General-Purpose AI Model Compliance Standard

Document Type: Standard Standard ID: STD-AI-018 Standard Title: General-Purpose AI Model Compliance Standard Version: 1.0 Effective Date: 2025-08-02 Next Review Date: 2026-08-02 Review Frequency: Annually or upon regulatory change Parent Policy: POL-AI-001 - Artificial Intelligence Policy Owner: AI Act Program Manager Approved By: AI Governance Committee Chair Status: Draft Classification: Internal Use Only

---

TABLE OF CONTENTS

1. Document History
2. Objective
3. Scope and Applicability
4. Control Standard
5. Supporting Procedures
6. Compliance
7. Roles and Responsibilities
8. Exceptions
9. Enforcement
10. Key Performance Indicators (KPIs)
11. Training Requirements
12. Definitions
13. Link with AI Act and ISO42001

---

DOCUMENT HISTORY

Version	Date	Author	Changes	Approval Date	Approved By
0.1	2025-07-15	AI Act Program Manager	Initial draft	-	-
0.2	2025-07-25	AI Act Program Manager	Added systemic risk controls and open-source exemptions	-	-
0.3	2025-07-30	AI Act Program Manager	Incorporated Legal and stakeholder feedback	-	-
1.0	2025-08-02	AI Act Program Manager	Final version approved - GRC restructured	2025-08-01	Jane Doe, AI Governance Committee Chair

---

OBJECTIVE

This standard defines requirements for providers of general-purpose AI (GPAI) models under EU AI Act Articles 51-56, covering technical documentation, downstream provider information, copyright compliance, training data summaries, and additional obligations for GPAI models with systemic risk including model evaluation, adversarial testing, incident reporting, and cybersecurity.

Primary Goals:

• Ensure complete Annex XI technical documentation for all GPAI models
• Provide downstream providers with Annex XII information enabling understanding of model capabilities and limitations
• Implement copyright compliance and publish training data summaries
• Classify GPAI models for systemic risk and notify the European Commission
• Conduct model evaluations and adversarial testing for systemic risk models
• Implement incident reporting and cybersecurity for systemic risk models

---

SCOPE AND APPLICABILITY

2.1 Mandatory Applicability

This standard is mandatory for:

• All GPAI models provided or made available on the Union market
• GPAI models classified as presenting systemic risk (Art. 51)
• GPAI model providers acting within the EU or whose models are used in the EU

2.2 Open-Source Exemption (Art. 53(2))

Obligations under Art. 53(1)(a) (Annex XI technical documentation) and Art. 53(1)(b) (Annex XII downstream provider information) do not apply to providers of GPAI models that:

• Are released under a free and open-source licence, AND
• Have publicly available parameters, weights, architecture, and usage information

Critical exception: This open-source exemption does not apply if the model presents systemic risk under Art. 51. Systemic risk models must comply with all obligations regardless of their open-source status.

All GPAI model providers, including open-source providers, must still comply with:

• Art. 53(1)(c): Copyright compliance policy
• Art. 53(1)(d): Training data summary publication

2.3 Recommended Applicability

This standard is recommended for:

• Organisations evaluating whether their AI models qualify as GPAI models
• Downstream providers integrating GPAI models into AI systems
• Organisations developing foundation models for internal use

2.4 GPAI Model Requirements Covered

• Annex XI technical documentation (Art. 53(1)(a))
• Annex XII downstream provider information (Art. 53(1)(b))
• Copyright compliance under Directive 2019/790 (Art. 53(1)(c))
• Training data summary publication (Art. 53(1)(d))
• Systemic risk classification and Commission notification (Art. 51-52)
• Model evaluation and adversarial testing (Art. 55(1)(a)-(b))
• Serious incident reporting (Art. 55(1)(c))
• Cybersecurity protections (Art. 55(1)(d))

2.5 Out of Scope

• High-risk AI system requirements (covered by STD-AI-001 through STD-AI-013)
• AI literacy training (covered by STD-AI-014)
• Prohibited AI practices (covered by separate standard)
• GPAI models used exclusively for research and development purposes before market placement

---

CONTROL STANDARD

Control GPAI-001: GPAI Model Technical Documentation

Control ID: GPAI-001 Control Name: GPAI Model Technical Documentation Control Type: Preventive Control Frequency: Per model release, annual review Risk Level: High

Control Objective

Draw up and maintain technical documentation per Annex XI for each GPAI model, ensuring comprehensive documentation of model architecture, training process, testing methodology, evaluation results, computational resources used, and known limitations (Art. 53(1)(a)).

Open-source models with publicly available parameters, weights, architecture, and usage information are exempt from this obligation unless the model presents systemic risk (Art. 53(2)).

Control Requirements

CR-001.1: Annex XI Technical Documentation

Prepare and maintain technical documentation containing all information required under Annex XI of the EU AI Act.

Annex XI Documentation Requirements:

Documentation Element	Description	Detail Level	Update Trigger
Model Architecture	Detailed description of model architecture and design	Full technical specification	Any architectural change
Training Process	Training data sources, methodology, parameters, decisions	Comprehensive process documentation	Any training change
Testing Methodology	Testing approach, benchmarks, evaluation frameworks	Full methodology with results	Any testing change
Evaluation Results	Performance metrics, benchmark scores, capability assessments	Complete results with analysis	Per evaluation cycle
Computational Resources	Resources used for training (FLOPs, hardware, duration)	Quantified resource accounting	Per training run
Known Limitations	Known limitations, failure modes, inappropriate use cases	Comprehensive limitation analysis	Ongoing discovery

Mandatory Actions:

• Document model architecture and design decisions
• Document training process including data sources and methodology
• Document testing methodology and evaluation results
• Document computational resources used for training
• Document known limitations and appropriate use cases
• Maintain and update documentation throughout model lifecycle
• Assess open-source exemption eligibility per Art. 53(2)

CR-001.2: Documentation Maintenance

Keep documentation current and update upon material changes.

