Policy Prototyping the AI Act: Transparency requirements – workshop report

The Knowledge Centre organised a workshop on 21 June 2023 to officially kick off its 'Policy Prototyping the AI Act: Transparency Requirements' project. A group of 18 interdisciplinary professionals worked on different prototypes for instructions for use (Art 13 AIA) and disclaimers (Art 52 AIA) based on five different use cases. The final result will be presented this autumn.

The EU emphasizes transparency as a fundamental value for the development, deployment, and utilization of AI systems. The topic was already given prominence in the policy documents that preceded the AI act, such as the Ethics Guidelines for Trustworthy AI, issued by the High-Level Expert Group on AI, or the White Paper on AI, issued by the European Commission. The AI Act continues that line. For this reason, the Knowledge Centre decided this year to focus its Policy Prototyping project on the transparency obligations in the EU AI Act proposal.

Those obligations are mainly translated into two articles in the regulation, namely Art. 13 'Transparency and provision of information to users' and Art. 52 'Transparency obligations for certain AI systems'. Through this project, the Knowledge Centre aims to examine the transparency obligations in detail and:

• Develop prototypes in the format of instructions for use for high-risk AI systems (under Art. 13 AIA) and disclaimers (under Art. 52 AIA).
• Collect feedback on the practical feasibility, practicability, and understandability of the transparency requirements.
• Provide feedback to policymakers and lessons learned to stakeholders.

To gather feedback on transparency requirements, the Knowledge Centre Data & Society organised a physical legal design workshop at the AI Experience Centre in Brussels. The 18 participants (professionals working with AI on a day-to-day basis and experts with experience in facilitating transparency and drafting policies) and 4 facilitators worked together for a whole day in five different groups to shape first versions of different prototypes. Through a legal design step process, the initial effort was to clarify transparency requirements in the context of a use case of a high-risk application (under the AI Act). This included identifying potential stakeholders of the use case and identifying pain points and opportunities. The next phase involved brainstorming and subsequently laying the first foundations of prototypes, considering a potential end-user or recipient..

A total of 21 people participated in the workshop: 4 facilitators from the Knowledge Centre Data & Society, 8 representatives of AI-developers (incl. start-ups) and -users, 2 consultants, 4 legal experts, 2 representatives of civil society organisations and 1 academic researcher.

Use case 1 – Medical AI device for disease detection

The use case focuses on an AI tool that analyzes eye retina photos in order to assist in identifying/diagnosing certain, specific diseases, such as diabetes.

Use case 2 – Medical device

The use case focuses on an algorithm that can predict cardiac arrhythmias, enabling early detection and better follow-up of high-risk patients. The prediction is based on an electrocardiogram and additional values of blood pressure and cholesterol.

Use case 3 - Chatbot

The use case focused on a chatbot which employees can use to ask HR-related questions. The bot uses a collection of documents (e.g. HR documentation) to answer these questions.

Use case 4 - Deepfake

The use case focused on the use of deep fake technology in documentaries. The use of deep fake technology protects the anonymity of people involved, while also retaining facial expressions and emotions. 
 

Use case 5 – HR Talent API

In this use case, a four-way talent matching API is used to connect job seekers with vacancies relevant to them. The AI system takes several criteria into account such as work experience, travel time, acquired skills and so on. The instructions of use prototype will be focused on users/deployers of the AI system.

During the workshop, rough drafts of the prototypes were conceived. These drafts were further developed by experts from various groups in accordance with the AI Act. In the autumn, the Knowledge Centre organised another online workshop where the participants and other interested parties gave feedback on the developed prototypes. This finally led to a report presenting the prototypes and policy recommendations. You can download the report below. 
 

About

Image: Cash Macanaya via Unsplash