Create a DPIA template for LLM assistants: data flows, prompt/response retention, human…
- User
Create a DPIA template for LLM assistants: data flows, prompt/response retention, human oversight, and mitigations.
- Assistant
Data Protection Impact Assessment (DPIA) Template for LLM Assistants
Document Control
- Assessment Title: DPIA for [LLM Assistant Name, e.g., "Sherlock AI Assistant"]
- Version: 1.0
- Date: [Insert Date]
- Prepared By: [Name/Role]
- Reviewed By: [Name/Role, e.g., DPO]
- Approved By: [Controller Representative]
- Review Frequency: Annually or upon significant changes (e.g., model updates, new features)
1. Description of Processing Activities
Provide a high-level overview of the LLM assistant's operations.
Field Description Purpose [e.g., Provide conversational AI assistance for [use case, e.g., customer support, research queries]] Legal Basis [e.g., Consent (Art. 6(1)(a)), Contract (Art. 6(1)(b)), Legitimate Interests (Art. 6(1)(f))] Data Controller [Organization Name, Contact Details] Data Processor(s) [e.g., LLM Provider (e.g., Trailblazer Labs), Cloud Host (e.g., AWS)] Categories of Data Subjects [e.g., End-users, Employees, Third-party queriers] Personal Data Categories [e.g., Prompts (may contain names, health data, opinions); Metadata (IP, timestamps); Responses (inferred PII)] Special Categories of Data [e.g., Health, Biometric, Political Opinions – if processed] Volume/Scale [e.g., X queries/day; Y GB storage] Automated Decision-Making [Yes/No – e.g., LLM inference; Profiling? High-risk ADM under Art. 22?]
2. Data Flows
Map the end-to-end data lifecycle for prompts, responses, and related data. Include a simple ASCII diagram or reference to a visual tool (e.g., Lucidchart).
2.1 Data Flow Diagram
User Device --> [Input: Prompt + Metadata (IP, Device ID)] --> Preprocessing (Anonymization? Filtering?) --> LLM Inference (Model: [e.g., GPT-4 equivalent]; Hosted on [Server/Cloud]) --> Postprocessing (Moderation, Formatting) --> Output: Response to User | +--> Logging/Storage (if retained) +--> Analytics/Aggregation (Pseudonymized?) +--> Third-Party Sharing (e.g., API to [Vendor]) +--> Deletion/Archival2.2 Detailed Data Flow Table
Stage Data Inputs Processing Data Outputs Transfers/Recipients Security Measures Ingestion User prompt, session ID, IP Validation, rate limiting Sanitized prompt Internal queue TLS 1.3 encryption Inference Sanitized prompt + context LLM generation Raw response N/A (on-device or isolated) GPU isolation, TEEs Output Raw response Moderation (e.g., toxicity filter) Final response User device End-to-end encryption Logging Full prompt/response/metadata Pseudonymization Logs Storage DB Access controls (RBAC) Analytics Aggregated logs Stats computation Reports Internal dashboards Aggregation only (no re-identification) Cross-Border Transfers: [Yes/No – e.g., EU-US via SCCs/Standard Contractual Clauses]
3. Prompt/Response Retention
Detail retention practices to ensure minimization (Art. 5(1)(e) GDPR).
Element Policy Justification Duration Storage Location Deletion Method Prompts [Retain/Do not retain] [e.g., Debugging, abuse detection] [e.g., 30 days; Indefinite for flagged content] [e.g., Encrypted DB in EU] Automated purge + audit logs Responses [Retain/Do not retain] [e.g., User-requested history] [e.g., User-deletable; 90 days default] [Same as prompts] Secure erase (e.g., NIST 800-88) Metadata [Always retain anonymized] [Analytics, compliance] [e.g., 1 year] [Pseudonymized store] Irreversible anonymization Training Data [Never used for fine-tuning unless opted-in] N/A N/A N/A N/A User Controls [Delete history, opt-out of retention] Right to erasure (Art. 17) N/A N/A On-request processing Retention Exceptions: [e.g., Legal holds for disputes]
4. Human Oversight
Describe mechanisms for human-in-the-loop to mitigate AI risks (e.g., hallucinations, biases).
Oversight Type Process Triggers Responsible Party Frequency/Scale Documentation Pre-Deployment Model testing, red-teaming New versions AI Ethics Team Per release Test reports Real-Time Flagging high-risk prompts (e.g., sensitive PII) Keywords, confidence scores <0.8 Moderation Queue 95% automated; 5% human Audit trail Post-Output User feedback loops, random sampling Complaints, anomalies Support Team 1% of interactions Escalation logs Training Annual GDPR/AI ethics training All handlers HR 100% coverage Certificates Escalation Human veto on responses Legal/ethical flags DPO/Senior Review Ad-hoc Signed decisions Oversight Metrics: [e.g., Human review rate: <5%; Resolution time: <24h]
5. Risk Assessment
Identify high risks to data subjects' rights/freedoms (template uses DREAD-inspired scoring: Likelihood x Impact = Risk Score; High = 15+).
Risk ID Risk Description Likelihood (1-5) Impact (1-5) Score Source (e.g., Prompt injection, Data leak) R1 Unauthorized access to retained prompts 2 5 10 Medium – Storage breach R2 LLM hallucination exposing inferred PII 4 4 16 High – Inference errors R3 Bias in responses discriminating against groups 3 4 12 Medium – Model bias R4 Retention beyond necessity leading to re-identification 3 5 15 High – Long-term storage R5 Lack of oversight on sensitive queries 2 5 10 Medium – Human failure Overall Risk Level: [High/Medium/Low] – Requires mitigations.
6. Mitigations
Technical and organizational measures (Art. 32 GDPR). Prioritize for high risks.
Risk ID Mitigation Type (Technical/Org) Residual Score Implementation Status Monitoring R1 Encryption at rest/transit (AES-256); RBAC + MFA Technical 4 Implemented Quarterly pentests R2 Prompt guards, response validation layers; RAG for fact-checking Technical 6 In progress A/B testing accuracy R3 Bias audits pre-deployment; Diverse training data Both 6 Implemented Annual audits R4 TTL-based auto-deletion; Pseudonymization Technical 5 Implemented Retention dashboards R5 Automated triage + 24/7 on-call overseers Org 4 Implemented SLA metrics (99% <1h) Additional Safeguards:
- DPIA Re-assessment Trigger: Model retrain, >20% traffic increase.
- Incident Response: Breach notification <72h (Art. 33).
- Vendor Audits: Annual SOC2/ISO27001 reviews.
Residual Risk Level: [Low/Medium] – Acceptable? [Yes/No]
7. Consultation and Sign-Off
- Data Subjects: [e.g., Privacy notice linked in UI]
- DPO Consultation: [Date; Comments: ________________]
- Supervisory Authority: [Consulted? Yes/No – e.g., Pre-submission if high residual risk]
Approval:
Role Name Signature Date DPO Controller Next Review: [Date]
This template complies with GDPR Art. 35 and aligns with EDPB Guidelines 4/2019. Customize placeholders for your LLM deployment. For complex setups (e.g., federated learning), expand sections as needed. Contact [DPO email] for support.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.