Documentation Maintenance Schedule:

Activity	Frequency	Trigger	Responsible
Comprehensive review	Annually	Calendar	AI Act Program Manager
Update on model change	Per change	Material model update	Model Development Team
Update on new evaluation	Per evaluation	New evaluation results	Model Evaluation Team
Version control	Continuous	Any documentation change	Documentation Owner

Evidence Required:

• Annex XI technical documentation
• Model architecture documentation
• Training process records
• Testing and evaluation reports
• Documentation update records
• Open-source exemption assessment (if applicable)

Audit Verification:

• Verify Annex XI documentation exists for each GPAI model
• Confirm documentation covers all required elements
• Check documentation is current and maintained
• Validate open-source exemption assessments where claimed
• Review documentation version history

---

Control GPAI-002: Downstream Provider Information

Control ID: GPAI-002 Control Name: Downstream Provider Information Control Type: Preventive Control Frequency: Per model release, upon model change Risk Level: High

Control Objective

Provide information and documentation to downstream AI system providers enabling them to understand capabilities and limitations per Annex XII (Art. 53(1)(b)). Open-source models with publicly available parameters, weights, architecture, and usage information are exempt unless the model presents systemic risk (Art. 53(2)).

Control Requirements

CR-002.1: Annex XII Information Package

Create and distribute an information package to downstream providers containing all elements required under Annex XII.

Annex XII Information Package Contents:

Information Element	Description	Purpose	Format
Model Capabilities	What the model can do, intended use cases	Enable appropriate integration	Technical specification
Model Limitations	Known limitations, failure modes, biases	Prevent misuse and inform risk assessment	Limitation report
Integration Guidance	Technical guidance for integration	Enable proper integration	Integration guide
Performance Characteristics	Performance metrics, benchmarks, accuracy	Set expectations for downstream use	Performance report
Safety Information	Safety considerations, guardrails, restrictions	Enable safe deployment	Safety documentation
Acceptable Use Policy	Permitted and prohibited uses	Clarify usage boundaries	Policy document

Mandatory Actions:

• Create downstream provider information package per Annex XII
• Document model capabilities and known limitations
• Provide integration guidance for downstream providers
• Update documentation when model changes materially
• Assess open-source exemption eligibility per Art. 53(2)

CR-002.2: Distribution and Update Management

Ensure downstream providers receive current information and are notified of material changes.

Distribution Requirements:

Activity	Timing	Method	Record
Initial package distribution	Before or at model provision	Secure delivery	Distribution log
Material change notification	Without undue delay	Direct notification	Notification record
Annual review notification	Annually	Standard communication	Review record
Version tracking	Continuous	Version control system	Version history

Evidence Required:

• Annex XII information packages
• Distribution records to downstream providers
• Model capability and limitation documentation
• Integration guidance documents
• Documentation update records

Audit Verification:

• Verify Annex XII information packages exist for each GPAI model
• Confirm distribution records to all downstream providers
• Check information packages are current and complete
• Validate update notifications sent for material changes
• Review downstream provider feedback mechanisms

---

Control GPAI-003: Copyright Compliance and Training Data Summary

Control ID: GPAI-003 Control Name: Copyright Compliance and Training Data Summary Control Type: Preventive Control Frequency: Per model release, ongoing monitoring Risk Level: High

Control Objective

Implement copyright compliance policy respecting rights reservations under Directive (EU) 2019/790 and publish a sufficiently detailed training data summary per AI Office template (Art. 53(1)(c)-(d)).

Note: Unlike GPAI-001 and GPAI-002, these obligations apply to all GPAI model providers, including open-source providers. There is no open-source exemption for copyright compliance or training data summary requirements.

Control Requirements

CR-003.1: Copyright Compliance Policy

Establish and implement a policy to comply with Union copyright law, in particular with respect to rights reservations expressed pursuant to Article 4(3) of Directive (EU) 2019/790.

Copyright Compliance Requirements:

Requirement	Description	Implementation	Verification
Copyright Policy	Formal policy for copyright compliance in training	Written policy approved by Legal	Annual review
Rights Reservation Identification	Process to identify opt-out reservations	Automated and manual screening	Per data acquisition
Opt-Out Compliance	Respect opt-out reservations from rights holders	Exclusion from training data	Audit trail
Record Keeping	Records of copyright compliance measures	Compliance log	Continuous
Dispute Resolution	Process for handling copyright disputes	Dispute handling procedure	Per dispute

Mandatory Actions:

• Establish and implement copyright compliance policy
• Identify and respect opt-out reservations under Art. 4(3) of Directive 2019/790
• Maintain records of copyright compliance measures
• Implement dispute resolution process for copyright claims

CR-003.2: Training Data Summary Publication

Prepare and publish a sufficiently detailed training data summary using the template provided by the AI Office.

Training Data Summary Requirements:

Element	Description	Detail Level	Publication
Data Sources	General description of training data sources	Sufficiently detailed summary	Public
Data Types	Types of data used (text, image, code, etc.)	Category level	Public
Data Preparation	Key data preparation and processing methods	Methodology overview	Public
Data Provenance	Origin and provenance of training data	Summary level	Public
AI Office Template	Compliance with AI Office template format	Full template completion	Public

Mandatory Actions:

• Create sufficiently detailed training data summary
• Use AI Office template for the summary
• Publish training data summary publicly
• Update summary when training data changes materially

Evidence Required:

• Copyright compliance policy
• Rights reservation identification and compliance records
• Published training data summary
• AI Office template completion records
• Copyright compliance audit trail

Audit Verification:

• Verify copyright compliance policy exists and is implemented
• Confirm opt-out reservations identified and respected
• Check training data summary published using AI Office template
• Validate training data summary is sufficiently detailed
• Review copyright dispute handling records

---

Control GPAI-004: Systemic Risk Classification and Notification

Control ID: GPAI-004 Control Name: Systemic Risk Classification and Notification Control Type: Preventive Control Frequency: Per model release, upon capability change Risk Level: Critical

Control Objective

Classify GPAI models for systemic risk based on high-impact capabilities or computational thresholds and notify the European Commission when a GPAI model meets systemic risk criteria (Art. 51-52).

