EU AI Act Art.93: AI Office Interim Measures for GPAI Systemic Risk — Urgency Procedure, Commission Decision, and Provider Response Guide (2026)

EU AI Act Article 93 is the AI Office's emergency power in the GPAI enforcement toolkit. Where Art.90 is a documentary investigation that unfolds over weeks and Art.91 is a physical or remote inspection that requires a Commission decision and advance notice, Art.93 is the provision that authorises the AI Office to request the Commission to act immediately — suspending, restricting, or requiring corrective action for a GPAI model with systemic risk before the full enforcement cycle is complete.

The urgency threshold is high: Art.93 requires that a GPAI model with systemic risk presents an imminent, serious risk to health, safety, or fundamental rights, and that the pace of the normal enforcement procedure is itself a threat — that waiting for Art.90 and Art.91 to run their course would allow harm that cannot be undone. But for providers of systemic-risk GPAI models, that threshold is not theoretical. A GPAI model that generates outputs enabling large-scale fraud, that introduces identifiable security vulnerabilities at scale, or that causes demonstrable discrimination across downstream applications at Union-wide reach could meet the Art.93 urgency standard — particularly if the Scientific Panel has already issued a qualified alert under Art.55.

For GPAI model developers, Art.93 is the provision that makes infrastructure location, response-time commitments, and corrective-action readiness operationally relevant. A provider who receives an Art.93 Commission decision has no grace period for document gathering or technical assessment: interim measures apply immediately, penalties under Art.99(4)(e) for non-compliance accrue from the decision date, and the provider's right to be heard operates under a compressed timeline adapted from the Art.89 procedure.

Art.93 became applicable on 2 August 2025, concurrent with all GPAI obligations in Chapter V.

Art.93 in the GPAI Enforcement Chain

Art.93 is the emergency layer of the GPAI regulatory stack, sitting beyond both the documentary investigation (Art.90) and the access-based inspection (Art.91):

Art.93 can be triggered without completing the Art.90 and Art.91 stages. It is not a sequential escalation that requires Art.91 inspection first — it is an independent urgency pathway that can be initiated when the AI Office has sufficient grounds to conclude that immediate action is necessary, regardless of where the investigation currently sits.

Art.93(1): Trigger Conditions — When the AI Office Can Act

Art.93(1) sets the threshold conditions that authorise the AI Office to request the Commission to adopt interim measures. Two triggering pathways exist.

Pathway 1: Art.55 Scientific Panel Qualified Alert

The most structured pathway to Art.93 is through the Scientific Panel under Art.55. When the Panel issues a qualified alert — its formal determination that a GPAI model with systemic risk presents a specific, credible risk requiring regulatory response — the AI Office may use that qualified opinion as the evidentiary basis for an Art.93 interim measures request to the Commission.

The Art.55 → Art.93 pathway is significant because it involves structured scientific assessment. A qualified alert from the Scientific Panel means that independent experts with access to technical evaluations of the model have concluded that:

• The model's systemic risk classification under Art.51 is confirmed or strengthened
• The risk is of a character that warrants enforcement response, not just monitoring
• The urgency condition is met — that the harm pathway is active or imminent, not speculative

For providers, this means that ongoing cooperation with Scientific Panel evaluations under Art.55 is not merely a compliance formality. Panel findings that cite a provider's model in connection with systemic risk are a direct precursor to Art.93 authority — and providers who have participated transparently in Art.55 evaluations, documented corrective measures proactively, and maintained updated technical documentation will be in a materially different position when a qualified alert is issued.

Pathway 2: AI Office Own Initiative

The second pathway is the AI Office acting on its own initiative, based on information gathered through Art.90 information requests, Art.91 inspections, Art.92 market monitoring, or other available sources. This pathway requires the AI Office to have, independently, sufficient grounds to conclude that:

• A GPAI model with systemic risk poses an imminent serious risk to health, safety, or fundamental rights
• The gravity of the risk justifies emergency measures before the standard enforcement procedure concludes
• The measures requested are proportionate to the nature and scale of the identified risk

The AI Office's own-initiative pathway reflects the enforcement reality that not every serious risk situation arrives via the Scientific Panel. Market monitoring under Art.92 may identify a sudden capability development, a coordinated misuse pattern, or a downstream deployment causing harm at scale. Art.90 documentary review may reveal compliance failures that create acute risk. Art.91 inspection may uncover technical characteristics that the provider failed to disclose. In each case, the AI Office can proceed to request Art.93 interim measures without waiting for a qualified Panel alert.

