sota.io
2026-04-16·12 min read·

EU AI Act Art.59 Personal Data Processing for AI Development in Regulatory Sandboxes — Developer Guide (2026)

EU AI Act Article 59 solves one of the hardest practical problems in AI development within the EU: where does the training data come from? High-quality AI systems require large, representative datasets — but the personal data that makes datasets representative is typically collected for purposes other than AI development. GDPR's purpose limitation principle (Art.5(1)(b)) normally prohibits using that data for new, incompatible purposes.

Art.59 creates a targeted exception: within AI regulatory sandboxes established under Art.57, providers may further process personal data originally collected for other purposes — provided the processing satisfies six cumulative conditions designed to protect data subjects while enabling legitimate AI innovation.

Art.59 is not a blanket GDPR derogation. It does not override GDPR's lawful basis requirements for the original collection, does not remove data subject rights, and does not permit processing outside the sandbox boundary. What it does is resolve the purpose compatibility question for a carefully bounded category of AI development activity — sandbox testing and validation — where the public interest in reliable, representative AI systems justifies the further use of existing data.

This guide covers Art.59(1)–(4) in full, the GDPR Art.6(4) compatibility framework, each of the six conditions, the sandbox-boundary constraint, special category data interaction, CLOUD Act jurisdiction risk for sandbox training data, and Python implementation for sandbox data processing governance.

Art.59 became applicable on 2 August 2025 as part of Chapter VI (Measures in Support of Innovation) of the EU AI Act (Regulation (EU) 2024/1689).

Art.59 in the Chapter VI Innovation Framework

ArticleMechanismApproach
Art.57AI Regulatory SandboxControlled environment, sandbox plan, authority partnership
Art.58Real-World Testing Outside SandboxNotification-based, 30-day implicit consent, independent testing
Art.59Personal Data Processing in SandboxFurther-processing exception for sandbox training/testing data
Art.60–63Further Innovation SupportSME measures, regulatory guidance, codes of practice

Art.59 is architecturally dependent on Art.57: it only operates within a formally established AI regulatory sandbox. Providers using the Art.58 real-world testing pathway outside the sandbox do not gain Art.59's further-processing exception — their data processing remains governed entirely by standard GDPR rules and original consent/lawful basis.

The Core Problem Art.59 Solves

Purpose Limitation Under GDPR

GDPR Art.5(1)(b) establishes the purpose limitation principle: personal data shall be collected for specified, explicit, and legitimate purposes and not further processed in a manner incompatible with those purposes. This principle is fundamental to GDPR and not subject to general exemption.

GDPR Art.6(4) provides a compatibility assessment framework for determining whether further processing is compatible with the original purpose. The Art.6(4) factors are:

  1. Link between purposes: Is there a link between the original and new purpose?
  2. Context of collection: What was the context in which data was collected, and reasonable expectations of data subjects?
  3. Nature of data: Is special category data (Art.9) or criminal conviction data (Art.10) involved?
  4. Consequences of further processing: What are the possible consequences for data subjects?
  5. Safeguards in place: Are there appropriate safeguards (encryption, pseudonymisation)?

If further processing is compatible under Art.6(4), no new lawful basis is needed. If incompatible, a new lawful basis is required. Art.59 provides that new lawful basis for sandbox AI development.

The AI Development Data Problem

Without Art.59, a provider wanting to train a high-risk AI system on hospital patient data (originally collected for treatment) would face:

Art.59 addresses this impasse for the sandbox context, recognising that high-quality AI systems that serve public interests require representative real-world data — and that the controlled sandbox environment provides sufficient safeguards to justify carefully bounded further processing.

Art.59(1): The Permission and Its Conditions

The Grant

Art.59(1) states that personal data lawfully collected for other purposes may be processed in the AI regulatory sandbox for the purpose of developing, training, testing, and validating AI systems — provided the following conditions are cumulatively satisfied.

The phrase "lawfully collected for other purposes" is important: Art.59 does not create a basis for the original collection. Data must have been collected lawfully (under Art.6(1) or Art.9(2) as applicable). Art.59 only addresses what can be done with that data once it is in the provider's possession.

