What is our legal framework?

All personal data are processed in accordance with European Union data protection law, that is to say in line with Regulation (EU) 2018/1725 (EUDPR).

Why do we process personal data?

Individuals can engage in a dialogue with the ECBot, which is found on selected pages of the ECB and Banking Supervision websites, to receive information on ECB-related topics in real time. The complete conversation between the end user and the ECBot is stored by the data processing company IBM. As in all “natural language processing” chatbots, conversational data is needed to train the ECBot. ‘Natural Language Processing’ is based on deep learning, which enables chatbots to acquire meaning from inputs given by end users. It assesses the users’ intent from the input and then creates responses based on a contextual analysis, similar to a human being.

Conversations are anonymous by default. We discourage users from entering sensitive data (e.g. information about religion or health). However, if any personal or sensitive data are voluntarily provided, they will be stored for 180 days, as part of the conversation content.

Personal data, if any, are processed in order to improve the chat function, to which end the ECB analyses all conversations in an automated manner to highlight areas for future improvement.

Only natural persons who are at least 18 years old are permitted to use the ECBot.

What is the legal basis for processing your personal data?

Your personal data are processed by the ECB because you consented to this processing by using the ECBot. You may withdraw your consent at any time by contacting the ECB Chatbot Team. All processing of your personal information will stop once you have withdrawn your consent; however, any processing that has already taken place remains lawful.

Who is responsible for processing your personal data?

The ECB is the controller of the processing of the personal data. Various teams are responsible for this processing, depending on the topic of the chatbot:

• The Directorate General Human Resources, Business Partnering Division, is responsible for the Recruitment ECBot;
• The Directorate General of Statistics, Banking Supervision Data Division, is responsible for the CASPER and Supervisory Fees ECBot;
• The Directorate General Corporate Services, Budgeting and Controlling Division, is responsible for the Supervisory Fees content in the CASPER and Supervisory Fees ECBot;
• The Directorate General Communications, Public Communication Division, is responsible for the Public Enquiries ECBot.

Who will be the recipients of your personal data?

The recipients of your personal data (including entities who have access to that personal data) are:

• The Directorate General Human Resources, Business Partnering Division;
• The Directorate General of Statistics, Banking Supervision Data Division;
• The Directorate General Corporate Services, Budgeting and Controlling Division;
• The Directorate General Communications, Public Communication Division;
• The Governance and Transformation Services Division.

The personal data will be stored in the IBM customer database.

What categories of personal data are collected?

By default, no personal data are collected when using the ECBot and we discourage you from including personal data in the conversation voluntarily.

Any personal data which the user chooses to include in the conversation will be stored. Personal data in the context of the conversation are provided voluntarily to the controller. The content of your conversations and the respective responses from the ECBot will be analysed.

Will your personal data (in a clear or encrypted form) be processed (e.g. transferred, accessed or stored) in third countries or by international organisations?

Your personal data are processed by IBM, located in Germany and Romania, and are stored on the IBM customer database. Processing is based on the inclusion of special clauses in the contract with IBM which ensure that the company complies with EU data protection standards (European Commission's website).

How long will the ECB keep personal data?

Any personal data included in the conversations will be stored for a maximum of 180 days before they are deleted.

What are your rights?

You have the right to access your personal data and correct any data that are inaccurate or incomplete. You also have (with some limitations) the right to delete your personal data and to object to or to restrict the processing of your personal data in line with the EUDPR. The ECB may restrict your rights to safeguard the interests and objectives referred to in Article 25(1) EUDPR.

Who can you contact for queries or requests?

You can exercise your rights by contacting the ECB Chatbot Team. You can also directly contact the ECB’s Data Protection Officer for all queries relating to your personal data.

Addressing the European Data Protection Supervisor

If you consider that your rights under the EUDPR have been infringed as a result of the processing of your personal data, you have the right to lodge a complaint with the European Data Protection Supervisor at any time.

