Professor Gregory O'Hare:

I thank Deputy O'Reilly. This is a really profoundly difficult question to answer. By way of trying to illustrate the colossal challenge it presents to science, not only is there typically not an algorithm and not only is there typically not a set of algorithmic steps that one could scrutinise, with a trained eye, AI and, in particular, deep AI, does not have an algorithmic basis. Even if I, a professor of artificial intelligence, looked at a particular AI application that was using deep learning, I would have great difficulty in being able to establish on the surface how it is arrived at, its deduction and recommendation or conclusion. Whenever we talk about transparency, we really mean things like auditability and one of the current watchwords at the moment, which is explainable AI. It is incumbent that the legislation mandates that such systems are able to support explainable AI. In other words, there may well be explanations in the spoken language or written language. These explanations should explain that, based on a statistical high correlation between concept X and concept Y, this resulted in an inference, and in turn, when this was combined with another strong correlation between some other phenomena, this resulted in the particular recommendation. That is the kind of explainability that we are moving towards. Some of these systems are huge in their extent and their complexity is enormous.

I would counsel people with regard to earlier questions that, of course, it is crucial that we have appropriate, considered engagement of all the stakeholders involved. I would counsel them that often such engagement ought not to be rushed and necessarily takes considerable time. The velocity at which the uptake and deployment of AI systems is occurring does not afford us that level of time.