Condition 1: AI Systems in the Public Interest

Art.59(1)(a): The AI system being developed must be in the public interest within a field listed in Art.57(1) or otherwise justified under the sandbox authorisation.

What counts as public interest for Art.59:

What does NOT qualify:

The public interest requirement creates a meaningful filter. A healthcare startup developing a patient triage AI qualifies; a fintech developing a private credit scoring model for competitive advantage likely does not.

Condition 2: Processing Necessary and Proportionate

Art.59(1)(b): The further processing must be necessary and proportionate to the AI development goal. This mirrors the proportionality principle applied throughout GDPR and the EU AI Act.

Necessity assessment:

Proportionality assessment:

Providers should document the necessity and proportionality assessment in the sandbox plan under Art.57(3). Sandbox authorities review this assessment as part of plan approval.

Condition 3: No Adverse Effects on Data Subjects

Art.59(1)(c): The further processing must not produce adverse effects for the data subjects whose data is used for AI training and testing.

What "adverse effects" means:

Practical safeguards to prevent adverse effects:

The "no adverse effects" condition is ongoing, not just assessed at the point of processing. Providers must monitor whether sandbox training produces outputs that adversely affect subjects — for example, by detecting whether outputs correlate with protected characteristics in ways that could harm individuals if deployed.

Condition 4: Subject to Appropriate Technical and Organisational Measures

Art.59(1)(d): Processing must be subject to appropriate technical and organisational measures (TOMs) protecting data subjects' rights and freedoms.

Required TOMs for Art.59 compliance:

TOM CategoryExamples
PseudonymisationTokenisation of identity attributes; key management separate from training
EncryptionData encrypted at rest and in transit; key management documented
Access controlRole-based access; training data accessible only to necessary personnel
Data minimisationFeature selection limited to required attributes; remove unnecessary columns
AggregationWhere possible, aggregate rather than use individual records
Audit loggingAll access to sandbox training data logged and retained
Incident responseBreach detection and response procedures for sandbox data
TestingRegular adversarial testing for re-identification risk

The TOM requirement interacts with GDPR Art.25 (data protection by design and by default), which applies independently. Art.59 TOMs are additional, AI-development-specific safeguards layered on top of standard GDPR Art.25 obligations.

Condition 5: Sandbox Plan Includes Data Processing

Art.59(1)(e): The data processing under Art.59 must be covered by the sandbox plan agreed with the competent authority under Art.57(3).

This is the most significant structural condition: Art.59 cannot be invoked unilaterally. The further processing of personal data must be disclosed to and accepted by the sandbox supervisory authority as part of the formal sandbox plan. This means:

Providers who wish to use Art.59 cannot simply process the data and claim the sandbox exception after the fact. The exception requires prior authorisation through the sandbox plan process.

Condition 6: No Transfer Outside the Union

Art.59(1)(f): Personal data processed under Art.59 must not be transferred to third countries or international organisations unless adequate protection equivalent to GDPR standards is ensured.

This condition reflects the EU's approach to maintaining the data protection level of the GDPR framework even in innovation contexts. The Art.59 exception is specifically designed for EU-supervised sandbox environments; extending the data to third-country jurisdictions would undermine the supervisory oversight that justifies the exception.

Practical implications:

Art.59(2): The Sandbox Boundary Constraint

Art.59(2) establishes the critical sandbox boundary rule: personal data processed under Art.59 must remain within the sandbox environment and must not be used outside it.

What "Remain Within the Sandbox" Means

The sandbox boundary is both a data boundary and a system boundary:

Data boundary: Personal data processed under Art.59 cannot be extracted from the sandbox environment. This includes:

System boundary: AI model outputs derived from Art.59 data cannot be used in production systems without a separate data protection assessment. The sandbox-trained model is an AI artifact — once deployed outside the sandbox, it processes new data under standard GDPR rules. But the weights trained on Art.59 data carry a taint: if the model memorises training data (a known ML risk), production use could re-expose Art.59 subjects.

