AI Transparency
EU AI Act Article 52 — How Pauhu discloses AI involvement
1. Overview
This page explains how Pauhu uses artificial intelligence and complies with Regulation (EU) 2024/1689 (the EU AI Act), specifically Article 52 (transparency obligations for certain AI systems).
Pauhu is a search and analysis platform for EU institutional data. We believe users have the right to know when AI is involved in producing the results they see, what kind of AI is used, and what its limitations are.
2. AI System Classification
Under the EU AI Act risk-based classification framework, Pauhu's AI components fall into the following categories:
| Component | AI Act Classification | Rationale |
|---|---|---|
| Search (retrieval) | Minimal risk | Information retrieval from public EU data sources — no decision-making, no profiling |
| Chat (generation) | Limited risk — Art. 52 | AI-generated text summaries require transparency disclosure per Art. 52(1) |
| Translation | Minimal risk | Neural machine translation of existing content — no content generation |
| Annotation | Minimal risk | Automated topic classification using EuroVoc taxonomy — deterministic mapping |
| Terminology lookup | Minimal risk | Exact match against IATE database — no generative component |
Pauhu does not operate any high-risk AI systems as defined in Annex III of the AI Act. Pauhu does not use biometric identification, emotion recognition, social scoring, or any prohibited practices listed in Article 5.
3. Where AI Is Used
3.1 Semantic Search (Vector Embeddings)
When you search, Pauhu converts your query into a mathematical representation (a 1024-dimensional vector using the BGE-M3 model) and finds documents with similar meaning. This is AI-assisted retrieval — the documents returned are real EU institutional documents, not AI-generated content.
Disclosure: Search results are ranked by semantic similarity. The ranking algorithm uses AI embeddings but the source documents are unmodified official publications.
3.2 Chat (Grounded Generation)
The chat feature uses a language model to generate natural-language answers grounded in retrieved documents. Every chat response includes citations to the source documents used.
Disclosure per Art. 52(1): Chat responses are clearly labelled as AI-generated. The interface displays:
- An "AI-generated" badge on every chat response
- Source citations with CELEX numbers or document identifiers
- A confidence indicator based on retrieval similarity and generation entropy
- A notice that AI responses should be verified against original sources
3.3 Automatic Topic Classification
Documents are automatically classified using the EuroVoc multilingual thesaurus maintained by the Publications Office of the EU. Classification uses a combination of keyword matching and vector similarity.
Disclosure: Topic labels shown as filter chips indicate automated classification. Users can verify by clicking through to the original EU source.
3.4 Translation
Pauhu offers neural machine translation for search queries and interface text across 24 EU official languages. Translation models run locally in the user's browser (ONNX Runtime) — no query data is sent to external servers.
Disclosure: Machine-translated content is labelled as such. Official EU translations (available for most legislative texts) are used when available and marked as "Official translation".
4. Where AI Is Not Used
The following Pauhu features do not involve AI:
- Source documents: All documents are sourced directly from EU institutional APIs (EUR-Lex, TED, IATE, etc.) and served unmodified
- IATE terminology: Exact database lookup — no generation or modification
- Document metadata: CELEX numbers, dates, authors, and procedural information come directly from source institutions
- Pricing and billing: Deterministic tier-based pricing with no algorithmic personalisation
- User authentication: Standard OAuth 2.0 — no behavioural analysis
5. User Notification
Pauhu notifies users of AI involvement through the following mechanisms:
| Feature | Notification Method |
|---|---|
| Chat responses | "AI-generated" badge + source citations + confidence score |
| Search results | "Ranked by AI" indicator in search bar |
| Topic labels | Tooltip: "Automatically classified" |
| Translations | "Machine translated" label vs. "Official translation" label |
| This page | Permanent transparency disclosure accessible from every page footer |
6. Human Oversight
Pauhu implements human oversight in accordance with Article 14 principles:
- No autonomous decisions: Pauhu does not make decisions on behalf of users. All AI outputs are informational — users decide how to act on search results and chat answers.
- Source verification: Every result links to the original EU institutional source. Users can always verify AI-assisted outputs against authoritative documents.
- Kill switch: System operators can disable AI features (chat, semantic ranking) instantly without affecting access to raw data. See incident response procedures.
- Feedback: Users can report incorrect AI outputs via the feedback mechanism in the chat interface.
7. Training Data and Sources
Pauhu's AI components use the following data:
| AI Component | Data Source | License |
|---|---|---|
| Search embeddings (BGE-M3) | Pre-trained on public multilingual corpora (BAAI) | MIT License |
| Topic classification | EuroVoc thesaurus (Publications Office) | CC BY 4.0 |
| Chat grounding | Retrieved EU documents (not fine-tuned on user data) | Per-source (see data catalog) |
| Translation | Helsinki-NLP models (University of Helsinki) | CC BY 4.0 / MIT |
No user data is used for training. Pauhu does not collect, store, or use user queries or interactions to train or fine-tune AI models. All inference runs in the user's browser (ONNX Runtime Web) — queries never leave the device.
8. User Rights
Under the EU AI Act and GDPR, users have the right to:
- Be informed that they are interacting with an AI system (Art. 52)
- Receive an explanation of how AI outputs were generated
- Access their data — request a copy of any personal data processed (GDPR Art. 15)
- Request erasure — delete their account and all associated data (GDPR Art. 17)
- Object to processing — opt out of AI-assisted features and use raw data access only
- Lodge a complaint with the Finnish Data Protection Ombudsman (tietosuoja.fi) or their national supervisory authority
9. Contact
For questions about AI transparency or to exercise your rights:
- Data Protection Officer: dpo@pauhu.ai
- General enquiries: info@pauhu.eu
- Registered address: Pauhu Ltd, Helsinki, Finland (EU)
This transparency disclosure was last updated on 2026-03-13 and is reviewed whenever Pauhu's AI components are materially changed.