EU AI Office Shifts from Guidance to Formal Scrutiny as August 2 High-Risk Deadline Closes In

The European Union's AI Office has entered an enforcement posture nine months after the EU AI Act's General-Purpose AI (GPAI) model obligations took legal effect on August 2, 2025. The transition — from the consultative Code of Practice process the Office ran through late 2024 into active compliance reviews of frontier model providers — marks the first time any multilateral AI regulator has moved past voluntary guidance into structured documentation demands and model evaluations. With the next major statutory deadline, full Annex III high-risk AI system compliance, arriving on August 2, 2026, builders deploying AI in employment screening, credit scoring, healthcare triage, and biometric access control have fewer than 90 days to complete conformity assessments.

The EU AI Office, housed within the European Commission's Directorate-General for Communications Networks, Content and Technology (DG CNECT) and granted supervisory authority over GPAI model providers under Regulation (EU) 2024/1689, is the world's first purpose-built multilateral AI regulator. Its enforcement trajectory in 2026 is being tracked closely by regulators in the UK (the ICO and CMA), Canada (ISED), Brazil (ANPD), and India (MeitY) as an operational template for post-voluntary AI oversight.

What the EU AI Office Is Doing

Regulation (EU) 2024/1689 — the EU AI Act — structures GPAI obligations across two tiers. Tier 1 applies to any provider releasing a GPAI model in or into the EU: publish technical documentation, comply with EU copyright law during training, make available a sufficiently detailed summary of training data content, and implement AI-generated content marking once the relevant delegated act specifies the technical standard. Tier 2 — systemic-risk model obligations — adds adversarial testing (red-teaming), 72-hour serious incident reporting to the EU AI Office, and cybersecurity measures proportionate to the model's capabilities. Both tiers have been active law since August 2, 2025.

The Systemic-Risk Threshold: 10²⁵ FLOPs

The Act creates a rebuttable presumption of systemic risk for any GPAI model trained with cumulative computation exceeding 10²⁵ floating-point operations (FLOPs). Based on disclosed parameter counts, published research, and hardware estimates, the models presumed to breach this threshold include: OpenAI's GPT-4 (March 2023), GPT-4o (May 2024), and o1 (September 2024); Anthropic's Claude 3 Opus (March 2024) and Claude 3.5 Sonnet (June 2024); Google DeepMind's Gemini 1.0 Ultra (December 2023) and Gemini 1.5 Pro (February 2024); and xAI's Grok 1 in its original 314B-parameter form (open-sourced March 2024).

Mistral AI sits in contested territory. The Paris-based company argued in Code of Practice working groups that its flagship models — Mistral Large (March 2024) and Mistral Large 2 (July 2024) — fall below the FLOPs threshold. The EU AI Office can separately designate a model as systemic risk even below the compute floor, based on capabilities, market reach, or a formal risk assessment. That designation power creates regulatory exposure for any mid-tier frontier provider whose models achieve broad deployment even without record-breaking compute.

The Open-Source Edge Case: Meta's Llama Family

Meta's release of Llama 3.1 405B on July 23, 2024 — under a commercial license restricted to entities with under 700 million monthly active users — created the Act's most complex enforcement puzzle. Article 53(2) carves out a modified regime for providers that make model weights publicly available: they retain Tier 1 documentation and copyright obligations, but receive partial relief from Tier 2 systemic-risk requirements, conditional on the EU AI Office's determination that the release is "genuinely open." The Office has not yet published a formal ruling on whether Llama 3.1 405B qualifies. That ruling, when it comes, will function as precedent for every open-weight release globally, affecting Mistral's openly released models, xAI's Grok 1, and any future frontier open-weight releases from companies including Google (Gemma) and Alibaba (Qwen).

What the Code of Practice Locked In

The GPAI Code of Practice, developed through four multi-stakeholder drafting rounds between July 2024 and early 2025, served as a transitional safe harbor: providers who participated in good faith were shielded from enforcement action while the code was being finalized. That safe harbor has since lapsed, and the code's commitments now function as the baseline against which the EU AI Office's evaluations will measure providers. Key commitments include:

• Training data transparency: Publish a summary of copyrighted material used in pretraining — detailed enough for rights-holders to assess potential infringement, but not necessarily a complete dataset manifest.
• Red-team testing: Systemic-risk models must be adversarially tested before deployment and after significant capability updates, using "sufficiently representative third parties" beyond internal safety teams.
• Incident reporting: Any "serious incident" — death, serious injury, serious harm to fundamental rights, or critical infrastructure disruption attributable to model outputs — must reach the EU AI Office within 72 hours of the provider becoming aware.
• Cybersecurity controls: Systemic-risk model providers must implement measures against model weight exfiltration, unauthorized access to training pipelines, and adversarial prompt manipulation at scale.