Control Requirements

CR-004.1: Systemic Risk Classification

Assess GPAI models against systemic risk criteria defined in Art. 51.

Systemic Risk Classification Criteria:

Criterion	Description	Threshold	Assessment Method
High-Impact Capabilities	Model has high-impact capabilities as determined by the Commission	Commission decision or designation	Capability assessment against Commission criteria
Computational Threshold	Cumulative amount of computation used for training exceeds threshold	10^25 FLOPs	Computational resource accounting
Commission Designation	Commission designates model as systemic risk based on criteria in Annex XIII	Commission decision	Commission notification receipt

GPAI Model Systemic Risk Decision Flow:

Step	Action	Responsible	Timeline
1. Initial Assessment	Assess model against systemic risk criteria	Model Development Team	Before market placement
2. FLOP Calculation	Calculate cumulative training computation	Model Development Team	Per training run
3. Capability Assessment	Evaluate for high-impact capabilities	AI Act Program Manager	Per model release
4. Classification Decision	Make formal classification determination	AI Governance Committee	Before market placement
5. Notification	Notify Commission if threshold met	AI Act Program Manager	Within 2 weeks

Mandatory Actions:

• Assess GPAI models for high-impact capabilities indicating systemic risk
• Monitor for 10^25 FLOP cumulative computational threshold
• Notify European Commission within 2 weeks of systemic risk threshold being met
• Maintain classification assessment records
• Reassess classification upon material model changes

CR-004.2: Commission Notification

Notify the European Commission when a GPAI model meets systemic risk criteria.

Notification Requirements:

Requirement	Description	Timeline	Method
Notification trigger	Systemic risk threshold met or Commission designation	Immediate awareness	Internal alert
Commission notification	Formal notification to European Commission	Within 2 weeks of threshold being met	Official communication channel
Documentation	Record of notification and Commission response	Upon notification	Notification register
Ongoing monitoring	Monitor for changes affecting classification	Continuous	Periodic review

Evidence Required:

• Systemic risk classification assessments
• Computational resource calculations (FLOP records)
• European Commission notification records
• Classification reassessment records
• High-impact capability assessment documentation

Audit Verification:

• Verify systemic risk classification performed for all GPAI models
• Confirm FLOP calculations documented and accurate
• Check Commission notifications sent within required timeline
• Validate classification reassessed upon material changes
• Review classification decision documentation

---

Control GPAI-005: Systemic Risk Model Evaluation and Adversarial Testing

Control ID: GPAI-005 Control Name: Systemic Risk Model Evaluation and Adversarial Testing Control Type: Preventive Control Frequency: Per model release, annually, upon material change Risk Level: Critical

Control Objective

Perform model evaluations using standardised protocols and conduct adversarial testing for GPAI models classified with systemic risk, assessing and mitigating risks at Union level (Art. 55(1)(a)-(b)).

Note: This control applies only to GPAI models classified as presenting systemic risk under Art. 51. These obligations apply regardless of open-source status.

Control Requirements

CR-005.1: Standardised Model Evaluation

Conduct model evaluations using standardised protocols, including benchmarks and testing methodologies established or referenced by the AI Office.

Model Evaluation Requirements:

Evaluation Type	Description	Methodology	Frequency
Benchmark Evaluations	Performance against standardised benchmarks	AI Office protocols and recognised benchmarks	Per model release
Capability Assessments	Assessment of model capabilities and emergent behaviours	Structured capability testing	Per model release + annually
Safety Evaluations	Assessment of safety-relevant properties	Safety testing protocols	Per model release + annually
Bias and Fairness	Assessment of systematic biases	Bias testing frameworks	Per model release + annually
Robustness Testing	Assessment of model robustness	Perturbation and stress testing	Per model release

Mandatory Actions:

• Conduct standardised model evaluations including benchmarks
• Use evaluation methodologies aligned with AI Office protocols
• Document all evaluation findings comprehensively
• Share evaluation results with AI Office as requested

CR-005.2: Adversarial Testing (Red-Teaming)

Conduct adversarial testing to identify and address vulnerabilities, including through red-teaming exercises.

Adversarial Testing Requirements:

Testing Area	Description	Method	Documentation
Prompt Injection	Resistance to prompt injection attacks	Automated and manual testing	Test results and mitigations
Jailbreaking	Resistance to safety bypass attempts	Red-team exercises	Findings and fixes
Misuse Scenarios	Testing for potential misuse pathways	Scenario-based testing	Risk assessment and mitigations
Emergent Risks	Testing for unexpected or dangerous capabilities	Exploratory testing	Capability documentation
Systemic Risks	Assessment of risks at Union level	Structured risk assessment	Risk mitigation plans

Mandatory Actions:

• Perform adversarial testing (red-teaming) to identify vulnerabilities
• Assess and mitigate systemic risks at Union level
• Document findings and implement mitigations
• Engage with AI Office on evaluation methodologies where applicable

Evidence Required:

• Model evaluation reports with standardised protocol results
• Adversarial testing (red-team) records and findings
• Risk mitigation plans and implementation records
• AI Office engagement records (if applicable)
• Systemic risk assessment documentation

Audit Verification:

• Verify model evaluations conducted using standardised protocols
• Confirm adversarial testing performed comprehensively
• Check systemic risk mitigation plans exist and are implemented
• Validate evaluation frequency meets requirements
• Review AI Office engagement and reporting

---

Control GPAI-006: Systemic Risk Incident Reporting and Cybersecurity

Control ID: GPAI-006 Control Name: Systemic Risk Incident Reporting and Cybersecurity Control Type: Detective Control Frequency: Continuous monitoring, upon incident Risk Level: Critical

Control Objective

Track and report serious incidents to the AI Office without undue delay and ensure adequate cybersecurity protections for GPAI models with systemic risk and their physical infrastructure (Art. 55(1)(c)-(d)).