The technology behind the ECBot

How it works

The ECBot uses a “decision tree approach”, in which a set of pre-defined answers is created and stored as a knowledge base. The chatbot is then trained to recognise users’ questions and provide the correct responses. It does so by means of natural language processing (NLP), a branch of artificial intelligence that focuses on the interaction between computers and humans through natural language. When a user interacts with the chatbot, the NLP algorithm analyses the question to discern the user’s intent. The chatbot then selects the most appropriate pre-defined answer from its knowledge base and provides it as a reply to the user.

This technology has recently been enhanced by the use of Retrieval Augmented Generation (RAG). While our chatbot will continue to use the decision tree approach, RAG will be gradually implemented as an additional component to improve its capabilities. Currently only the Public Enquiries ECBot uses RAG.

RAG and LLM-powered chatbots

What is RAG?

RAG is a type of technology that combines two processes to improve responses to users’ questions. First it retrieves relevant information from a large database or collection of documents. Then, the large language model (LLM) uses this information to generate a coherent and contextually appropriate response, in much the same way as a person would summarise or explain something after looking it up in a library. This approach helps to formulate more accurate and informative answers by using existing knowledge while also allowing for more natural and human-like responses.

What is an LLM?

LLMs are advanced computer programs designed to understand and generate texts in natural language. When a user asks a question or gives a prompt, these models analyse the input and create responses that are coherent and contextually relevant, much as a person would respond in a conversation. Essentially, LLMs help machines communicate with people more naturally.

How our RAG-powered chatbot works

The RAG functionality enhances the chatbot solution based on the decision tree approach. The chatbot’s knowledge base comprises a mix of documents (or web links) and pre-defined answers for topics that need a carefully worded answer or that are not explicitly covered in the documents in the knowledge base. Whenever possible, the chatbot will use the pre-defined answers to provide a response to users. But if a user asks a question that is not covered by these pre-defined answers, the chatbot will use the RAG component to search for relevant information in the documents in its knowledge base and will then instantly generate a reply for the user.

In this way, we are using a multi-layered approach for the ECBot:

Layer 1: the chatbot is able to give pre-defined answers using the decision tree approach. This type of answer provides a high level of control over the chatbot’s response, making it suitable for sensitive topics that require precise and carefully worded answers.

Layer 2: if the chatbot is not able to find an appropriate pre-defined answer using the decision tree, it will use the RAG component to search the documents in its knowledge base and then use an LLM to generate an answer.

The model we use

The LLM component of the ECBot was built with Meta Llama 3.1. It is one of the latest iterations of Meta’s open-source language models with advanced capabilities for understanding and generating language. Meta Llama 3.1 is licensed under the META LLAMA 3.1 community licence.

Watson Discovery is the component used to enable RAG in the ECBot. It uses artificial intelligence-powered searching and text analytics to extract information from documents and texts, improving the accuracy and relevance of information retrieval for more precise and contextually appropriate responses.

How will I know if the ECBot is using an LLM?

Whenever an ECBot answer is generated using an LLM, this will be stated below the response to notify the user.

Guardrails

The ECBot employs guardrails to ensure that its LLM maintains appropriate language and stays within the boundaries of its intended purpose. Guardrails are like safety barriers that guide the chatbot’s interactions, preventing it from straying into inappropriate or irrelevant topics. They help the chatbot understand what types of conversation are acceptable, ensuring that it provides helpful and relevant responses while avoiding sensitive subjects or off-topic discussions. So users can be confident that when they interact with the chatbot it will adhere to the established guidelines. The guardrails were built using Meta Llama 3.1.

Limitations and considerations

While an LLM is a very advanced type of technology, it is important to remember that the answers it provides may not always be factually correct, as it might sometimes generate information that appears accurate but is not actually contained in the knowledge base. Therefore, we encourage you to check the sources of the generated answers by clicking on the button labelled “Here’s the information I based my answer on”. This will provide links to the documents used to generate the answers, allowing you to verify that the information is correct. Please note that only the three main sources are shown and the chatbot might have used material from more sources.

Feedback and improvement

We value your feedback! If you have suggestions or encounter issues when using our chatbot, please let us know via the ECBot comment function, or by sending an email to the ECB Chatbot team. Your input helps us improve and enhance the service.