IGF 2023 Day 0 Event #21 Under the Hood: Approaches to Algorithmic Transparency

Civil society and policymakers are increasingly calling for greater algorithmic transparency and AI explainability for a range of reasons: concerns about filter bubbles and echo chambers, seemingly “black box” demotions, and bias. The OECD AI Principles stress the importance of “transparency and responsible disclosure around AI systems to ensure that people understand when they are engaging with them and can challenge outcomes.” Greater understanding of the details and nuances of this topic will help foster more developed policy conversations and approaches. This interactive session will explore the complexities of algorithmic transparency and the following key policy questions: 1. What does algorithmic transparency and AI explainability mean? Why is this important? 2. What are tech companies doing to meet calls for algorithmic transparency and AI explainability? 3. What are key considerations for policymakers when considering legislation about algorithmic transparency and AI explainability? 4. How do users benefit from greater algorithmic transparency and AI explainability? This session is designed to elicit multistakeholder participant feedback about practical proposals for algorithmic transparency. By reviewing already-established efforts to provide more transparency around algorithms – from transparency reporting to in-product features – we hope to discuss with participants what efforts are already underway, while using their feedback to determine if more is needed, and why. The EU Digital Service Act requires audits, calling for auditors who have “the necessary expertise in the area of risk management and technical competence to audit algorithms.” The UK Online Safety Bill requires the release of “information about systems and processes which a provider uses to deal with illegal content.” Other bills and laws propose similar requirements, but often these are high level and non specific. Zoe Darme (Google) will give an “under the hood” peak at the complex systems that power tools like Google Search. Participants will workshop a new experience called “Life of a Query,” which is a streamlined version of training for all new Search engineers. We will then engage in a frank and open discussion about technical complexities and mitigations for abuse. Whether end users, policymakers, academic experts or software engineers, we share the same goal – increased trust and accountability in the use of algorithms in decision-making processes. This session will help us develop a better understanding of the needs, limitations, challenges and opportunities.

Using Zoom will allow both onsite and online participants to see and hear each other. We will ask all participants, both in person and remote to be logged in so we can manage the question queue in a neutral manner, but when in doubt will defer to remote participants as sometimes they are more difficult to spot. Our onsite and online moderators will be in constant communication to ensure that we can facilitate questions and comments from both onsite and online participants. We will also consider the unique challenges and opportunities that remote participants face, such as time zone differences, technical limitations, and differences in communication styles. We will urge our speakers to use clear and concise language, avoid technical jargon, and provide context for all information discussed during the session to ensure that both onsite and online participants can follow along and understand the content. Finally we will explore the use of a polling tool, such as Mentimeter or Poll Everywhere, to ask questions and get feedback from both onsite and online participants in real-time.