Note: This control applies only to GPAI models classified as presenting systemic risk under Art. 51. These obligations apply regardless of open-source status.

Control Requirements

CR-006.1: Serious Incident Tracking and Reporting

Implement processes to track, assess, and report serious incidents related to GPAI models with systemic risk.

Incident Reporting Requirements:

Requirement	Description	Timeline	Responsible
Incident Detection	Mechanisms to detect serious incidents	Continuous	Model Operations Team
Incident Assessment	Assess severity and systemic implications	Within 24 hours of detection	AI Act Program Manager
AI Office Notification	Report serious incidents to AI Office	Without undue delay	AI Act Program Manager
Incident Documentation	Comprehensive incident documentation	Throughout incident lifecycle	Incident Manager
Corrective Actions	Implement and document corrective actions	Per incident	Model Development Team

Serious Incident Categories:

Category	Description	Reporting Priority
Safety Incidents	Incidents causing or potentially causing harm to health, safety, or fundamental rights	Immediate
Security Incidents	Breaches or vulnerabilities with systemic impact	Immediate
Capability Incidents	Unexpected or dangerous emergent capabilities	Urgent
Misuse Incidents	Significant misuse causing or risking harm	Urgent
Infrastructure Incidents	Failures affecting model availability or integrity at scale	High

Mandatory Actions:

• Implement incident tracking and detection mechanisms
• Report serious incidents to AI Office without undue delay
• Document all incidents comprehensively
• Implement corrective actions and track remediation

CR-006.2: Cybersecurity Protections

Ensure adequate cybersecurity for GPAI models with systemic risk and their physical infrastructure.

Cybersecurity Requirements:

Requirement	Description	Implementation	Verification
Model Security	Protect model weights, parameters, and configuration	Access controls, encryption, integrity verification	Quarterly assessment
Infrastructure Security	Secure physical and cloud infrastructure	Infrastructure security controls	Quarterly assessment
Supply Chain Security	Secure model supply chain	Vendor security assessment, code signing	Per vendor, annually
Access Control	Restrict access to model and infrastructure	Role-based access, multi-factor authentication	Continuous monitoring
Monitoring and Detection	Detect security threats and anomalies	Security monitoring, intrusion detection	Continuous
Incident Response	Respond to cybersecurity incidents	Incident response plan and team	Per incident

Mandatory Actions:

• Implement cybersecurity measures for model and physical infrastructure
• Document all protective measures taken
• Conduct regular cybersecurity assessments
• Maintain incident response capability

Evidence Required:

• Incident tracking and detection system records
• AI Office serious incident reports
• Cybersecurity assessment records
• Protective measures documentation
• Incident response and remediation records

Audit Verification:

• Verify incident tracking mechanisms are operational
• Confirm AI Office reports submitted for all serious incidents
• Check cybersecurity assessments conducted regularly
• Validate cybersecurity measures implemented and documented
• Review incident response capability and readiness

---

SUPPORTING PROCEDURES

This standard is implemented through the following detailed procedures:

Procedure PROC-AI-GPAI-001: GPAI Model Documentation Procedure

Purpose: Define step-by-step process for GPAI model technical documentation and downstream provider information Owner: AI Act Program Manager Implements: Controls GPAI-001, GPAI-002

Procedure Steps:

1. Identify GPAI models requiring documentation
2. Prepare Annex XI technical documentation - Control GPAI-001
3. Prepare Annex XII downstream provider information - Control GPAI-002
4. Assess open-source exemption eligibility
5. Distribute information to downstream providers
6. Maintain and update documentation
7. Review documentation annually

Outputs:

• Annex XI technical documentation
• Annex XII information packages
• Open-source exemption assessments
• Distribution records

---

Procedure PROC-AI-GPAI-002: Copyright Compliance and Training Data Summary Procedure

Purpose: Define process for copyright compliance and training data summary publication Owner: Legal / AI Act Program Manager Implements: Control GPAI-003

Procedure Steps:

1. Establish copyright compliance policy
2. Implement opt-out reservation identification process
3. Screen training data for rights reservations
4. Create training data summary using AI Office template
5. Publish training data summary
6. Monitor for new opt-out reservations
7. Handle copyright disputes

Outputs:

• Copyright compliance policy
• Rights reservation records
• Published training data summary
• Dispute handling records

---

Procedure PROC-AI-GPAI-003: Systemic Risk Classification and Notification Procedure

Purpose: Define process for systemic risk classification and Commission notification Owner: AI Act Program Manager Implements: Control GPAI-004

Procedure Steps:

1. Calculate cumulative training computation (FLOPs)
2. Assess model for high-impact capabilities
3. Determine systemic risk classification
4. Prepare Commission notification (if applicable)
5. Submit notification within 2-week deadline
6. Monitor for changes affecting classification
7. Reassess upon material model changes

Outputs:

• Classification assessment records
• FLOP calculations
• Commission notification records
• Reassessment records

---

Procedure PROC-AI-GPAI-004: Systemic Risk Model Evaluation and Incident Management Procedure

Purpose: Define process for model evaluation, adversarial testing, incident reporting, and cybersecurity for systemic risk models Owner: AI Act Program Manager Implements: Controls GPAI-005, GPAI-006

Procedure Steps:

1. Plan model evaluations per standardised protocols
2. Conduct benchmark evaluations and capability assessments
3. Perform adversarial testing (red-teaming)
4. Document findings and implement mitigations
5. Monitor for serious incidents
6. Report serious incidents to AI Office
7. Conduct cybersecurity assessments
8. Implement and document protective measures

Outputs:

• Model evaluation reports
• Adversarial testing records
• Incident reports
• Cybersecurity assessment records

---

COMPLIANCE

5.1 Compliance Monitoring

Monitoring Approach: Continuous automated monitoring of GPAI model compliance supplemented by quarterly manual reviews and annual comprehensive audits.