What Builders Must Do Right Now

If you are building on top of a commercial API — OpenAI, Anthropic, Google, Mistral — your direct GPAI obligations are limited. You are a deployer, not a provider, and the heavy documentation and red-team testing burden falls on the model company. Your obligation is not to strip required content labels, not to modify safety systems in ways that violate your API agreement, and to cooperate with any audit the provider requires for its own compliance. This is not zero work, but it is manageable.

The August 2, 2026 deadline changes the calculus for deployers. If your product makes or substantially influences a consequential decision in any of the eight Annex III categories, you are a high-risk AI system deployer subject to the full conformity regime — regardless of which underlying model you use:

• Employment: CV screening, interview scoring, worker performance monitoring, automated task allocation
• Credit and financial services: creditworthiness assessment, insurance risk classification, loan origination
• Education: student evaluation, admissions scoring, professional qualification assessments
• Law enforcement: risk scoring of individuals, predictive crime tools, biometric surveillance in public spaces
• Healthcare: diagnostic, triage, or treatment recommendation systems
• Critical infrastructure: AI-driven grid management, water treatment, or transport control
• Migration and border control: asylum application risk scoring, visa assessment tools
• Administration of justice: legal research tools that materially influence judicial or arbitration decisions

The Conformity Assessment Sprint: Six Steps Before August 2

1. Technical documentation per Annex IV: system architecture description, training data provenance, accuracy and robustness metrics, and a risk assessment specific to the deployment context
2. Conformity assessment: self-assessment suffices for most Annex III categories; mandatory third-party audit by an EU-designated notified body is required for real-time biometric identification and law-enforcement use cases
3. EU Declaration of Conformity: a signed legal declaration that the system meets all applicable Act requirements — this document must be retained for 10 years
4. Registration in the EU's AI Act database, operated by the European Commission, before placing the system on the EU market
5. Post-market monitoring plan: documented ongoing data collection and periodic review of real-world system performance against the original technical documentation
6. Human oversight measures: technical and operational controls allowing a qualified human operator to override, intervene, or shut down the system — logs of override events must be retained

Penalty Structure and Scale

The EU AI Act's penalty tiers are calibrated against global annual turnover, not EU-specific revenue, which means the exposure figures for frontier AI companies are large:

• Violations of prohibited AI practices (in force since February 2, 2025 — covers real-time remote biometric surveillance in public spaces, AI-based social scoring, manipulation of vulnerable individuals): up to €35 million or 7% of global annual turnover, whichever is higher
• GPAI and high-risk AI system non-compliance: up to €15 million or 3% of global annual turnover
• Providing incorrect or misleading information to the EU AI Office or national authorities: up to €7.5 million or 1.5% of global annual turnover

Broader Regulatory Context: FTC and State Moves

The EU AI Office's shift toward enforcement is not happening in a vacuum. The U.S. Federal Trade Commission's September 2024 Operation AI Comply — five simultaneous enforcement actions against companies making false or unsubstantiated AI performance claims, including a $193,000 settlement with DoNotPay (styled as the "world's first robot lawyer") — established that AI-capability misrepresentation is prosecutable deception under Section 5. Colorado's SB 205, signed by Governor Jared Polis on May 17, 2024 and effective February 1, 2026, created the US's first comprehensive state-level AI deployer obligation: impact assessments, algorithmic discrimination protections, and consumer notice rights for high-stakes automated decisions. US builders facing the EU's August 2 deadline are simultaneously inside Colorado's enforcement window — the two regimes share enough structural DNA that a single conformity program can often satisfy both.

What Comes Next

The EU AI Office has signaled that Q2 2026 will bring formal documentation requests to the major systemic-risk model providers, with evaluations potentially including direct model testing by independent technical experts retained by the Office under Article 92. Builders deploying high-risk AI systems have until August 2 to complete conformity assessments; national market surveillance authorities across the 27 member states will begin active enforcement sweeps shortly after. The window for voluntary remediation without enforcement exposure is now measured in weeks.