Módulo 5: GDPR for AI Systems
Legitimate Interest vs Consent
Descripción
En M5-05 tocamos brevemente la elección entre consent y legitimate interest. Esta cápsula profundiza: cuándo cada uno aplica, cómo hacer el balancing test (LIA — Legitimate Interest Assessment) defensible, y errores que sancionan.
Al terminar vas a poder:
- Realizar un balancing test para legitimate interest
- Decidir entre consent y LI para AI features específicas
- Documentar el LIA defendible en audit
- Identificar cuándo LI NO es apropiado (deber recurrir a consent)
El balancing test (LIA)
Legitimate interest requiere demostrar que tu interés legítimo no override los rights del data subject. Es un test de 3 pasos:
Step 1: Purpose Test
¿Tu purpose es legitimate?
- ✅ Legítimo: fraud prevention, security, internal analytics aggregated, R&D
- ⚠️ Borderline: marketing, ad personalization
- ❌ No legítimo: discriminación, manipulación, sale to third parties sin disclosure
Step 2: Necessity Test
¿Es necesario procesar para lograr ese purpose? ¿O hay manera less intrusive?
Pregunta clave: ¿podrías lograr el purpose con menos datos o sin processing? Si sí, NO es necesario, fail.
Step 3: Balancing Test
¿Tu interés legítimo es proporcional al impacto sobre el usuario?
Factores que pesan a favor del data subject:
- Datos sensibles (salud, religious, sexual orientation)
- Procesamiento que el sujeto no esperaba
- Impactos en derechos fundamentales
- Children o vulnerable groups
Factores que pesan a favor del controller (vos):
- Beneficios para el data subject
- Public interest
- Expectativas razonables del usuario
- Safeguards implementados (encryption, access controls)
El template del LIA
# Legitimate Interest Assessment — Fraud Detection AI
**Date**: 2026-05-11
**Reviewer**: [Tech Lead + DPO]
## Purpose
Detect fraudulent transactions automatically using ML model trained on
historical transaction data + current transaction features.
## Lawful basis claim: Legitimate Interest (Art. 6(1)(f) GDPR)
## Step 1: Purpose Test
✅ LEGITIMATE
- Fraud prevention is legitimate business interest
- Protects both us (financial loss) and users (account safety)
- Common industry practice; users expect this
## Step 2: Necessity Test
✅ NECESSARY
- Manual review would be impossibly slow at our scale (50K txns/day)
- Rule-based detection has high false negative rate (35%)
- AI model achieves 90% detection with 5% false positive
- No less-intrusive alternative achieves required performance
## Step 3: Balancing Test
### Impact on data subject:
- Transactions reviewed automatically (low impact, no decision yet)
- High-risk transactions flagged for human review
- Account holds applied only after human approval
### Safeguards implemented:
- Model trained on anonymized data
- Decisions logged with explainability features
- Right to challenge: manual review within 24h
- No sale or sharing with third parties
### Expectations:
- Users informed in privacy policy
- Industry standard practice
- Reasonable expectation given service
### Conclusion:
Legitimate interest CAN be claimed. Impact on data subject is
proportionate and balanced by safeguards. User retains right to
object via [process link].
## Documentation
- Privacy policy updated 2025-01-15 with fraud AI disclosure
- Internal procedure for handling Art. 21 objection requests
- DPIA conducted (M7) and approved
- Annual review scheduled for 2027-01-15
Decision framework: consent vs LI
| Situation | Recommended basis |
|---|---|
| Marketing emails | Consent (explicit opt-in) |
| Personalized ads | Consent (granular) |
| Fraud detection | Legitimate Interest |
| Security monitoring | Legitimate Interest |
| Internal analytics aggregated | Legitimate Interest |
| AI training on user data — service improvement | Either (LI possible if anonymized + balancing test passes) |
| AI training — model commercialized to third parties | Consent (necessary for commercial use) |
| Profile inference for monetization | Consent |
| Recommended products to user (own data) | Either (LI if reasonable expectation) |
| Sensitive data processing (health, etc.) | Explicit Consent (Art. 9 requires it) |
Sensitive data: special rules (Art. 9)
Datos "special category" requieren excepciones específicas del Art. 9:
- Health
- Religious beliefs
- Sexual orientation
- Political opinions
- Race/ethnicity
- Biometric (identification)
- Genetic
Para estos, legitimate interest NO basta. Necesitás:
- Explicit consent
- Necessary for employment/social security obligations
- Necessary for vital interests
- Necessary for public interest
- Etc.
