EU AI Act GPAI Enforcement and Fines: How the Commission Polices Model Providers

A Centralised Enforcement Model

For most of Regulation (EU) 2024/1689, enforcement is decentralised: national market surveillance authorities police AI systems in their territories. For general-purpose AI models the architecture is deliberately different. Article 88 confers on the Commission exclusive powers to supervise and enforce Chapter V, exercised through the AI Office. The reasoning is structural: GPAI models are placed on a single Union market, their providers are often global, and fragmented national supervision would produce inconsistent demands on identical artefacts. One model, one supervisor.

This design has a practical consequence providers sometimes miss: national authorities cannot fine you for model-level Chapter V breaches, but they remain fully competent for AI systems built on your model — including your own products. A model provider that also ships applications answers to the AI Office for the model and to national authorities for the systems, in parallel.

The Supervisory Toolkit: Articles 91 to 93

Information requests (Article 91)

The AI Office may request from a provider the documentation drawn up under Articles 53 and 55 — Annex XI files, evaluation records, incident logs — and any additional information needed to assess compliance. Requests carry deadlines; supplying incorrect, incomplete or misleading information is itself a fining ground.

Model evaluations (Article 92)

Where information proves insufficient, or following a qualified alert from the scientific panel under Article 90, the Commission may conduct evaluations of the model itself — including through independent experts — to assess compliance or investigate Union-level systemic risks. Evaluations can include requests for model access through appropriate means, structured to protect trade secrets. An evaluation is the GPAI equivalent of a regulatory inspection: external experts reproducing your safety claims against your own model.

Requests for measures (Article 93)

Where warranted, the Commission may request providers to take appropriate measures to comply with their obligations, to implement mitigation measures where an evaluation has raised serious and substantiated concerns of systemic risk, or to restrict the making available of the model, withdraw it, or recall it. Measures escalate; recall of a model distributed as open weights is acknowledged to be partly irreversible, which is precisely why supervision concentrates on prevention and documentation.

The Fines: Article 101

Article 101 empowers the Commission to impose fines on GPAI model providers not exceeding 3 percent of annual total worldwide turnover in the preceding financial year or 15 million euros, whichever is higher, where the provider intentionally or negligently: infringes the relevant provisions of the regulation — the Article 53 and 55 obligations themselves; fails to comply with an information request under Article 91 or supplies incorrect, incomplete or misleading information; fails to comply with a measure requested under Article 93; or fails to make available access to the model for an evaluation under Article 92. Fines must be effective, proportionate and dissuasive, taking into account the nature, gravity and duration of the infringement, and providers receive procedural protections: the right to be heard before adverse decisions, reasoned decisions, and review by the Court of Justice of the European Union, which has unlimited jurisdiction to cancel, reduce or increase fines.

The timing is critical: although Chapter V obligations have applied since August 2, 2025, Article 113 defers the application of Article 101, so the Commission's fining power over GPAI providers applies from August 2, 2026. The first year was supervision without monetary sanction; from August 2026 the regime is complete.

How GPAI Fines Compare

Context clarifies severity. Under Article 99, which applies to AI systems: prohibited practices attract up to 35 million euros or 7 percent of worldwide turnover; breaches of most operator obligations, including high-risk and transparency duties, up to 15 million euros or 3 percent; supplying misleading information to authorities, up to 7.5 million euros or 1 percent. The GPAI ceiling thus matches the mid-tier system penalty — but with two aggravating features: it is levied by a single well-resourced supervisor rather than a patchwork of national authorities, and for frontier-scale providers 3 percent of worldwide turnover is a number measured in billions, not millions.

What Triggers Enforcement in Practice

The realistic pathways into an enforcement file, in rough order of likelihood: rightsholder complaints about missing or inadequate training data summaries and copyright policies — the most visible, public-facing obligations; downstream provider complaints about absent Annex XII documentation, often surfacing through high-risk conformity assessments that cannot be completed; scientific panel qualified alerts grounded in published capability research or incidents; serious incidents reaching the AI Office through Article 55 reporting or press coverage; and inconsistencies between public claims and filings — a marketing page boasting frontier capabilities alongside a filing arguing the model is modest invites questions. Notification failures under Article 52 — crossing the 10^25 FLOPs threshold silently — are uniquely dangerous because compute budgets leak: cloud providers, investors and papers all reveal scale.

Procedure and Timing Realities

Enforcement under Article 101 follows Commission administrative procedure familiar from competition law: investigation, a statement of the Commission's preliminary findings, the provider's right to be heard in writing and orally, then a reasoned decision subject to full review by the Court of Justice. Two timing realities shape strategy. First, investigations are slow but their subject matter is fixed early: the documentation state on the day of the first Article 91 request defines the file, and improvements afterwards count as mitigation, not absolution. Second, cooperation is priced in: the gravity-and-duration calculus rewards providers who respond completely and on time, and punishes the drip-feed of partial answers that turns a documentation gap into a misleading-information ground. The Commission has also signalled, through the Code of Practice process, that signatories acting in good faith within agreed timelines can expect proportionality in how the first enforcement cycles are run — a posture, not a promise, but one with institutional weight behind it.

A Concrete Example

A provider launches a capable model in early 2026 without a training data summary, reasoning that enforcement is soft. A publishers' association files a complaint citing the missing summary and suspected disregard of opt-outs. The AI Office sends an Article 91 request for the Annex XI file and copyright policy with a six-week deadline. The provider's reconstruction effort produces documentation that contradicts an engineering blog post about dataset composition. The file now contains three potential grounds — the Article 53(1)(d) breach, an inadequate copyright policy, and arguably misleading information in the response. After being heard, and with the breach period extending past August 2, 2026, the provider negotiates commitments: a conformant summary, crawler changes, third-party verification — and absorbs a fine calibrated to negligence rather than intent. Every element of that file was avoidable at a cost measured in engineer-weeks.

Reducing Your Enforcement Surface

• Publish the visible artefacts first: the training data summary and a copyright contact point eliminate the most common complaint grounds.
• Treat Article 91 readiness as an SLA: a standing internal commitment to deliver the documentation set within weeks, rehearsed annually.
• Never let public statements and filings diverge; route capability claims through the same review as regulatory documents.
• If the model approaches 10^25 FLOPs, file the Article 52 notification on the planning-stage trigger, not at launch.
• Consider Code of Practice adherence: signatories work from shared templates and a cooperative supervisory posture, which shapes both the probability and the tone of enforcement.
• Document remediation of every known gap with dates — negligence assessments turn on what you knew and when you acted.

Action Plan

Build the enforcement scenario into your compliance programme rather than treating it as a remote tail risk: identify which of your obligations are complaint-exposed, rehearse the information request, and keep the documentation coherent across public and regulatory audiences. The Commission built a supervision machine with long levers and a single point of aim; the providers it will fine first are not those with imperfect models, but those with absent paperwork, missed notifications and contradictory stories.

For boards and audit committees, the governance translation is simple: GPAI enforcement exposure is now a quantifiable operational risk with a known supervisor, known triggers and a known ceiling, and it belongs on the risk register with an owner, a readiness metric and an annual rehearsal — exactly as data protection enforcement was absorbed after 2018. Companies that made that institutional move early under the GDPR spent less, litigated less and recovered faster; the same curve is available here, and it starts with treating the AI Office as a counterparty you prepare for rather than an abstraction you hope to avoid.

This article is for informational purposes only and does not constitute legal advice. Regulatory requirements change frequently — verify current rules with official sources. Built by Sawai Gyoseishoshi Office, Hiroshima, Japan.