Back to agendaABSTRACT
Speakers: ICREA Research Professors Marco Baroni and Carlos Castillo, both from the Universitat Pompeu Fabra (UPF)
When: 27th of June 2023, 18:00h
Where: Zoom & Auditorium FCRI, Passeig de Lluís Companys, 23, 08010 Barcelona
Abstract:
As AI systems start being deployed in a large number of everyday contexts, it is particularly important to study the transparency and trustfulness of these systems from different perspectives.
In the first part of our joint talk, we will consider whether AI systems can have disadvantageous consequences for people based on the social group to which they belong, i.e., whether AI can engage in indirect discrimination, specifically statistical group discrimination. We will consider various examples of this phenomenon, which seems quite pervasive and particularly affects high-risk applications of AI, such as those related to housing, employment, and the administration of justice.
In the second part, we will consider to what extent large AI systems trained to “speak English” by being exposed to large amounts of text are really understanding natural language like humans do, and to what extent they have instead developed an alternative communication code that is only superficially resembling English. We will discuss some surprising results that suggest that the second hypothesis is correct, and the implications that this might have in terms of the usability and safety of these systems.
The ICREA colloquia are a great way to learn about remote fields of research from our best experts. We usually have two speakers, who offer their opinions on the same subject from different angles. They are open to all ICREAs and their guests.