Compliance Metrics:

Metric	Target	Measurement Method	Frequency	Owner
Annex XI Documentation Completeness	100%	% of GPAI models with complete documentation	Quarterly	AI Act Program Manager
Annex XII Information Distribution	100%	% of downstream providers with current information	Quarterly	AI Act Program Manager
Copyright Compliance Rate	100%	% of models with copyright policy in place	Quarterly	Legal
Training Data Summary Publication	100%	% of models with published summaries	Quarterly	AI Act Program Manager
Systemic Risk Evaluation Completeness	100%	% of systemic risk models with completed evaluations	Quarterly	AI Act Program Manager
Incident Reporting Timeliness	100%	% of serious incidents reported without undue delay	Per incident	AI Act Program Manager

Monitoring Tools:

• GPAI Model Register
• Documentation Management System
• Compliance Dashboard
• Quarterly AI Governance Committee reviews

---

5.2 Internal Audit Requirements

Audit Frequency: Annually (minimum)

Audit Scope:

• GPAI model documentation completeness (Annex XI and XII)
• Copyright compliance policy implementation
• Training data summary publication
• Systemic risk classification accuracy
• Model evaluation and adversarial testing completeness
• Incident reporting timeliness
• Cybersecurity measures adequacy
• Controls effectiveness (GPAI-001 through GPAI-006)

Audit Activities:

• Review 100% of GPAI model documentation
• Verify copyright compliance records
• Test systemic risk classification process
• Review model evaluation reports
• Check incident reporting records
• Assess cybersecurity measures

Audit Outputs:

• Annual GPAI Model Compliance Audit Report
• Findings and recommendations
• Corrective action plans for deficiencies

---

5.3 External Audit / Regulatory Inspection

Preparation:

• Maintain audit-ready GPAI documentation at all times
• Designate AI Act Program Manager and Legal as regulatory liaisons
• Prepare standard response procedures for AI Office and authority requests

Provide to Auditors/Regulators:

• Annex XI technical documentation
• Annex XII downstream provider information
• Copyright compliance records
• Published training data summaries
• Systemic risk classification assessments
• Model evaluation and adversarial testing reports
• Incident reports
• Cybersecurity assessment records
• Internal audit reports
• Evidence of controls execution

Authority Request Response:

• Acknowledge request within 1 business day
• Provide requested documentation within 5 business days
• Coordinate through Legal and AI Act Program Manager
• Document all interactions with authorities

---

ROLES AND RESPONSIBILITIES

6.1 RACI Matrix

Activity	AI Act Program Manager	Legal	Model Development Team	Model Evaluation Team	CISO	AI Governance Committee
Annex XI Documentation	A	C	R	C	I	I
Annex XII Information	A	C	R	C	I	I
Copyright Compliance	C	R/A	C	I	I	I
Training Data Summary	R/A	C	R	I	I	I
Systemic Risk Classification	R/A	C	R	R	I	A
Model Evaluation	A	I	C	R	I	I
Adversarial Testing	A	I	C	R	R	I
Incident Reporting	R/A	C	C	C	C	I
Cybersecurity	C	I	C	I	R/A	I

RACI Legend:

• R = Responsible (does the work)
• A = Accountable (ultimately answerable)
• C = Consulted (provides input)
• I = Informed (kept up-to-date)

---

6.2 Role Descriptions

AI Act Program Manager

• Primary Responsibility: Owns GPAI compliance framework, coordinates all GPAI model compliance activities
• Key Activities:
  • Oversees Annex XI and XII documentation
  • Manages systemic risk classification process
  • Coordinates Commission notifications
  • Manages incident reporting to AI Office
  • Reports to AI Governance Committee
• Required Competencies: EU AI Act GPAI provisions (Art. 51-56), model documentation, regulatory engagement

Legal

• Primary Responsibility: Owns copyright compliance, advises on regulatory obligations
• Key Activities:
  • Establishes copyright compliance policy
  • Manages rights reservation compliance
  • Advises on open-source exemption eligibility
  • Handles copyright disputes
• Required Competencies: EU copyright law, Directive 2019/790, EU AI Act GPAI provisions

Model Development Team

• Primary Responsibility: Creates and maintains GPAI model documentation
• Key Activities:
  • Prepares Annex XI technical documentation
  • Prepares Annex XII downstream provider information
  • Documents training data and computational resources
  • Supports systemic risk classification
• Required Competencies: AI model development, technical documentation, model architecture

Model Evaluation Team

• Primary Responsibility: Conducts model evaluations and adversarial testing
• Key Activities:
  • Performs standardised model evaluations
  • Conducts adversarial testing (red-teaming)
  • Documents evaluation findings
  • Supports systemic risk assessment
• Required Competencies: Model evaluation, adversarial testing, safety assessment, benchmarking

CISO (Chief Information Security Officer)

• Primary Responsibility: Owns cybersecurity for GPAI models with systemic risk
• Key Activities:
  • Implements cybersecurity measures for models and infrastructure
  • Conducts cybersecurity assessments
  • Supports adversarial testing
  • Manages security incident response
• Required Competencies: Cybersecurity, AI system security, incident response

AI Governance Committee

• Primary Responsibility: Provides governance oversight and approves systemic risk classifications
• Key Activities:
  • Approves systemic risk classification decisions
  • Reviews GPAI compliance reports
  • Oversees incident resolution
  • Provides strategic direction
• Required Competencies: AI governance, EU AI Act, risk management

---

EXCEPTIONS

7.1 Exception Philosophy

GPAI model compliance is a critical regulatory obligation under the EU AI Act. Exceptions are granted restrictively and only where compensating controls adequately mitigate risks. Non-compliance with GPAI obligations may result in penalties of up to EUR 15 million or 3% of global annual turnover, whichever is higher.

