize the thresholds in the implementing act (expected Q1 2026) and then complete model evaluations under Article 51; only four labs currently publish ≥10^26 FLOP training runs, and none have disclosed public evidence meeting the 10^25 cumulative-compute systemic-risk criterion set out in the December 2023 trilogue text. The Commission’s March 2024 impact assessment projected the first classification decision no earlier than late 2026, giving a compressed nine-month window after the Office becomes operational in August 2025.