sota.io
2026-04-14·13 min read·sota.io team

EU AI Act Art.102: Penalties for Natural Persons — Member State Criminal Sanctions Developer Guide (2026)

Articles 99, 100, and 101 of the EU AI Act establish administrative fine frameworks for organizations — providers, deployers, importers, EU institutions, and GPAI model providers. They specify maximum fines, calculation factors, and the enforcement authorities (national MSAs, EDPS, AI Office) that impose those fines. But every fine in Art.99-101 lands on a legal entity. Article 102 is different: it mandates penalties against individuals.

Article 102 requires each EU Member State to lay down rules on penalties — including criminal sanctions — applicable to natural persons who infringe the Regulation in conduct not already covered by Art.99-101. This provision closes a gap that Art.99 deliberately leaves open: organizational fines do not automatically transfer to the individuals who made the decisions that caused the violation.

For developers, data scientists, compliance officers, product managers, and engineering leaders, Art.102 is the provision that converts corporate AI Act liability into personal professional and criminal risk. Understanding Art.102 is not optional once you are making decisions about AI system design, deployment configuration, oversight mechanisms, or documentation completeness.

What Article 102 Actually Says

Article 102 is a delegation provision. It does not itself establish specific penalty amounts or criminal offence definitions — instead, it mandates Member States to create those rules at the national level. The provision requires that:

  1. Member States shall lay down rules on penalties applicable to infringements of the Regulation that are not already subject to Art.99, Art.100, or Art.101
  2. Penalties shall include criminal sanctions — Member States cannot limit their Art.102 implementation to administrative fines only
  3. Penalties shall be effective, proportionate, and dissuasive — the same standard applied to corporate fines under Art.99
  4. Member States shall notify the Commission of their adopted penalty rules by the applicable date

The key phrase is "not already subject to Art.99, 100 or 101." This means Art.102 fills three gaps:

Who Is a "Natural Person" Under Art.102?

The natural persons most exposed under Art.102 are individuals whose professional decisions directly affect AI Act compliance outcomes:

Software Engineers and ML Engineers: Engineers who implement prohibited practices — manipulation techniques, exploitation of vulnerabilities under Art.5 — or who deliberately disable required safety features create direct personal exposure. If an organization is fined under Art.99 for a prohibited practice violation and the engineer who implemented that feature made the decision knowing it was prohibited, Art.102 creates a parallel national-level enforcement pathway against that individual.

Product Managers and AI System Owners: Product managers who define AI system scope, configure intended purposes, or override safety constraints are responsible for classification decisions. A product manager who knowingly classifies a high-risk AI system as non-high-risk to avoid conformity assessment obligations — and that misclassification is later discovered — faces Art.102 exposure at the national level.

Compliance Officers and Legal Counsel: Compliance officers who certify conformity documentation they know to be incomplete or inaccurate face Art.102 exposure beyond the organizational fine under Art.99. Legal counsel who advises providers to misrepresent AI system capabilities to national competent authorities compound their client's Art.99 liability with their own potential Art.102 exposure.

CTOs, VPs of Engineering, and C-Suite: Executive-level individuals who set AI development policy, approve deployment decisions for high-risk AI systems, or determine compliance resource allocation are decision-makers in the enforcement chain. Art.102's criminal sanction pathway is specifically designed to reach individuals at this level — the same individuals who are otherwise shielded by corporate liability structures.

Third-Party Auditors and Notified Bodies: Auditors who issue conformity certificates for systems they know do not meet EU AI Act requirements face national criminal exposure under Art.102. The Regulation's conformity assessment architecture depends on auditor integrity — Art.102 is the enforcement backstop when that integrity fails.

The Art.102 vs Art.99 Coverage Matrix

Art.99 and Art.102 are complementary, not duplicative. Understanding what each covers — and how they interact — determines personal liability exposure:

DimensionArt.99Art.102
Who is liableLegal entities: providers, deployers, importers, authorised representativesNatural persons: engineers, managers, compliance officers, executives
Who enforcesNational market surveillance authorities (NCA/MSA)National courts and criminal prosecutors (per Member State law)
Fine typeAdministrative: up to €35M or 7% global turnoverAdministrative + criminal (per Member State rules)
Maximum penalty€35M / 7% turnover (prohibited practices)Varies by Member State — criminal sanctions include imprisonment
CoordinationEU AI Office coordinates via EAIBMember State criminal procedure rules
Double jeopardyMay run in parallel with Art.102May run in parallel with Art.99 against same organization
Applicable dateAugust 2, 2026 (general application)Member State implementation deadline (same)

Key interaction: Art.99 and Art.102 can apply simultaneously. An organization found by its national MSA to have deployed a prohibited AI practice (Art.99 fine up to €35M) can simultaneously face Art.102 proceedings against the individuals who implemented and approved that practice. The organizational fine does not absorb or satisfy the individual's criminal exposure.