---

7.2 Allowed Exceptions

The following exceptions may be granted with proper justification and approval:

Exception Type	Justification Required	Maximum Duration	Approval Authority	Compensating Controls
Extended Documentation Timeline	Resource constraints prevent timely completion	30 days	AI Act Program Manager	Interim documentation; Accelerated plan
Alternative Evaluation Method	Standardised protocol not yet available for model type	Until protocol available	AI Governance Committee	Alternative rigorous methodology; Document rationale
Open-Source Exemption Claim	Model meets all Art. 53(2) criteria	Permanent (subject to review)	AI Act Program Manager + Legal	Document exemption basis; Monitor for systemic risk

---

7.3 Prohibited Exceptions

The following exceptions cannot be granted under any circumstances:

• Skipping copyright compliance - Mandatory per Art. 53(1)(c) for all GPAI models, no exceptions including open-source
• Skipping training data summary publication - Mandatory per Art. 53(1)(d) for all GPAI models, no exceptions including open-source
• Skipping Commission notification for systemic risk - Mandatory per Art. 52, 2-week deadline, no exceptions
• Skipping incident reporting for systemic risk models - Mandatory per Art. 55(1)(c), no exceptions
• Claiming open-source exemption for systemic risk models - Art. 53(2) exemption does not apply to systemic risk models

---

7.4 Exception Request Process

Step 1: Submit Exception Request

• Complete Exception Request Form (FORM-AI-EXCEPTION-001)
• Include business justification
• Propose compensating controls
• Specify duration requested
• Attach risk assessment including regulatory penalty risk

Step 2: Risk Assessment

• AI Act Program Manager assesses risk of granting exception
• Evaluates adequacy of compensating controls
• Assesses regulatory exposure (EUR 15 million / 3% turnover)
• Documents residual risk

Step 3: Approval

• Route to appropriate approval authority based on exception type
• AI Act Program Manager approval: Minor documentation exceptions
• AI Governance Committee: Significant exceptions or systemic risk matters
• AI Governance Committee + Legal: Exceptions with regulatory exposure

Step 4: Documentation and Monitoring

• Document exception in Exception Register
• Assign exception owner
• Set review date
• Monitor compensating controls
• Report exceptions quarterly to AI Governance Committee

Step 5: Exception Review and Closure

• Review exception at specified review date
• Assess if exception still needed
• Close exception when compliance achieved
• Document lessons learned

---

ENFORCEMENT

8.1 Non-Compliance Consequences

Violation	Severity	Consequence	Remediation Required
Missing Annex XI documentation	Critical	Immediate escalation; Model market access review	Complete documentation within 10 business days
Missing Annex XII information	Critical	Downstream provider notification; Escalation	Complete and distribute within 10 business days
Copyright non-compliance	Critical	Legal review; Potential model suspension	Implement compliance measures within 5 business days
Training data summary not published	High	Immediate publication required	Publish within 5 business days
Systemic risk notification missed	Critical	Immediate Commission notification; Legal review	Notify immediately; Document delay
Model evaluation not completed	Critical	Model availability review; Escalation	Complete evaluation within 15 business days
Incident not reported	Critical	Immediate AI Office report; Investigation	Report immediately; Root cause analysis
Cybersecurity measures inadequate	Critical	Immediate security review; Potential suspension	Implement measures within 10 business days

---

8.2 Escalation Procedures

Level 1: AI Act Program Manager

• Minor documentation gaps
• Administrative delays < 5 days
• Action: Written warning, corrective action required

Level 2: AI Act Program Manager + AI Governance Committee

• Material documentation gaps
• Missed notification deadlines
• Evaluation or testing gaps
• Action: Formal review, corrective action plan, management notification

Level 3: AI Governance Committee + Legal

• Systemic risk notification failures
• Copyright non-compliance
• Incident reporting failures
• Action: Immediate investigation, model market access review, regulatory strategy

Level 4: Executive Management + Legal

• Potential regulatory enforcement action
• Significant legal liability (EUR 15 million / 3% turnover exposure)
• Reputational risk
• Action: Executive crisis management, legal strategy, regulatory engagement

---

8.3 Immediate Escalation Triggers

Escalate immediately to AI Governance Committee + Legal if:

• Systemic risk GPAI model operating without required evaluations
• Serious incident not reported to AI Office
• Commission notification deadline at risk of being missed
• Regulatory inquiry or inspection related to GPAI compliance
• Copyright infringement claim related to training data

---

8.4 Regulatory Penalties

Non-compliance with GPAI model obligations under Articles 51-56 may result in:

• Administrative fines of up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher
• Orders to bring the GPAI model into compliance
• Restrictions on market access
• Reputational damage

---

KEY PERFORMANCE INDICATORS (KPIs)

9.1 GPAI Model Compliance KPIs

KPI ID	KPI Name	Definition	Target	Measurement Method	Frequency	Owner	Reporting To
KPI-GPAI-001	Technical Documentation Completeness	% of GPAI models with complete Annex XI documentation	100%	(# complete / # total models) x 100	Quarterly	AI Act Program Manager	AI Governance Committee
KPI-GPAI-002	Downstream Provider Information Rate	% of GPAI models with complete Annex XII information for downstream providers	100%	(# complete / # total) x 100	Quarterly	AI Act Program Manager	AI Governance Committee
KPI-GPAI-003	Copyright Compliance Rate	% of GPAI models with copyright compliance policy in place	100%	(# compliant / # total) x 100	Quarterly	Legal	AI Governance Committee
KPI-GPAI-004	Training Data Summary Publication	% of GPAI models with published training data summary	100%	(# published / # total) x 100	Quarterly	AI Act Program Manager	AI Governance Committee
KPI-GPAI-005	Systemic Risk Model Evaluation Rate	% of systemic risk GPAI models with completed model evaluations	100%	(# evaluated / # systemic risk models) x 100	Quarterly	AI Act Program Manager	AI Governance Committee