Art.93(2): Types of Interim Measures

The Commission decision under Art.93 may impose interim measures from a defined set of enforcement tools. These are protective measures, not punitive ones: their purpose is to prevent ongoing harm while the full enforcement procedure runs its course.

Requiring Corrective Action

The Commission may order the provider to take specific corrective measures with respect to the GPAI model. Corrective action orders are the least restrictive interim measure and are preferred where targeted fixes can address the risk without broader market impact.

Corrective action orders under Art.93 may require:

• Immediate adjustment or disabling of specific model capabilities identified as risk sources
• Addition of safety guardrails or refusal mechanisms for identified harm categories
• Modification of the model's terms of service to restrict downstream deployment contexts
• Implementation of monitoring mechanisms to track specific output patterns
• Provision of additional documentation or technical disclosures to the AI Office within compressed timelines

The proportionality principle governs corrective action orders: the measures must be calibrated to the specific risk identified, not used as a vehicle for comprehensive model redesign. A provider who receives a corrective action order and implements the required measures in good faith will have demonstrated compliance even if the full Art.90/91 investigation later uncovers broader issues.

Access Restriction

The Commission may restrict access to the GPAI model — including restricting access by specific user categories, geographic markets, application domains, or deployment contexts. Access restriction is a graduated measure that targets the risk pathway while preserving model availability for contexts where the risk does not materialise.

Access restriction orders are particularly relevant for GPAI models where the systemic risk arises not from the model's general capabilities but from specific deployment patterns. A GPAI model used broadly in creative and research contexts may present systemic risk specifically when deployed in financial fraud, disinformation generation, or critical infrastructure management. An Art.93 access restriction can target the harmful deployment context while leaving non-harmful use cases operational.

For providers with API access models, access restriction means implementing technical enforcement of the Commission decision: blocking API access by use case category, imposing user verification requirements that gate access to restricted contexts, or modifying the model's deployment terms to exclude restricted applications. The Commission decision will specify the restriction scope; providers are responsible for implementing it with the technical precision the decision requires.

Temporary Market Withdrawal

The most severe interim measure is temporary withdrawal of the GPAI model from the Union market — suspension of all placing on market, making available, or use within the EU pending the outcome of the full enforcement procedure.

Market withdrawal orders are reserved for situations where the identified risk cannot be contained by corrective action or access restriction — where the model's core characteristics create systemic risk that pervades its available use cases. Given the downstream integration complexity of widely-deployed GPAI models (API integrations, downstream system dependencies, third-party applications built on the model), market withdrawal has cascading effects that the Commission must assess as part of its proportionality evaluation.

For providers, temporary market withdrawal is the Art.93 scenario requiring the most extensive advance preparation: maintaining clear documentation of downstream integrators, having an established communication protocol for suspension events, and having technically prepared for a withdrawal implementation that can be executed rapidly. A provider who receives a market withdrawal decision and takes weeks to implement it faces ongoing Art.99(4)(e) penalty exposure for every day of continued market presence during the interim period.

The Commission Decision: Process and Timeline

Art.93 interim measures require a Commission decision. Unlike Art.90 information requests (which the AI Office can issue on its own authority), Art.93 interim measures require the Commission to act — because they directly restrict or suspend the provider's market activity, which is an enforcement action with significant legal and economic consequences.

The Commission decision process under Art.93 operates under urgency constraints, but is not instantaneous. The AI Office's request to the Commission must include:

• The factual basis for the urgency finding — why the risk is imminent and why the standard enforcement timeline is insufficient
• The evidentiary basis — the Art.55 qualified alert, Art.90/91 findings, or Art.92 monitoring findings that support the systemic risk assessment
• The proposed interim measures — what the AI Office considers necessary and proportionate
• The duration — how long interim measures should remain in force before the Commission reassesses

The Commission evaluates the AI Office's request against the urgency threshold and proportionality requirements. Where the Commission decides to act, the decision specifies the interim measures, their duration, and any conditions or monitoring requirements that accompany them.