Member State Implementation Variation

Art.102 is a mandate to Member States, not a harmonized rule. The actual penalty rules — including whether a criminal offence requires intent, negligence, or strict liability; what the maximum imprisonment term is; and whether corporate compliance programs constitute a defence — differ by jurisdiction.

High-enforcement jurisdictions (Germany, France, Netherlands): These Member States have strong existing frameworks for corporate criminal liability and are likely to implement Art.102 provisions with meaningful criminal penalty thresholds. Germany's existing data protection enforcement culture (DSGVO/GDPR) suggests aggressive Art.102 implementation.

Civil law jurisdictions with criminal corporate prosecution tradition: France (where legal persons can already face criminal liability under the Code Pénal) and Italy will likely extend their existing frameworks to cover AI Act individuals.

Common law jurisdictions (Ireland): Ireland's implementation matters disproportionately because many US tech companies are headquartered there for EU purposes. The Irish Art.102 rules will govern individual liability for engineers and executives at these companies for their EU AI Act conduct.

Practical consequence: If you are a developer making AI compliance decisions for a product deployed across the EU, you may simultaneously be subject to multiple Member State Art.102 frameworks. Your primary exposure is in the Member State where you work and where the affected users are located — but cross-border enforcement under mutual legal assistance treaties is possible for serious criminal matters.

Criminal Sanction Risk for Developers: What Triggers Prosecution

Not every Art.99 organizational violation creates Art.102 criminal risk for individuals. Criminal prosecution requires:

1. Intent or gross negligence: Most Member State criminal frameworks require that the individual either knew their conduct was prohibited and continued anyway (intent) or failed to take obvious precautions that a reasonable professional would have taken (gross negligence). A developer who makes a genuine mistake in good faith based on legal advice has a substantially lower criminal risk than a developer who deliberately circumvents a required safety feature.

2. Prohibited practice involvement: Art.5 violations — deploying AI that manipulates people through subliminal techniques, exploits vulnerabilities of specific groups, conducts social scoring, or enables real-time remote biometric identification — create the highest criminal exposure under Art.102. These are not technical compliance failures; they are deliberate capability deployments that the Regulation prohibits outright.

3. False certification: Knowingly certifying false documentation — whether as an internal compliance officer, an external auditor, or a technical lead signing off on conformity assessment — creates criminal exposure in virtually every Member State's intended Art.102 implementation. False certification combines the organizational Art.99 violation with individual criminal conduct.

4. Obstruction: Individuals who obstruct national MSA investigations — hiding documentation, providing misleading answers during Art.88-92 proceedings, or instructing others to withhold information — face Art.102 criminal exposure for the obstruction conduct, independent of the underlying organizational violation.

5. Whistleblower retaliation: Under Directive 2019/1937 (Whistleblower Protection Directive), individuals who retaliate against employees who report AI Act violations to national authorities may face both Directive 2019/1937 civil liability and Art.102 criminal exposure for the retaliation conduct as an AI Act-related infringement.

Personal Liability Documentation Strategy

The most effective protection against Art.102 criminal exposure is professional conduct documentation that establishes good faith compliance intent — created before any investigation opens, not as a reactive measure.

Document your compliance escalations: When you identify a potential AI Act compliance issue and escalate it to management, document that escalation in writing. A compliance concern email with a timestamp is powerful evidence that you raised the issue; if the organization chose to proceed despite your concern, the decision was organizational, not individual.

Record your legal advice reliance: If you received legal counsel that a specific AI capability was compliant with the Regulation, document that advice and your reliance on it. Reliance on qualified legal advice is a recognized defence to criminal intent claims in most jurisdictions.

Maintain your technical decision records: For high-risk AI systems, document why specific safety features were implemented the way they were. Architecture decision records (ADRs) that show considered compliance trade-offs are evidence of professional diligence, not concealment.

Create personal compliance records separate from organizational records: Your individual compliance documentation should not be stored exclusively on systems your employer controls. A personal record of your professional compliance conduct — stored on personal hardware or EU-sovereign personal cloud storage — cannot be deleted by an employer to protect the organization at your expense.

The CLOUD Act intersection: If your professional compliance records are stored on a US cloud provider's infrastructure — Microsoft 365, Google Workspace, iCloud — they are subject to US government access under the CLOUD Act independently of any EU Art.102 investigation. In a scenario where the US and EU are coordinating investigation of an AI Act violation with cross-border dimensions, your personal compliance records may reach both authorities through separate legal channels before you have an opportunity to present them voluntarily.