Art.59(2) and Model Deployment

Art.59(2) creates a compliance gap that many providers overlook: the model trained on Art.59 data in the sandbox needs a separate legal basis for deployment outside the sandbox.

The Art.59 exception covers the training/testing process. When the model is deployed in production:

Recommended practice: Before deploying a sandbox-trained model, conduct a DPIA (GDPR Art.35) that specifically addresses the Art.59 training data heritage and the risk of training data memorisation in the deployed model.

Art.59(3): Data Subject Rights During Sandbox Processing

Art.59(3) confirms that data subjects' rights under GDPR continue to apply during Art.59 processing. Art.59 does not suspend access rights (Art.15), rectification rights (Art.16), erasure rights (Art.17), or objection rights (Art.21).

The Right to Erasure Challenge

The right to erasure (GDPR Art.17) creates a practical challenge for AI training: if a data subject exercises their erasure right, the provider must delete their data — but what does "deletion" mean for a trained model?

The EU AI Act does not resolve the machine unlearning question directly, but the Art.59 framework implies:

  1. Training data must be deleted from the sandbox storage
  2. Trained model weights may need to be retrained if the deleted data had material influence
  3. Derived artefacts (embeddings, feature representations traceable to the subject) must be deleted

Machine unlearning — the ability to "forget" a training example from a model — is an active research area. Providers using Art.59 should implement:

Notification Under GDPR Art.13/14

Where providers process Art.59 data, GDPR Art.13 (data collected from the subject) or Art.14 (data not collected from the subject) notification obligations apply. The notification must disclose the further processing purpose (AI development in the sandbox) unless an Art.14(5) exception applies (disproportionate effort, impossible, or would seriously impair the purposes).

For large-scale healthcare or financial datasets where individual notification is impractical, providers may rely on the Art.14(5)(b) disproportionate effort exception — but must make the further processing information publicly available and take other appropriate measures (GDPR Art.14(5)(b)).

Art.59(4): Supervisory Authority Cooperation

Art.59(4) requires that sandbox competent authorities cooperate with national data protection authorities (DPAs) in supervising the data processing aspects of Art.59.

What DPA Cooperation Means in Practice

The competent authority for the AI regulatory sandbox (typically a market surveillance or sector regulator) has AI expertise but not necessarily GDPR expertise. DPA cooperation ensures:

For providers, DPA cooperation means that sandbox plans involving significant personal data processing may be subject to DPA review — either proactively (as part of plan approval) or reactively (in response to complaints or incidents).

Member State Implementation Variation

Some Member States may require DPA pre-approval as part of the sandbox authorisation process. Others may rely on DPA consultation during plan review. Providers should check the specific requirements of the national sandbox authority in each Member State where they operate or intend to seek sandbox participation.

GDPR Interaction Deep Dive

Art.59 and GDPR Art.6(4) Compatibility

Art.59 supplements (but does not replace) the GDPR Art.6(4) compatibility analysis. A provider should conduct both:

Step 1: GDPR Art.6(4) analysis

Step 2: Art.59 exception

Providers should not skip Step 1. If the Art.6(4) analysis shows compatibility, Art.59 is not needed and the processing is simpler. Art.59 is the safety valve for cases where compatibility cannot be established.

Art.59 and Special Category Data (GDPR Art.9)

Art.59 does not create a new Art.9(2) basis for special category data. If sandbox training involves health data, biometric data, genetic data, or other Art.9 categories, the provider must establish an independent Art.9(2) basis — typically:

Art.59 then operates as the purpose extension mechanism for the further processing, while the Art.9(2) basis covers the special category dimension. Both must be satisfied.

Data TypeOriginal Collection BasisArt.59 Further ProcessingArt.9(2) Basis (if needed)
Medical recordsArt.6(1)(c) legal obligationArt.59 exceptionArt.9(2)(g) public interest
Biometric dataArt.6(1)(a) consentArt.59 exceptionArt.9(2)(a) explicit consent
Employment dataArt.6(1)(b) contractArt.59 exceptionArt.9(2)(b) employment law
Financial dataArt.6(1)(c) legal obligationArt.59 exceptionNot applicable (non-special)

CLOUD Act Jurisdiction Risk for Sandbox Training Data

Art.59(1)(f) prohibits transfers outside the EU — but the CLOUD Act creates a different risk: US authorities may compel production of data held on US cloud infrastructure even without a transfer.

