Dr. Ciarán Seoighe:

That is a great question. There is no obvious immediate answer, but researchers are doing work on occasion to unpick the biases in the training material that is used. There are some great examples of Irish researchers who discovered - I am not sure if members are aware of this - that biases were built into the large language models being used in the MIT training data. There were images and then words associated with the images that had inherent biases built into those. Without knowing those and without checking those, the biases are then taught to the machine through the machine learning. Therefore, we do need to be doing that level of research and engagement to check at all times where those biases are.

The other thing is transparency, that is, trying to be as transparent as we can in the outcomes and keeping humans in the loop. Another factor on which there is a lot of research under way at the moment is on how we often talk about the black box nature of AI, which some members mentioned earlier, and the sense that it does this thing and we do not quite know how it has arrived at those conclusions. Some research that needs to be done next is into being able to interpret that black box activity so that we can understand exactly what came out of it and then interrogate that because one of the scariest parts is when something goes into that black box and we cannot fully understand how it arrived at a conclusion. Therefore, we need the research and data into it such that we can then translate back and track back how that particular decision was arrived at.