---

9.2 KPI Dashboards and Reporting

Real-Time Dashboard (AI Act Program Manager access)

• Current GPAI model compliance status
• Documentation completeness scores
• Systemic risk model evaluation status
• Open incidents and resolution progress
• Commission notification status

Monthly Management Report

• KPI-GPAI-001, 002, 003, 004
• Trend analysis (vs. previous month)
• Issues and risks
• Planned actions

Quarterly AI Governance Committee Report

• All KPIs
• GPAI model compliance assessment
• Systemic risk model status
• Internal audit findings (if conducted)
• Exception register review

Annual Executive Report

• Full-year KPI performance
• GPAI compliance maturity assessment
• Regulatory engagement summary
• Strategic recommendations

---

9.3 KPI Thresholds and Alerts

KPI	Green (Good)	Yellow (Warning)	Red (Critical)	Alert Action
Documentation Completeness	100%	90-99%	< 90%	Red: Immediate escalation to AI Governance Committee Chair
Downstream Provider Information	100%	90-99%	< 90%	Red: Escalate to AI Governance Committee
Copyright Compliance Rate	100%	95-99%	< 95%	Yellow: Improvement plan; Red: Escalate to Legal + AI Governance Committee
Training Data Summary Publication	100%	90-99%	< 90%	Red: Immediate publication required
Systemic Risk Evaluation Rate	100%	-	< 100%	Red: Immediate escalation; any gap is critical

---

TRAINING REQUIREMENTS

10.1 Training Program Overview

All personnel involved in GPAI model compliance must complete role-specific training to ensure competency in documentation, classification, evaluation, incident reporting, and cybersecurity requirements.

---

10.2 Role-Based Training Requirements

Role	Training Course	Duration	Content	Frequency	Assessment Required
AI Act Program Manager	GPAI Compliance Expert Training	16 hours	GPAI obligations; Art. 51-56; Annex XI/XII; Systemic risk; AI Office engagement	Initial + annually	Yes - Written exam (>=90%)
Legal	GPAI Copyright and Regulatory Training	12 hours	Copyright compliance; Directive 2019/790; Open-source exemptions; Penalties	Initial + annually	Yes - Written exam (>=90%)
Model Development Team	GPAI Documentation Training	8 hours	Annex XI requirements; Annex XII requirements; Documentation standards	Initial + annually	Yes - Practical exercise
Model Evaluation Team	GPAI Evaluation and Testing Training	12 hours	Standardised evaluation protocols; Adversarial testing; Red-teaming	Initial + annually	Yes - Practical exercise
CISO	GPAI Cybersecurity Training	8 hours	GPAI cybersecurity requirements; Model security; Infrastructure protection	Initial + annually	Yes - Written exam (>=90%)

---

10.3 Training Content by Topic

GPAI Regulatory Framework

• EU AI Act Articles 51-56
• Annex XI and XII requirements
• Open-source exemption criteria (Art. 53(2))
• Penalty framework (EUR 15 million / 3% turnover)

Technical Documentation

• Annex XI documentation elements
• Model architecture documentation
• Training process documentation
• Computational resource documentation

Systemic Risk

• Classification criteria (Art. 51)
• 10^25 FLOP threshold
• Commission notification process
• Model evaluation and adversarial testing requirements

Copyright and Training Data

• Directive 2019/790 requirements
• Opt-out reservation compliance
• Training data summary preparation
• AI Office template usage

---

10.4 Training Delivery Methods

Initial Training:

• Instructor-led classroom or virtual training
• Includes interactive exercises and case studies
• Hands-on practice with documentation templates
• Group discussions of systemic risk scenarios

Annual Refresher:

• E-learning modules for core content review
• Live update sessions for regulatory changes
• Case study reviews of recent GPAI compliance activities
• Knowledge assessment

On-the-Job Training:

• Mentoring for new team members
• Supervised documentation preparation for first 2 models
• Supervised evaluation for first systemic risk assessment

Just-in-Time Training:

• Quick reference guides for Annex XI/XII requirements
• Systemic risk classification decision aids
• Incident reporting checklists
• Copyright compliance job aids

---

10.5 Training Effectiveness Measurement

Assessment Methods:

• Written exams for knowledge retention
• Practical exercises for documentation skill application
• On-the-job observations for competency validation
• Feedback surveys for training quality

Competency Validation:

• Model Development Team: Must demonstrate ability to prepare 1 complete Annex XI documentation package with 100% completeness before independent work
• Model Evaluation Team: Must participate in 1 supervised model evaluation before independent work
• All staff: Must pass knowledge assessments with minimum required scores

Training Metrics:

Metric	Target	Frequency
Training completion rate	100%	Quarterly
Assessment pass rate (first attempt)	>= 90%	Per training
Training effectiveness score (survey)	>= 4.0/5.0	Per training
Time to competency (new staff)	< 60 days	Per person

---

10.6 Training Records

Records Maintained:

• Training attendance records
• Assessment scores
• Competency validations
• Refresher training completion
• Individual training transcripts

Retention: 10 years (to align with EU AI Act documentation retention)

Access: AI Act Program Manager, HR, Internal Audit, Competent Authorities (upon request)

---

DEFINITIONS

Term	Definition	Source
General-Purpose AI Model (GPAI Model)	An AI model, including where such a model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks	EU AI Act Article 3(63)
Systemic Risk	A risk that is specific to the high-impact capabilities of GPAI models, having a significant effect on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole	EU AI Act Article 3(65)
GPAI Model with Systemic Risk	A GPAI model classified as presenting systemic risk based on high-impact capabilities or exceeding the 10^25 FLOP computational threshold	EU AI Act Article 51
Downstream Provider	A provider of an AI system that integrates a GPAI model into their system	EU AI Act
Annex XI	Technical documentation requirements for GPAI models	EU AI Act Annex XI
Annex XII	Information requirements for downstream providers of GPAI models	EU AI Act Annex XII
Adversarial Testing	Testing designed to identify vulnerabilities, weaknesses, and potential misuse pathways in AI models, including red-teaming exercises	EU AI Act Art. 55(1)(b)
AI Office	The EU body established to oversee GPAI model compliance and enforcement	EU AI Act
Open-Source GPAI Model	A GPAI model with publicly available parameters, weights, architecture, and usage information released under a free and open-source licence	EU AI Act Art. 53(2)
Training Data Summary	A sufficiently detailed summary of training data used for the GPAI model, prepared using the AI Office template	EU AI Act Art. 53(1)(d)