The CLOUD Act Compellability Problem

If an EU provider uses AWS, Azure, or Google Cloud for sandbox AI training under Art.59, the training data — pseudonymised healthcare records, financial histories, biometric datasets — is potentially subject to CLOUD Act orders:

  1. US government serves CLOUD Act order on the US cloud provider (AWS/Microsoft/Google)
  2. Cloud provider must produce data held by it or under its control — including EU data centre data
  3. GDPR Art.48 conflict: GDPR prohibits transfers ordered by third-country authorities unless under MLAT or similar instruments
  4. Art.59(1)(f) violation: the production to a US authority constitutes a transfer outside the EU

The CLOUD Act risk is particularly acute for Art.59 data because:

Mitigation: EU-sovereign infrastructure with no US corporate parent is the cleanest Art.59 infrastructure solution. Providers using EU-only cloud providers (such as European PaaS platforms without US corporate affiliates) avoid the CLOUD Act compellability risk entirely for training workloads.

Practical Infrastructure Requirements for Art.59

Art.59 Compliant Infrastructure Architecture:

[Training Data Storage]
├── EU sovereign cloud or on-premises only
├── Encryption at rest (provider-managed keys, EU HSM)
├── Access logging: all reads auditable
└── No replication to non-EU regions

[Compute Environment]
├── EU-located training nodes
├── No telemetry to US-controlled cloud services
├── Isolated network: no outbound to third-country services
└── Ephemeral environments: clean-up after training complete

[Model Artefacts]
├── Weights stored in EU-sovereign storage
├── No export to non-EU evaluation or inference environments
└── If deployed: separate DPIA required before production use

[Access Control]
├── Role-based: only authorised ML engineers
├── Multi-factor authentication
├── No remote access from non-EU locations (or VPN with EU exit)
└── Third-party contractors: data processing agreements required

Python Implementation: Sandbox Data Processing Governance

"""
EU AI Act Art.59 Sandbox Data Processing Governance
Implements compliance tracking for further processing of personal data
in AI regulatory sandboxes under Art.59(1) conditions.
"""

from dataclasses import dataclass, field
from datetime import datetime, date
from enum import Enum
from typing import List, Optional, Dict
import hashlib
import json


class DataCategory(Enum):
    """GDPR data categories relevant to Art.59"""
    GENERAL = "general"
    SPECIAL_CATEGORY = "special_category"  # Art.9
    CRIMINAL = "criminal"  # Art.10


class OriginalPurposeLegalBasis(Enum):
    """GDPR Art.6(1) bases for original collection"""
    CONSENT = "art_6_1_a_consent"
    CONTRACT = "art_6_1_b_contract"
    LEGAL_OBLIGATION = "art_6_1_c_legal_obligation"
    VITAL_INTERESTS = "art_6_1_d_vital_interests"
    PUBLIC_TASK = "art_6_1_e_public_task"
    LEGITIMATE_INTERESTS = "art_6_1_f_legitimate_interests"


class Art9Basis(Enum):
    """GDPR Art.9(2) bases for special category data"""
    EXPLICIT_CONSENT = "art_9_2_a"
    EMPLOYMENT_LAW = "art_9_2_b"
    VITAL_INTERESTS = "art_9_2_c"
    LEGITIMATE_ACTIVITIES = "art_9_2_d"
    MANIFESTLY_PUBLIC = "art_9_2_e"
    LEGAL_CLAIMS = "art_9_2_f"
    PUBLIC_INTEREST = "art_9_2_g"
    PREVENTIVE_MEDICINE = "art_9_2_h"
    PUBLIC_HEALTH = "art_9_2_i"
    RESEARCH = "art_9_2_j"