Art.102 and the Employment Law Intersection

Art.102 creates a structural tension in the employment relationship for AI professionals. An organization facing an Art.99 investigation has an incentive to identify individuals who made the non-compliant decisions — not to protect those individuals, but to shift narrative responsibility away from organizational policy toward individual misconduct.

What this means in practice:

D&O insurance for AI decisions: Directors and Officers insurance policies are beginning to address AI regulatory liability explicitly. If you are in a management role making AI deployment decisions, verify that your D&O coverage addresses EU AI Act enforcement proceedings — both the cost of defence and the potential penalty liability.

Python Tooling for Individual Liability Assessment

from dataclasses import dataclass, field
from typing import Optional
from datetime import date

@dataclass
class Art102IndividualProfile:
    """
    Individual natural person Art.102 exposure profile.
    Maps professional role to personal liability risk dimensions.
    """
    name: str
    role: str  # e.g., "Senior ML Engineer", "CTO", "Compliance Officer"
    member_state: str  # Primary jurisdiction for Art.102 exposure
    
    # AI system involvement
    prohibited_practice_involvement: bool = False  # Art.5 exposure
    high_risk_classification_decisions: bool = False  # Annex III scope decisions
    conformity_documentation_sign_off: bool = False  # False cert exposure
    msa_investigation_interactions: bool = False  # Obstruction risk
    safety_feature_override_authority: bool = False  # Override documentation
    
    # Mitigation evidence
    compliance_escalations_documented: int = 0  # Number of documented escalations
    legal_advice_relied_upon: bool = False  # Qualified legal counsel engaged
    adr_records_maintained: bool = False  # Architecture decision records
    personal_compliance_log: bool = False  # Separate from employer systems
    
    def criminal_risk_score(self) -> str:
        """
        Returns HIGH/MEDIUM/LOW Art.102 criminal sanction risk.
        Reflects exposure under most Member State implementations.
        """
        score = 0
        if self.prohibited_practice_involvement:
            score += 40  # Art.5 violations: highest criminal exposure
        if self.conformity_documentation_sign_off:
            score += 25  # False certification: universal criminal risk
        if self.msa_investigation_interactions:
            score += 20  # Obstruction: direct criminal conduct
        if self.high_risk_classification_decisions:
            score += 10  # Classification error: lower individual exposure
        if self.safety_feature_override_authority:
            score += 15  # Override without documentation: high exposure
        
        # Mitigations
        if self.legal_advice_relied_upon:
            score -= 20
        if self.compliance_escalations_documented >= 3:
            score -= 15
        if self.adr_records_maintained and self.personal_compliance_log:
            score -= 10
        
        if score >= 50:
            return "HIGH"
        elif score >= 25:
            return "MEDIUM"
        else:
            return "LOW"
    
    def priority_protective_actions(self) -> list[str]:
        """Returns ordered list of priority actions to reduce Art.102 exposure."""
        actions = []
        if self.prohibited_practice_involvement:
            actions.append("CRITICAL: Document Art.5 compliance analysis for each implicated capability. Obtain legal opinion.")
        if self.conformity_documentation_sign_off and not self.legal_advice_relied_upon:
            actions.append("HIGH: Ensure all signed conformity documents were reviewed by qualified legal counsel before signature.")
        if not self.personal_compliance_log:
            actions.append("HIGH: Create personal compliance record on EU-sovereign personal storage, separate from employer systems.")
        if self.compliance_escalations_documented < 3:
            actions.append("MEDIUM: Begin documenting compliance escalations in writing. Create paper trail of concerns raised.")
        if not self.adr_records_maintained:
            actions.append("MEDIUM: Establish Architecture Decision Records for AI compliance decisions. Include Art. references.")
        return actions


@dataclass  
class Art102MemberStateVariance:
    """
    Tracks Art.102 implementation variance across key EU jurisdictions.
    Update as Member States publish their implementing rules.
    """
    member_state: str
    criminal_sanctions_available: bool
    max_imprisonment_years: Optional[int]
    intent_requirement: str  # "intent", "negligence", "strict"
    corporate_compliance_defence: bool
    whistleblower_protection_aligned: bool  # Directive 2019/1937
    implementation_date: Optional[date]
    notes: str