---

LINK WITH AI ACT AND ISO42001

12.1 EU AI Act Regulatory Mapping

This standard implements the following EU AI Act requirements:

EU AI Act Provision	Article	Requirement Summary	Implemented By (Controls)
GPAI Model Classification	Article 51	Classification of GPAI models as systemic risk	GPAI-004
Systemic Risk Presumption	Article 51(2)	Presumption of systemic risk at 10^25 FLOPs	GPAI-004
Commission Notification	Article 52	Notification when systemic risk threshold met	GPAI-004
GPAI Model Obligations	Article 53	Obligations for all GPAI model providers	GPAI-001, GPAI-002, GPAI-003
Technical Documentation	Article 53(1)(a)	Annex XI technical documentation	GPAI-001
Downstream Provider Info	Article 53(1)(b)	Annex XII downstream provider information	GPAI-002
Copyright Compliance	Article 53(1)(c)	Copyright policy per Directive 2019/790	GPAI-003
Training Data Summary	Article 53(1)(d)	Publish training data summary	GPAI-003
Open-Source Exemption	Article 53(2)	Exemption for open-source models (Art. 53(1)(a)-(b) only)	GPAI-001, GPAI-002
Authorised Representatives	Article 54	Appointment for non-EU providers	All controls
Systemic Risk Obligations	Article 55	Additional obligations for systemic risk models	GPAI-005, GPAI-006
Model Evaluation	Article 55(1)(a)	Standardised model evaluations	GPAI-005
Adversarial Testing	Article 55(1)(b)	Adversarial testing including red-teaming	GPAI-005
Incident Reporting	Article 55(1)(c)	Serious incident reporting to AI Office	GPAI-006
Cybersecurity	Article 55(1)(d)	Adequate cybersecurity protections	GPAI-006
Codes of Practice	Article 56	Compliance via codes of practice	All controls

---

12.2 ISO/IEC 42001:2023 Alignment

This standard aligns with ISO/IEC 42001:2023 as follows:

ISO 42001 Clause	Requirement	Implementation in This Standard
Clause 6.1: Actions to address risks	Risk identification and treatment	GPAI-004, GPAI-005
Clause 7.5: Documented information	Documentation management	GPAI-001, GPAI-002, GPAI-003
Clause 8.1: Operational planning and control	Operational controls	All controls
Clause 9.1: Monitoring, measurement, analysis and evaluation	Performance monitoring	All KPIs

---

12.3 Relationship to Other Standards

This GPAI model compliance standard integrates with other AI Act standards:

Related Standard	Integration Point	Rationale
STD-AI-001: Classification	GPAI model classification feeds into AI system classification	Downstream AI systems using GPAI models may be high-risk
STD-AI-002: Risk Management	Systemic risk assessment methodology	Risk management framework applies to GPAI systemic risk
STD-AI-004: Technical Documentation	Annex XI documentation aligns with Annex IV	Documentation standards complement each other
STD-AI-008: Accuracy, Robustness, Security	Model evaluation and cybersecurity	Evaluation and security requirements overlap
STD-AI-012: Post-Market Monitoring	Incident monitoring and reporting	Post-market monitoring feeds into GPAI incident reporting
STD-AI-013: Incident Management	Serious incident reporting	Incident management processes support GPAI incident reporting

---

12.4 References and Related Documents

EU AI Act (Regulation (EU) 2024/1689):

• Article 51: Classification of GPAI models with systemic risk
• Article 52: Notification of GPAI models with systemic risk
• Article 53: Obligations for providers of GPAI models
• Article 53(1)(a)-(d): Specific GPAI model obligations
• Article 53(2): Open-source exemption
• Article 54: Authorised representatives for GPAI model providers
• Article 55: Obligations for providers of GPAI models with systemic risk
• Article 55(1)(a)-(d): Specific systemic risk obligations
• Article 56: Codes of practice
• Annex XI: Technical documentation for GPAI models
• Annex XII: Information for downstream providers
• Annex XIII: Criteria for designation of GPAI models with systemic risk

EU Copyright Directive:

• Directive (EU) 2019/790, Article 4(3): Text and data mining opt-out

Internal Documents:

• POL-AI-001: Artificial Intelligence Policy (parent policy)
• STD-AI-001: AI System Classification Standard
• STD-AI-002: AI Risk Management Standard
• STD-AI-004: AI Technical Documentation Standard
• STD-AI-008: AI Accuracy, Robustness, and Security Standard
• STD-AI-012: AI Post-Market Monitoring Standard
• STD-AI-013: AI Incident Management Standard
• PROC-AI-GPAI-001 through -004: GPAI compliance procedures

---

APPROVAL AND AUTHORIZATION

Role	Name	Title	Signature	Date
Prepared By	AI Act Program Manager	AI Act Program Manager	_________________	________
Reviewed By	Sarah Johnson	AI Act Program Manager	_________________	________
Reviewed By	Jane Doe	Chief Strategy & Risk Officer	_________________	________
Approved By	Jane Doe	AI Governance Committee Chair	_________________	________

Effective Date: 2025-08-02 Next Review Date: 2026-08-02 Review Frequency: Annually or upon regulatory change

---

END OF STANDARD STD-AI-018

---

This standard is a living document. Feedback and improvement suggestions should be directed to the AI Act Program Manager.