@dataclass
class Art6_4CompatibilityAssessment:
    """
    GDPR Art.6(4) compatibility analysis.
    Must be completed before invoking Art.59 exception.
    """
    link_between_purposes: str          # Factor 1: link analysis
    context_of_collection: str          # Factor 2: data subject expectations
    nature_of_data_risk: str            # Factor 3: sensitivity assessment
    consequences_for_subjects: str      # Factor 4: impact analysis
    safeguards_in_place: str            # Factor 5: technical/org measures
    compatibility_conclusion: bool       # True = compatible (Art.59 not needed)
    compatibility_rationale: str
    assessed_by: str
    assessed_date: date

    def requires_art59(self) -> bool:
        """Returns True if Art.59 exception is needed"""
        return not self.compatibility_conclusion


@dataclass
class Art59Condition:
    """Single Art.59(1) condition assessment"""
    condition_id: str
    condition_text: str
    satisfied: bool
    evidence: str
    assessed_date: date


@dataclass
class SandboxDataProcessingRecord:
    """
    Art.59 compliance record for personal data processing
    in an AI regulatory sandbox.
    """
    record_id: str
    sandbox_reference: str              # Art.57 sandbox identifier
    sandbox_plan_reference: str         # Art.57(3) approved plan reference
    ai_system_description: str
    public_interest_justification: str  # Art.59(1)(a)

    # Data source characterisation
    data_source_description: str
    data_category: DataCategory
    original_collection_purpose: str
    original_legal_basis: OriginalPurposeLegalBasis
    art9_basis: Optional[Art9Basis] = None  # Required if special category

    # Subject scope
    approximate_subject_count: int = 0
    subject_population_description: str = ""

    # Art.6(4) assessment
    compatibility_assessment: Optional[Art6_4CompatibilityAssessment] = None

    # Art.59(1) conditions
    conditions: List[Art59Condition] = field(default_factory=list)

    # Processing boundaries
    processing_start: Optional[date] = None
    processing_end: Optional[date] = None
    sandbox_boundary_confirmed: bool = False  # Art.59(2)

    # Infrastructure
    processing_location: str = ""
    eu_sovereign_infrastructure: bool = False
    cloud_act_risk_assessed: bool = False

    # Subject rights mechanism
    erasure_mechanism_documented: bool = False
    notification_approach: str = ""

    # Approval
    dpa_consulted: bool = False
    competent_authority_approved: bool = False
    approved_date: Optional[date] = None

    def all_conditions_satisfied(self) -> bool:
        return all(c.satisfied for c in self.conditions)

    def compliance_summary(self) -> Dict:
        issues = []

        if not self.compatibility_assessment:
            issues.append("GDPR Art.6(4) compatibility assessment not completed")

        if self.data_category == DataCategory.SPECIAL_CATEGORY and not self.art9_basis:
            issues.append("Art.9(2) basis required for special category data")

        if not self.all_conditions_satisfied():
            failed = [c.condition_id for c in self.conditions if not c.satisfied]
            issues.append(f"Art.59(1) conditions not satisfied: {failed}")

        if not self.sandbox_boundary_confirmed:
            issues.append("Art.59(2) sandbox boundary not confirmed")

        if not self.eu_sovereign_infrastructure:
            issues.append("WARNING: Non-EU infrastructure may violate Art.59(1)(f)")

        if not self.cloud_act_risk_assessed:
            issues.append("CLOUD Act compellability risk assessment missing")

        if not self.erasure_mechanism_documented:
            issues.append("Data subject erasure mechanism not documented")

        if not self.competent_authority_approved:
            issues.append("Competent authority approval pending")

        return {
            "record_id": self.record_id,
            "compliant": len(issues) == 0,
            "issues": issues,
            "conditions_satisfied": self.all_conditions_satisfied(),
            "sandbox_boundary_ok": self.sandbox_boundary_confirmed,
            "infrastructure_ok": self.eu_sovereign_infrastructure,
        }


class Art59ConditionChecker:
    """
    Systematic checker for all six Art.59(1) conditions.
    """