Provider Rights: The Art.89 Urgency Adaptation

Art.93 does not eliminate the provider's right to be heard under Art.89. But the standard Art.89 procedure — which provides for a meaningful preparation period before enforcement action — is adapted under Art.93 to fit the urgency context.

Under the Art.93 urgency adaptation, the provider's right to be heard is implemented with compressed timelines. The AI Office will notify the provider of the interim measures request before the Commission decision where feasible, but the urgency finding means that the opportunity to respond is time-constrained. In the most acute scenarios, the Commission may adopt interim measures and notify the provider simultaneously, with the right to be heard operating as a post-decision review right rather than a pre-decision consultation.

For providers, the practical implication is that Art.89 preparation should not be deferred to the moment of an Art.93 notification. A provider who has already documented:

• Its understanding of why its model does and does not meet the Art.51 systemic risk criteria
• The corrective measures it has proactively implemented to address identified risk categories
• Its technical capacity to implement access restriction or market withdrawal rapidly
• Its downstream integrator notification protocols

…will be in a substantially stronger position in an Art.93 urgency procedure than a provider who begins that documentation only when the AI Office contacts it.

Art.91 vs Art.93: Enforcement Distinction

Art.91 and Art.93 are both provider-specific enforcement tools, both requiring Commission decisions, both applicable to GPAI models with systemic risk. The distinction is operational:

The most significant operational distinction is that Art.91 generates information and Art.93 restricts activity. A provider can cooperate fully with an Art.91 inspection while continuing to operate its GPAI model normally during the investigation period. An Art.93 Commission decision applies immediately — the provider cannot continue market activity that the decision restricts while the decision stands.

Art.93 and Art.92 Interaction: Intelligence to Emergency

Art.93 and Art.92 represent opposite ends of the AI Office's GPAI enforcement function — Art.92 is the market-wide intelligence layer and Art.93 is the emergency response layer. The two interact in a structured way.

Art.92 market monitoring identifies patterns and developments that feed the AI Office's ongoing risk assessment of the GPAI market. When monitoring findings identify a specific provider's model as presenting emerging systemic risk, those findings can inform an AI Office decision to initiate Art.90 information requests or, in urgent cases, to request Art.93 interim measures directly.

For providers, this interaction means that Art.92 compliance — transparency in Art.53(3) notifications, participation in codes of practice under Art.56, proactive disclosure of capability updates — is not only a monitoring-layer obligation. It is also the primary mechanism through which the AI Office develops its Art.93 risk assessment for individual providers. A provider who is transparent with the AI Office's monitoring function is less likely to generate the kind of information gap that triggers an urgency finding; a provider who is opaque is more likely to have its model surface in the Art.92 market monitoring as a systemic risk concern that requires immediate action.

CLOUD Act Exposure Under Art.93

Art.93 interim measures create a specific CLOUD Act conflict scenario for GPAI providers whose model infrastructure sits on US cloud services or within US jurisdiction.

When an Art.93 Commission decision requires market withdrawal or access restriction of a GPAI model, the provider must implement that restriction across all deployment infrastructure — including infrastructure that may be subject to US jurisdiction. If US law enforcement has served, or subsequently serves, a CLOUD Act demand requiring the provider to maintain access or cooperation, the provider faces an irreconcilable conflict: EU law requires restriction of access, US law may require maintenance of access.

EU-resident model infrastructure under Art.93 interim measures reduces but does not eliminate this exposure. Where the provider maintains model weights, training data, and API routing within EU jurisdiction, an Art.93 interim measure requiring access restriction is enforceable under EU law without requiring the provider to act contrary to US legal obligations. This is the structural argument for EU-hosted GPAI infrastructure as a systemic risk management strategy.

Penalty Exposure: Art.99(4)(e)

Non-compliance with an Art.93 Commission decision — whether by continuing restricted market activity, failing to implement required corrective measures, or obstructing the interim measures regime — triggers penalty exposure under Art.99(4)(e).

