AI Transparency & EU AI Act Compliance

Last Updated: December 27, 2025

EU AI Act Regulation: 2024/1689

1. AI System Classification

Risk Level: Minimal Risk

This AI system is classified as minimal risk under Article 5 and Annex III of the EU AI Act. It is an educational chatbot for Historical European Martial Arts (HEMA) and does not fall into prohibited or high-risk categories.

2. Purpose and Intended Use

Primary Purpose: Provide information and answer questions about Historical European Martial Arts, including historical fencing treatises, techniques, terminology, and historical context.

Intended Users: HEMA practitioners, researchers, historians, martial artists, and anyone interested in learning about historical European swordsmanship.

NOT Intended For:

Professional medical advice or injury diagnosis
Legal advice or formal historical authentication
Employment decisions or educational assessments
Critical decision-making without human verification

3. AI Provider Information

3.1 WebLLM (Local Browser-Based Models)

Classification: General-Purpose AI Model

Data Processing: 100% Local (in your browser via WebGPU)

Models Available:

Llama-3.2-1B-Instruct (1 billion parameters)
Llama-3.2-3B-Instruct (3 billion parameters)
Hermes-2-Pro-Llama-3-8B (8 billion parameters)
Hermes-2-Pro-Mistral-7B (7 billion parameters)

Privacy: No data sent to external servers. Model files downloaded once and cached locally.

Training Data Cutoff: Varies by model (2023-2024)

Capabilities: General text generation, historical knowledge, function calling (select models)

Limitations:

Requires WebGPU-compatible browser (Chrome 113+, Edge 113+, Firefox 121+)
Large initial download (1-8 GB per model)
Performance depends on your device's GPU
May produce less accurate responses than larger cloud models

Provider Documentation: webllm.mlc.ai

3.2 OpenAI (App-Hosted Cloud Provider)

Classification: General-Purpose AI Model (GPAI)

Data Processing: Cloud-based (United States)

Models Available:

GPT-4o-mini (default) - Optimized for speed and cost
GPT-4o - Advanced reasoning and multimodal capabilities
GPT-4-turbo - Previous generation flagship model

Training Data Cutoff: October 2023 (as of December 2025)

Rate Limit: 40 requests per 24 hours per user (app-hosted free tier)

Capabilities:

Advanced natural language understanding
Function calling for knowledge retrieval
Large context windows (128k tokens for GPT-4 models)
Multilingual support

Limitations:

Knowledge cutoff: Information only up to October 2023
May generate plausible-sounding but incorrect information ("hallucinations")
Cannot access real-time information or external sources beyond provided context
Requires internet connection
Data transmitted to OpenAI servers in USA

Data Retention by OpenAI: API data retained for 30 days for abuse monitoring, then deleted

EU AI Act Compliance: OpenAI is registered as GPAI provider and subject to Articles 53-56

Provider Documentation:

4. How the AI System Works

4.1 Basic Operation

You type a question about HEMA
The system may call specialized functions to retrieve relevant historical context
Your question + retrieved context is sent to the selected AI model
The AI generates a response based on its training and provided context
The response is streamed back to you in real-time

4.2 Function Calling (Context Retrieval)

To improve accuracy and reduce token usage, the system uses function calling to retrieve specific information on-demand:

get_weapon_techniques: Retrieves guards, strikes, and cuts for specific weapons
get_historical_context: Provides biographical and historical information
search_hema_knowledge: Searches the knowledge base for relevant content
get_fencing_concepts: Explains timing, distance, and tactical principles

Benefit: This approach reduces token usage by 50-70% and improves response accuracy by providing relevant historical context only when needed.

4.3 Knowledge Sources

The AI system's knowledge about HEMA is derived from:

Joachim Meyer's Treatises (1570s): Fechtbuch and Arte de Combate
Wiktenauer: Online repository of historical fencing manuscripts
Public domain historical sources
AI model training data (up to October 2023 for OpenAI models)

5. Accuracy and Limitations

⚠️ Important Accuracy Disclaimer

AI systems can make mistakes. This includes:

Hallucinations: Generating plausible-sounding but factually incorrect information
Misinterpretations: Incorrectly interpreting historical texts or martial arts terminology
Outdated Information: Training data cutoffs mean recent research may not be reflected
Contextual Errors: Mixing concepts from different historical periods or fencing traditions
Translation Issues: Historical treatises are often in German, Italian, or Latin - translations may vary

Always verify information with:

Original historical sources (Wiktenauer, primary manuscripts)
Qualified HEMA instructors and recognized experts
Peer-reviewed academic research
Multiple independent sources

6. Safety Considerations

🛡️ Critical Safety Warning

DO NOT practice martial arts techniques without proper supervision.

Always practice under qualified HEMA instructors
Use appropriate safety equipment (masks, gloves, protective gear)
Ensure adequate training space and controlled environment
Start with basic techniques and progress gradually
Understand that historical martial arts carry inherent injury risks

We are not liable for injuries resulting from practicing techniques discussed in AI-generated responses.

7. Human Oversight

This system operates as a human-in-the-loop AI system:

User Control: You choose when to use the system and can stop generation at any time
Critical Thinking Required: Users must evaluate all responses for accuracy and applicability
Expert Consultation: Users are expected to consult qualified instructors for practical application
No Autonomous Decisions: The system provides information only; all decisions are made by users

8. Transparency Obligations (EU AI Act Article 50)

In compliance with Article 50 of the EU AI Act, we inform you that:

You are interacting with an AI system. This is disclosed via banner notification on first use and throughout the interface.
AI-generated content is clearly marked. All responses are identified as AI-generated in the chat interface.
System capabilities and limitations are explained. This page provides comprehensive information about AI models, accuracy, and limitations.
Data processing is transparent. See our Privacy Policy for details on data collection, processing, and your rights.

9. Your Rights and Responsibilities

9.1 Your Rights

Right to Information: Access detailed information about AI systems used (this page)
Right to Choose: Select between local (WebLLM) or cloud (OpenAI) AI providers
Right to Privacy: See our Privacy Policy for GDPR rights
Right to Complain: Lodge complaints with your national data protection authority

9.2 Your Responsibilities

Verify Information: Always fact-check AI responses with authoritative sources
Practice Safely: Only practice techniques under qualified supervision
Use Appropriately: Do not rely on AI for critical decisions without human expert review
Consult Experts: Discuss any AI-generated information with qualified HEMA instructors before practical application

10. Accountability and Governance

Provider Responsibility:

OpenAI is the GPAI model provider responsible for their models' compliance
WebLLM/MLC provide local model implementations
Fecht.guru is the deployer responsible for system integration and user interface

Governance Structure:

Continuous monitoring and improvement of AI systems
Regular updates to documentation and compliance materials
User feedback collection and issue resolution

11. Contact and Complaints

For questions or concerns about our AI system:

Technical Issues: Consult with your HEMA instructor or community for guidance on interpreting AI responses
Privacy Questions: See Privacy Policy for data processing information and rights
Data Protection Authority: You may lodge complaints with your national DPA (see EDPB directory)

12. Updates and Changes

This AI Transparency document is updated when:

New AI providers or models are added
EU AI Act guidance changes
System capabilities are significantly modified
Accuracy or safety considerations require clarification

Check the "Last Updated" date at the top of this page. Material changes will be announced to users via the application interface.