    CONDITION_TEMPLATES = [
        ("C1", "AI system serves public interest in an Art.57(1) qualifying field"),
        ("C2", "Processing is necessary and proportionate to the AI development goal"),
        ("C3", "Processing produces no adverse effects for data subjects"),
        ("C4", "Appropriate technical and organisational measures are in place"),
        ("C5", "Processing is covered by the approved sandbox plan under Art.57(3)"),
        ("C6", "No transfer of personal data outside the EU/EEA"),
    ]

    @staticmethod
    def create_conditions_template(assessed_by: str) -> List[Art59Condition]:
        """Create blank condition records for assessment"""
        today = date.today()
        return [
            Art59Condition(
                condition_id=cid,
                condition_text=text,
                satisfied=False,
                evidence="[ASSESSMENT REQUIRED]",
                assessed_date=today,
            )
            for cid, text in Art59ConditionChecker.CONDITION_TEMPLATES
        ]

    @staticmethod
    def validate_record(record: SandboxDataProcessingRecord) -> List[str]:
        """Return list of compliance gaps"""
        gaps = []

        # Check all 6 conditions exist
        condition_ids = {c.condition_id for c in record.conditions}
        required = {cid for cid, _ in Art59ConditionChecker.CONDITION_TEMPLATES}
        missing = required - condition_ids
        if missing:
            gaps.append(f"Missing condition assessments: {missing}")

        # Check all satisfied
        if not record.all_conditions_satisfied():
            failed = [
                f"{c.condition_id}: {c.condition_text}"
                for c in record.conditions if not c.satisfied
            ]
            for f in failed:
                gaps.append(f"Condition not satisfied: {f}")

        return gaps


class SandboxDataSubjectTracker:
    """
    Tracks data subject contributions to sandbox training.
    Enables erasure right compliance under GDPR Art.17.
    """

    def __init__(self, record_id: str):
        self.record_id = record_id
        self.subject_contributions: Dict[str, List[str]] = {}
        # Maps pseudonymised_id -> list of training_batch_ids

    def register_contribution(
        self,
        subject_pseudoid: str,
        batch_id: str,
    ) -> None:
        """Record that a subject's data was used in a training batch"""
        if subject_pseudoid not in self.subject_contributions:
            self.subject_contributions[subject_pseudoid] = []
        self.subject_contributions[subject_pseudoid].append(batch_id)

    def subject_in_training(self, subject_pseudoid: str) -> bool:
        """Check if a subject's data was used in any training batch"""
        return subject_pseudoid in self.subject_contributions

    def get_affected_batches(self, subject_pseudoid: str) -> List[str]:
        """Get all training batches that included the subject's data"""
        return self.subject_contributions.get(subject_pseudoid, [])

    def generate_erasure_impact(self, subject_pseudoid: str) -> Dict:
        """
        Assess erasure request impact for a subject.
        Returns remediation requirements.
        """
        batches = self.get_affected_batches(subject_pseudoid)
        if not batches:
            return {
                "subject_in_training": False,
                "erasure_scope": "data_only",
                "retraining_required": False,
            }

        return {
            "subject_in_training": True,
            "training_batches": batches,
            "erasure_scope": "data_and_model_weights_review",
            "retraining_required": True,
            "retraining_rationale": (
                "Subject's data contributed to model training. "
                "Model must be retrained without subject data, or "
                "differential privacy techniques applied to demonstrate "
                "subject data has been effectively forgotten."
            ),
        }


@dataclass
class Art59AuditLog:
    """
    Immutable audit log for Art.59 data processing.
    """
    timestamp: datetime
    record_id: str
    action: str
    actor: str
    data_accessed: str
    purpose: str
    ip_location: str

    def to_log_entry(self) -> str:
        return json.dumps({
            "ts": self.timestamp.isoformat(),
            "record": self.record_id,
            "action": self.action,
            "actor": self.actor,
            "data": self.data_accessed,
            "purpose": self.purpose,
            "location": self.ip_location,
        })