The Art.99(4)(e) penalty structure applies to non-compliance with Commission decisions:

• Maximum fine: EUR 3,000,000 or 1% of total worldwide annual turnover in the preceding financial year, whichever is higher
• Daily periodic penalty payments: Up to 1.5% of total worldwide annual turnover — these continue for each day of continued non-compliance after the Commission decision

For GPAI providers with significant global turnover, the 1.5%/day periodic penalty is the operationally material risk. A provider who takes ten days to implement a market withdrawal decision faces periodic penalty exposure of 15% of global annual turnover — separate from any base fine. This makes immediate implementation the only rational commercial response to an Art.93 Commission decision, regardless of whether the provider intends to challenge the decision before the CJEU.

The Art.99(4)(e) penalties for Art.93 non-compliance are equal in maximum severity to the penalties for obstruction of an Art.91 inspection. The penalty structure signals that the Commission treats non-compliance with interim protective measures as equal in gravity to obstructing an investigation — both are viewed as active evasion of the regulatory enforcement function.

Python: Art93InterimMeasureTracker

from dataclasses import dataclass, field
from datetime import date, timedelta
from enum import Enum
from typing import Optional


class InterimMeasureType(Enum):
    CORRECTIVE_ACTION = "corrective_action"
    ACCESS_RESTRICTION = "access_restriction"
    MARKET_WITHDRAWAL = "market_withdrawal"


class TriggerPathway(Enum):
    SCIENTIFIC_PANEL_ALERT = "art55_scientific_panel_qualified_alert"
    AI_OFFICE_OWN_INITIATIVE = "ai_office_own_initiative"
    ART90_ESCALATION = "art90_information_request_escalation"
    ART91_ESCALATION = "art91_inspection_escalation"
    ART92_MONITORING = "art92_market_monitoring_finding"


class ComplianceStatus(Enum):
    NOT_TRIGGERED = "not_triggered"
    AI_OFFICE_REQUEST_PENDING = "ai_office_request_to_commission_pending"
    COMMISSION_DECISION_PENDING = "commission_decision_pending"
    MEASURES_ACTIVE = "interim_measures_active"
    MEASURES_IMPLEMENTED = "provider_measures_implemented"
    MEASURES_LIFTED = "interim_measures_lifted"
    CJEU_CHALLENGED = "cjeu_challenge_pending"


@dataclass
class InterimMeasureDecision:
    decision_date: date
    measure_type: InterimMeasureType
    trigger_pathway: TriggerPathway
    scope_description: str
    duration_days: int
    corrective_actions_required: list[str] = field(default_factory=list)
    restriction_scope: Optional[str] = None
    status: ComplianceStatus = ComplianceStatus.COMMISSION_DECISION_PENDING

    @property
    def expiry_date(self) -> date:
        return self.decision_date + timedelta(days=self.duration_days)

    @property
    def daily_penalty_rate_pct(self) -> float:
        return 1.5

    def days_to_expiry(self, as_of: date = None) -> int:
        ref = as_of or date.today()
        return (self.expiry_date - ref).days


@dataclass
class Art93InterimMeasureTracker:
    provider_name: str
    model_name: str
    systemic_risk_classification: bool
    eu_infrastructure_pct: float  # 0-100: % of model infra in EU jurisdiction
    decisions: list[InterimMeasureDecision] = field(default_factory=list)
    corrective_actions_completed: list[str] = field(default_factory=list)

    def add_decision(self, decision: InterimMeasureDecision) -> None:
        self.decisions.append(decision)

    def active_decisions(self, as_of: date = None) -> list[InterimMeasureDecision]:
        ref = as_of or date.today()
        return [
            d for d in self.decisions
            if d.status == ComplianceStatus.MEASURES_ACTIVE
            and d.decision_date <= ref <= d.expiry_date
        ]

    def penalty_exposure_assessment(self) -> dict:
        active = self.active_decisions()
        if not active:
            return {"status": "no_active_decisions", "periodic_penalty_risk": "none"}

        most_severe = max(
            active,
            key=lambda d: [
                InterimMeasureType.MARKET_WITHDRAWAL,
                InterimMeasureType.ACCESS_RESTRICTION,
                InterimMeasureType.CORRECTIVE_ACTION,
            ].index(d.measure_type)
            if d.measure_type in [
                InterimMeasureType.MARKET_WITHDRAWAL,
                InterimMeasureType.ACCESS_RESTRICTION,
                InterimMeasureType.CORRECTIVE_ACTION,
            ] else 0
        )