AI implication: si tu modelo procesa o infiere sensitive data (e.g., facial recognition que infiere ethnicity, NLP que infiere political opinions), necesitás explicit consent o otra excepción específica del Art. 9.
Right to object — el límite del legitimate interest
Aunque uses LI, el data subject tiene right to object (Art. 21).
User: "Quiero que paren de procesar mis datos para fraud detection."
Tu respuesta:
- Si LI legitimate basis y no hay compelling override → debés parar
- "Compelling legitimate grounds" para continuar: legal obligations,
security threat documented, etc.
- Para direct marketing, objection es absoluto (no podés override)
Para AI features bajo LI: tenés que implementar mechanism para procesar objection requests. SLA típico: 1 mes.
Cuando combinás bases legales
Un sistema AI a menudo tiene múltiples processing activities con bases legales distintas:
## Processing Activities — AI Customer Support Bot
| Activity | Lawful Basis | Notes |
|----------|--------------|-------|
| User authentication | Contract (Art. 6(1)(b)) | Necessary for service |
| Storing chat history | Contract + LI (analytics) | LI for aggregated analytics |
| Real-time RAG retrieval | Contract | Necessary for service |
| Model improvement training | Consent (Art. 6(1)(a)) | Opt-in checkbox in settings |
| Fraud detection on misuse | Legitimate Interest | LIA documented |
| Profile inference for marketing | Consent | Separate granular consent |
Documentá cada activity con su basis. Diferentes activities pueden tener diferentes lawful bases.
Trampas comunes
Trampa 1 — LI por default sin LIA. Asumís "LI" porque parece más conveniente. Sin LIA documentado, regulador asume invalid.
Trampa 2 — Re-purposing data bajo nueva base. Collected con consent para marketing → ahora usás para AI training "bajo LI". Cambio de purpose requiere re-evaluation y posiblemente re-consent.
Trampa 3 — LI para sensitive data. Imposible. Sensitive data requiere Art. 9 specific basis.
Trampa 4 — Ignorar right to object. "Bajo LI, podés objetar" pero tu sistema no implementa mechanism. Violation.
Trampa 5 — LIA outdated. LIA hecho en 2018, sistema cambió mucho, LIA nunca actualizado. Required: revisión periódica (annual minimum).
Ejercicio
Para tu sistema de AI:
- Identificá 5-7 processing activities distintas
- Para cada una, decide consent vs LI vs otra base
- Para las que elegís LI, write un LIA brief siguiendo el template
- Para las de consent, especificá UX del consent flow
Ver solución (ejemplo)
Sistema: AI assistant en e-commerce
| Activity | Basis | Justification |
|---|---|---|
| User account auth | Contract | Necessary for service |
| Order processing | Contract | Necessary for service |
| Saving purchase history | Contract | Necessary to fulfill orders |
| Aggregated sales analytics | LI | Standard business analytics, anonymized |
| Fraud detection in checkout | LI | Industry standard, protects users |
| AI product recommendations (using user data) | Consent | Opt-in feature; user controls |
| Personalized marketing emails | Consent | Direct marketing |
| Inferring sensitive preferences (health products) | Explicit Consent | Sensitive data, Art. 9 |
LIA brief para fraud detection:
Purpose Test: ✅ (fraud prevention legitimate)
Necessity: ✅ (volume requires automation)
Balancing: ✅ (low individual impact, high safeguards)
- Decisions logged
- Right to object honored within 30 days
- Annual review scheduled
Consent UX para AI recommendations:
- Just-in-time when feature is first used
- Modal explaining what data is used + benefit
- Toggle in settings to revoke
- Persistent UI indicator when active
Resumen
Aprendiste:
- ✅ Los 3 pasos del balancing test (purpose, necessity, balancing)
- ✅ Template del LIA documentable
- ✅ Decision framework consent vs LI por situación
- ✅ Sensitive data requiere Art. 9 specific basis
- ✅ Right to object y mechanism para implementarlo
- ✅ Multi-activity systems: diferentes bases legales por activity
Checkpoint: si podés escribir un LIA defensible para cada activity de tu sistema, estás listo.
Siguiente cápsula
07 — Documentation y DPO en contextos AI. Qué documentar específicamente para AI compliance, y cuándo necesitás un Data Protection Officer.