# Reference implementations (update as Member States publish rules)
known_implementations = [
    Art102MemberStateVariance(
        member_state="Germany",
        criminal_sanctions_available=True,
        max_imprisonment_years=3,
        intent_requirement="intent",
        corporate_compliance_defence=True,
        whistleblower_protection_aligned=True,
        implementation_date=None,  # Pending publication
        notes="Strong GDPR enforcement culture. OWiG administrative offence framework + StGB criminal. Compliance programs (Compliance Management Systems) reduce individual exposure."
    ),
    Art102MemberStateVariance(
        member_state="France",
        criminal_sanctions_available=True,
        max_imprisonment_years=5,
        intent_requirement="intent",
        corporate_compliance_defence=True,
        whistleblower_protection_aligned=True,
        implementation_date=None,
        notes="Code Pénal already covers legal and natural person criminal liability. Sapin II compliance frameworks reduce exposure."
    ),
    Art102MemberStateVariance(
        member_state="Ireland",
        criminal_sanctions_available=True,
        max_imprisonment_years=None,  # TBD in implementing legislation
        intent_requirement="negligence",
        corporate_compliance_defence=True,
        whistleblower_protection_aligned=True,
        implementation_date=None,
        notes="Key jurisdiction: US Big Tech EU HQ. Companies Acts directors' liability framework will likely extend to AI Act officers."
    ),
]


def assess_cross_border_exposure(
    member_states_of_activity: list[str],
    has_us_cloud_personal_records: bool,
    role_seniority: str  # "developer", "manager", "executive"
) -> dict:
    """
    Returns Art.102 cross-border exposure assessment for multi-jurisdiction operators.
    """
    exposure = {
        "primary_jurisdiction_count": len(member_states_of_activity),
        "cloud_act_personal_record_risk": has_us_cloud_personal_records,
        "mutual_legal_assistance_risk": len(member_states_of_activity) > 2,
        "highest_exposure_profile": role_seniority,
        "recommendation": ""
    }
    
    if role_seniority == "executive" and len(member_states_of_activity) > 3:
        exposure["recommendation"] = "Engage specialist EU AI Act criminal defence counsel proactively. Cross-border Art.102 exposure at executive level warrants D&O audit."
    elif has_us_cloud_personal_records:
        exposure["recommendation"] = "Migrate personal compliance records to EU-sovereign personal storage before any Art.102-related investigation opens. CLOUD Act reach to personal records is a compounding risk."
    else:
        exposure["recommendation"] = "Document compliance role boundaries and escalation records. Individual exposure is manageable with proper documentation."
    
    return exposure

The Art.102 Enforcement Sequence

Unlike Art.99's clear MSA-driven investigation procedure (Art.88-94), Art.102 enforcement follows Member State criminal procedure. There is no single EU-level sequence. However, the typical sequence involves:

Art.99 organizational investigation opens (national MSA or AI Office)
        ↓
Organization identifies individuals involved in non-compliant conduct
        ↓ (organization may cooperate with authorities by naming individuals)
National criminal prosecutor assesses Art.102 referral
        ↓ (if criminal threshold met: intent, false certification, obstruction)
Art.102 criminal investigation opens against named individual(s)
        ↓
Member State criminal procedure (arrest, charge, trial)
        ↓
Criminal penalty: fine, suspended sentence, imprisonment, or professional prohibition

The Art.99 investigation creates Art.102 exposure. When a national MSA opens an Art.99 investigation against an organization, the investigation process includes requests for documentation, interviews of employees (voluntarily or under Art.92 adapted procedures), and inspection of records. The individuals interviewed during an Art.99 investigation may simultaneously be creating their Art.102 record. This is the mechanism by which organizational enforcement converts to individual criminal risk.

Professional prohibition: Beyond fines and imprisonment, Art.102 Member State implementations may include professional prohibition orders — prohibitions on serving as a director, officer, or AI system compliance authority for a period of years. For AI professionals, this is potentially more career-damaging than a monetary penalty.

30-Item Personal AI Act Compliance Checklist

Role and Authority Documentation (1–8):

Compliance Escalation Records (9–16):

Personal Record Infrastructure (17–22):

Criminal Risk Reduction (23–30):

Summary

Article 102 is where the EU AI Act stops being a corporate compliance matter and becomes a personal professional risk. Art.99 fines land on organizations; Art.102 penalties — including criminal sanctions — land on the individuals who made the decisions behind those organizations' violations.

The Regulation delegates Art.102 implementation to Member States, creating a patchwork of national criminal frameworks that will evolve over time. What will not vary is the fundamental principle: natural persons who knowingly implement prohibited practices, certify false documentation, obstruct MSA investigations, or retaliate against whistleblowers face individual penalties beyond whatever fine their employer pays under Art.99.

For developers and AI professionals, the Art.102 protection strategy is documentation-first: escalate concerns in writing, rely on documented legal advice, maintain personal compliance records on EU-sovereign storage, and ensure your role boundaries are clear before any investigation opens. The individuals with the highest Art.102 risk are not those who make good-faith mistakes — they are those who make consequential compliance decisions without creating any paper trail of their reasoning.

Art.102 is not yet fully implemented across EU Member States. Watch for national implementing legislation in the primary Member States where your AI systems are deployed, and re-evaluate your individual exposure profile as those rules become clear.