        pending_actions = [
            a for a in most_severe.corrective_actions_required
            if a not in self.corrective_actions_completed
        ]

        return {
            "active_decisions": len(active),
            "most_severe_measure": most_severe.measure_type.value,
            "days_to_expiry": most_severe.days_to_expiry(),
            "pending_corrective_actions": pending_actions,
            "implementation_status": most_severe.status.value,
            "periodic_penalty_rate": f"{most_severe.daily_penalty_rate_pct}%/day of global turnover",
            "cloud_act_risk": "HIGH" if self.eu_infrastructure_pct < 50 else "MEDIUM" if self.eu_infrastructure_pct < 80 else "LOW",
            "max_base_fine": "EUR 3M or 1% global annual turnover (higher)",
        }

    def urgency_readiness_score(self) -> dict:
        score = 0
        checks = []

        if self.systemic_risk_classification:
            checks.append(("Systemic risk classified — Art.93 exposure confirmed", False, 0))
        else:
            checks.append(("No systemic risk classification — Art.93 scope limited", True, 15))
            score += 15

        if self.eu_infrastructure_pct >= 80:
            checks.append(("≥80% EU infrastructure — CLOUD Act conflict risk low", True, 20))
            score += 20
        elif self.eu_infrastructure_pct >= 50:
            checks.append(("50-79% EU infrastructure — CLOUD Act risk moderate", True, 10))
            score += 10
        else:
            checks.append(("<50% EU infrastructure — CLOUD Act conflict risk HIGH", False, 0))

        if self.corrective_actions_completed:
            checks.append(("Corrective action protocol exists", True, 20))
            score += 20
        else:
            checks.append(("No corrective action protocol documented", False, 0))

        return {
            "readiness_score": score,
            "max_score": 55,
            "readiness_pct": round((score / 55) * 100, 1),
            "checks": checks,
        }

    def interim_measure_summary(self) -> str:
        active = self.active_decisions()
        lines = [
            f"Art.93 Interim Measure Tracker — {self.provider_name} / {self.model_name}",
            f"Systemic Risk: {'YES' if self.systemic_risk_classification else 'NO'}",
            f"EU Infrastructure: {self.eu_infrastructure_pct}%",
            f"Active Decisions: {len(active)}",
            f"Total Decisions: {len(self.decisions)}",
        ]
        if active:
            for d in active:
                lines.append(
                    f"  [{d.measure_type.value}] expires {d.expiry_date} "
                    f"({d.days_to_expiry()}d) — {d.status.value}"
                )
        readiness = self.urgency_readiness_score()
        lines.append(f"Urgency Readiness: {readiness['readiness_pct']}%")
        return "\n".join(lines)

15-Item Art.93 Urgency Preparedness Checklist

Classification and Monitoring

1. Systemic risk classification confirmed — Verify whether your GPAI model meets or approaches the Art.51 FLOP threshold or capability-based systemic risk criteria; Art.93 scope is limited to systemic-risk models.
2. Art.55 Scientific Panel engagement tracked — Monitor Scientific Panel publications and qualified alerts affecting your model or comparable models; a qualified alert is the primary Art.93 trigger.
3. Art.92 monitoring reports reviewed quarterly — AI Office market monitoring reports may contain findings that indicate systemic risk concerns for your model before any direct enforcement contact.
4. Art.90 response history documented — Prior Art.90 information requests and your responses are the evidentiary record the AI Office uses to assess urgency; document completeness is protective.

Infrastructure and Implementation Readiness

1. EU infrastructure percentage quantified — Calculate and document what percentage of your model's training, inference, and API infrastructure sits within EU jurisdiction; this is the CLOUD Act risk baseline for Art.93 scenario planning.
2. Market withdrawal implementation plan exists — Draft a technical protocol for implementing EU market withdrawal within 24 hours of a Commission decision, including API shutdown, downstream integrator notification, and user communication.
3. Access restriction capability deployed — Ensure your infrastructure can implement granular access restrictions by user category, geographic market, and application context without full market withdrawal; this is the intermediate Art.93 response.
4. Corrective action library maintained — Maintain a documented catalogue of corrective measures that can be activated rapidly: guardrail additions, capability restrictions, output filtering — these are the Art.93 corrective action options you can offer proactively.