# Example: Healthcare AI development under Art.59
def example_healthcare_ai_art59():
    record = SandboxDataProcessingRecord(
        record_id="ART59-2025-001",
        sandbox_reference="DE-SANDBOX-BfArM-2025-07",
        sandbox_plan_reference="SANDBOX-PLAN-v2.1-APPROVED-2025-08-01",
        ai_system_description=(
            "High-risk AI system for early detection of diabetic retinopathy "
            "in ophthalmology imaging (Annex III Category 5b: essential services)"
        ),
        public_interest_justification=(
            "Diabetic retinopathy is the leading cause of preventable blindness in the EU. "
            "Early AI-assisted detection enables treatment before irreversible damage. "
            "System serves public health interest under Art.57(1)."
        ),
        data_source_description=(
            "Retinal imaging records from University Hospital clinical database, "
            "originally collected under patient treatment consent and legal obligation "
            "to maintain medical records."
        ),
        data_category=DataCategory.SPECIAL_CATEGORY,
        original_collection_purpose="Medical treatment and clinical record-keeping",
        original_legal_basis=OriginalPurposeLegalBasis.LEGAL_OBLIGATION,
        art9_basis=Art9Basis.PUBLIC_INTEREST,  # Art.9(2)(g)
        approximate_subject_count=12000,
        subject_population_description="Adult diabetic patients in hospital care 2020-2024",
        processing_location="Frankfurt, Germany (EU sovereign colocation)",
        eu_sovereign_infrastructure=True,
        cloud_act_risk_assessed=True,
        erasure_mechanism_documented=True,
        notification_approach=(
            "Art.14(5)(b) disproportionate effort: 12,000 historical patients. "
            "Public notice published on hospital website and data register."
        ),
        dpa_consulted=True,
        competent_authority_approved=True,
        approved_date=date(2025, 8, 1),
        sandbox_boundary_confirmed=True,
        processing_start=date(2025, 8, 15),
        processing_end=date(2026, 2, 14),
    )

    # Add compatibility assessment
    record.compatibility_assessment = Art6_4CompatibilityAssessment(
        link_between_purposes=(
            "AI development purpose (diabetic retinopathy detection) is closely linked "
            "to the original treatment purpose — both aim at improving patient outcomes."
        ),
        context_of_collection=(
            "Data collected in clinical context. Patients would reasonably expect "
            "their data to be used for improving healthcare quality, including AI tools."
        ),
        nature_of_data_risk=(
            "Special category health data (Art.9). High sensitivity requires "
            "strict pseudonymisation and access controls."
        ),
        consequences_for_subjects=(
            "No direct decisions are taken about training subjects. "
            "Model does not feed back into clinical records of contributing patients."
        ),
        safeguards_in_place=(
            "Pseudonymisation with separate key management. "
            "Encryption at rest and in transit. Role-based access. "
            "Audit logging. No transfer outside EU."
        ),
        compatibility_conclusion=False,  # Not compatible without Art.59
        compatibility_rationale=(
            "AI model training is a new purpose not within scope of original treatment "
            "purpose. Art.59 exception required."
        ),
        assessed_by="DPO, Chief Privacy Officer",
        assessed_date=date(2025, 7, 15),
    )

    # Add Art.59(1) conditions
    record.conditions = [
        Art59Condition("C1", "AI system serves public interest in Art.57(1) qualifying field",
                      True, "Diabetic retinopathy detection: preventable blindness prevention", date(2025, 7, 15)),
        Art59Condition("C2", "Processing necessary and proportionate",
                      True, "Clinical images necessary for training; 12,000 subjects proportionate to accuracy target", date(2025, 7, 15)),
        Art59Condition("C3", "No adverse effects for data subjects",
                      True, "Model not deployed for clinical decisions on training subjects; pseudonymised", date(2025, 7, 15)),
        Art59Condition("C4", "Appropriate TOMs in place",
                      True, "Pseudonymisation, encryption, RBAC, audit logs, EU infrastructure documented", date(2025, 7, 15)),
        Art59Condition("C5", "Processing covered by approved sandbox plan",
                      True, "SANDBOX-PLAN-v2.1 Section 4: Data Processing, approved BfArM 2025-08-01", date(2025, 8, 1)),
        Art59Condition("C6", "No transfer outside EU",
                      True, "Frankfurt colocation, no US cloud, no third-country transfer", date(2025, 7, 15)),
    ]

    summary = record.compliance_summary()
    print(f"Art.59 Compliant: {summary['compliant']}")
    if summary['issues']:
        for issue in summary['issues']:
            print(f"  ISSUE: {issue}")
    return record

Art.59 vs General GDPR Research Exemptions

GDPR Art.89 provides a research and statistics processing framework with derogations from certain data subject rights. How does Art.89 relate to Art.59?

