In AI We Need To Trust; Not There Yet

Friday 10.Nov

THEME: AI

Back in 2018 in his book “21 lessons for the 21st Century”, historian Yuval Harari emphatically noted that “Already today, ‘truth’ is defined by the top results of the Google search.”. The recent developments and disruptions caused by foundational AI models (to put it simplistically) such as ChatGPT denote the gravity and criticality of integrating the AI technology in every aspect of our lives. However, how feasible is it to trust AI systems? In 2022, journalist Karen Hao’s article about Facebook’s AI systems, reveals an interesting truth; tech organizations may not be in control of their AI systems. What seems to be even more worrisome, is that tech organizations may not want to be in control of their AI if this impedes their growth, or profits (or both).

AI systems tend to be highly dependent on data, so their error-proneness and potential impact rely on how they treat those data. If we consider that data are socially constructed or simply put, data reflect how we, humans, live, then this means essentially two things:

• AI systems need to be designed in a way that ensures they can be trusted,
• The organizations that own and operate AI systems should be mature enough for governing them responsibly.

All these beg several questions such as: What are the best practices when evaluating how an organization governs an AI system? What are the upcoming regulations dictating how we can audit AI systems? What are the essential aspects of an AI system that we as engineers need to test? How is it possible for an organization to find a balance between building Trustworthy AI versus optimizing for efficiency and for maximizing their business profits?

In his speech, Yiannis Kanellopoulos will elaborate on the above mentioned questions in a practical way and provide real-life examples on how an AI system can be tested in terms of its Performance, Fairness, Accountability, Transparency and Safety/Security.