Legal and Procedural Readiness

1. Art.89 urgency procedure reviewed — Your legal team should have reviewed how the Art.89 right-to-be-heard procedure applies under Art.93 urgency conditions; compressed timelines mean preparation must precede the notification, not follow it.
2. CJEU challenge pathway assessed — Understand the conditions under which challenging a Commission Art.93 decision before the CJEU is viable; note that challenge filing does not automatically suspend the interim measures.
3. Penalty calculator maintained — Know your global annual turnover and the Art.99(4)(e) penalty exposure: EUR 3M base cap or 1% global turnover, plus 1.5%/day for continued non-compliance.
4. Authorised representative briefed — If you use an EU authorised representative under Art.54, ensure they are briefed on Art.93 urgency procedures and have authority to respond on the provider's behalf within compressed timelines.

Downstream and Communication

1. Downstream integrator notification protocol — If your GPAI model is accessed via API by downstream providers or deployers, maintain a current list and a notification protocol for Art.93-triggered access changes.
2. Art.53(3) notification current — Ensure your Art.53(3) systemic risk notification to the AI Office is current and accurately reflects your model's capability profile; stale notifications create information gaps that increase urgency-finding risk.
3. Incident response team identified — Designate a cross-functional team (legal, technical, communications, executive) who can convene within hours of an Art.93 notification and execute the provider's response protocol without coordination delays.

Art.90 → Art.91 → Art.93: Enforcement Escalation Scenarios

Art.93 is not always preceded by Art.90 and Art.91. The three enforcement tools operate in parallel as much as in sequence, depending on the urgency and nature of the identified risk.

Sequential escalation (most common): Art.90 information request reveals incomplete disclosure → Art.91 inspection confirms systemic risk characteristics → Art.92 monitoring finds downstream harm at scale → Art.93 interim measures requested to prevent further harm during ongoing investigation.

Parallel urgency pathway (acute risk): Art.92 monitoring identifies sudden capability development posing imminent risk → Art.55 Scientific Panel issues qualified alert → Art.93 interim measures requested without prior Art.90 or Art.91 action.

Own-initiative urgency (rare but structurally available): AI Office receives third-party information (market surveillance NCA report, downstream deployer incident notification, civil society alert) indicating imminent serious risk → AI Office assesses sufficiency of information → Art.93 request filed based on available evidence.

For GPAI providers, the parallel urgency pathway is the most difficult to prepare for: it can arise without the warning signals of an Art.90 information request or Art.91 inspection notice. The only structural preparation is ensuring that the Art.92 monitoring intelligence function does not generate surprises — that the provider's compliance posture, disclosed to the AI Office through Art.53(3) notifications and CoP participation, accurately reflects the model's actual capabilities and risk profile.

See Also

• Art.92: AI Office Monitoring of GPAI Market Developments — Reporting, Public Disclosure, and International Standards Developer Guide — Art.92 market monitoring is the intelligence function that feeds Art.93 urgency findings; monitoring reports can identify systemic risk patterns that trigger Art.93 without prior Art.90/91 escalation.
• Art.91: AI Office Inspections of GPAI Providers — On-Site and Remote Access Developer Guide — Art.91 inspections are the investigative tool that precedes or runs parallel to Art.93 interim measures; inspection findings can accelerate an Art.93 urgency determination.
• Art.90: AI Office Information Requests to GPAI Providers — Developer Response Guide — Art.90 information requests are the documentary stage that may precede Art.93; incomplete Art.90 responses increase the likelihood of AI Office urgency assessment.
• Art.55: AI Office Evaluation of GPAI Models and Scientific Panel Powers — Art.55 Scientific Panel qualified alerts are the primary structured trigger for Art.93 interim measures; Panel findings are the evidentiary basis for AI Office urgency requests.
• Art.89: Right to Be Heard in GPAI Enforcement Proceedings — Developer Response Guide — Art.89 right-to-be-heard applies to Art.93 urgency procedures in adapted form; understanding the standard procedure is prerequisite to understanding the urgency adaptation.
• Art.94: AI Office Commitments and Settlement Decisions for GPAI — Provider Exit Strategy and Binding Decision Framework — Art.94 is the cooperative alternative to Art.93 emergency measures; providers who engage the commitment pathway before Art.93 is invoked retain control over the terms of enforcement resolution.