Ethics & AI

Arumoy Shome

2021-10-08 Fri 19:17

I attended the ICAI session on Ethics & AI. It was a really informative session which touched upon some very interesting and intriguing topics. In this article, I share some of my thoughts on the matter.

  1. We know that if we use an AI system and then utilize its output as input to retrain the model, we may end up with a “self-fulfilling prophecy” situation. But what can we do to prevent this? I think this comes back to the data where we must put checks in place to prevent retraining with data that was directly generated by the AI system. But how do we do this?
  2. AI has been used to solve problems in healthcare, produce self driving cars and predict risk to society for convicts. But these systems suffer from serious problems of bias. This begs the question: “Are there certain problems where we should not be using AI”? If we put an engineering lens on how AI systems work, we can consider it to be a smart automation system. And then we can perhaps begin to see that not all tasks are worth automating.
  3. Self driving cars may work in the roads of certain parts of North America. There is bright light, roads are wide, and traffic consists mostly of cars and trucks. But it will (and we have evidence that) fail to operate in say the streets of Amsterdam. This is a known problem of AI (think Andrew’s example of fault detection on factory floors). This means that we not only have to consider the car but also the environment in which it operates. This lead me to think: What if we monitor the internal rather than its external environment of a car?” The internals of a car are much more uniform which may be easier to monitor and thus produce less “noise” and edge cases for the autonomous driving system to learn from.
  4. On the notion of “What is ethics”? I think that culture plays an important role. Take the example of the Indian child what was taken away from his parents because a neighbour saw him sleeping in his parent’s bedroom. This was in a “western” country where children sleep in their own room (or in a crib) from a very early age. However in India, children sleep with their parents until they are quite old.
  5. All the ethics questions raised against AI systems makes me wonder how engineering can help solve this problem. In engineering, “divide and conquer” a highly adopted principle. I wonder if we are trying to do too much with a single AI system. Perhaps we can break the problem of monitoring a patient’s health into several sub-problems and solve them individually with smaller AI systems. This also helps with the explainability of the AI systems since they are very focused on a single task. The results from these sub-systems then can be aggregated to answer a larger problem.