DimensionGDPR Art.89EU AI Act Art.59
ScopeScientific/historical research, statisticsAI development/testing in regulatory sandbox
EnvironmentAnyAI regulatory sandbox only
Rights derogationsArt.15/16/18/21 may be restrictedNo derogation — all rights maintained
Purpose linkResearch purpose established independentlyAI development in public interest
Supervisory oversightDPA oversightJoint: sandbox authority + DPA (Art.59(4))
Transfer prohibitionNot explicitExplicit: Art.59(1)(f)
Model deploymentN/ASeparate assessment required

Art.59 is narrower than Art.89 in some respects (no rights derogations, sandbox-only) but more enabling in others (targeted exception for purpose extension, multi-authority oversight framework). For AI development that does not occur within a formal sandbox, Art.89 remains relevant where the development has a genuine research character.

Art.59 Compliance Checklist (35 Items)

Pre-Processing: Legal Foundation

Sandbox Plan Requirements

Art.59(1) Conditions — All Six

Sandbox Boundary (Art.59(2))

Data Subject Rights

CLOUD Act and Infrastructure

Common Mistakes

Mistake 1: Invoking Art.59 without sandbox membership Art.59 only applies within a formally established Art.57 AI regulatory sandbox. There is no standalone Art.59 exception for AI development outside the sandbox. Providers who process personal data for AI development outside a sandbox must establish a standard GDPR basis.

Mistake 2: Skipping the Art.6(4) compatibility assessment If further processing is actually compatible under Art.6(4), Art.59 is not needed and the simpler route is to document compatibility. Only invoke Art.59 when the Art.6(4) analysis shows incompatibility.

Mistake 3: Treating Art.59 as a suspension of GDPR rights Data subjects retain all GDPR rights during Art.59 processing. Art.59 does not restrict access, rectification, erasure, or objection rights. Providers must have procedures for exercising these rights during sandbox processing.

Mistake 4: Deploying a sandbox-trained model without a separate DPIA Art.59 covers the training/testing phase. Deploying a model trained on Art.59 data in production requires a separate legal analysis, including a DPIA that addresses training data memorisation risk.

Mistake 5: Using US cloud infrastructure for Art.59 processing Art.59(1)(f) prohibits transfers outside the EU. CLOUD Act compellability orders against US cloud providers are a realistic risk for sensitive training data. EU-sovereign infrastructure is the compliant choice.

Mistake 6: Omitting Art.9(2) basis for special category training data Art.59 supplements but does not replace Art.9(2). If training data includes health, biometric, genetic, or other special category data, a separate Art.9(2) basis must be established.

Key Takeaways

  1. Art.59 is sandbox-bound: it only applies within a formally established Art.57 AI regulatory sandbox — not for general AI development.

  2. Six conditions are cumulative: all must be satisfied. A single failed condition invalidates the Art.59 exception.

  3. Prior authorisation is required: the data processing must be in the sandbox plan approved by the competent authority. Retroactive invocation is not possible.

  4. Data subject rights survive: Art.59 does not suspend any GDPR rights. Erasure requests require machine unlearning procedures.

  5. Deployment requires a new legal assessment: Art.59 covers training, not production use. Models trained on Art.59 data need a separate DPIA before deployment.

  6. EU-sovereign infrastructure is the safe choice: Art.59(1)(f) prohibits transfers outside the EU, and CLOUD Act compellability creates a risk on US cloud infrastructure even without explicit transfer.

  7. DPAs retain authority: Art.59(4) requires DPA cooperation. DPAs can review and can enforce GDPR regardless of sandbox participation.

See Also