Navigating Your Organisation’s AI Ethics Journey

By far the greatest barrier that stagnates wider discourse and engagement in AI ethics is the sheer complexity of AI itself which makes it difficult to understand its ethical implications. This dilemma can be illustrated by using AI Ethics Activist Deborah Raj’s analogy of engineering responsibility in the automobile industry. If we think about accountability in this industry, there is a clear allocation of responsibility and defined processes for fixing problems. For example, when a car is found to have faulty brakes there is no question as to who is responsible. The manufacturer has a duty of care to ensure the safety of their vehicles and in the interim between identifying and fixing the problem, cars with faulty brakes are recalled. At the same time we can easily distinguish discussions about responsibility at the micro level, such as ensuring the overall safety of cars before they are allowed on the road, and responsibility at the macro level, such as the need to completely rethink how we make cars to make them more environmentally friendly. Working on both these areas simultaneously is essential towards ensuring that we are optimising the short and long term benefits of cars while limiting the negative impacts they have on our future environmental security. 

Having similar simultaneous ethical discussions with AI is, by comparison, very difficult – for a number of reasons. For one, we can get stuck muddling through the micro level responsibility issues because sometimes we can’t even measure the “brakes” in some systems. For example it may be difficult to identify subpopulations in a dataset which you need to do to ensure the model is working appropriately across all groups. Additionally, the ambiguous accountability caused by the lack of transparency when black box models are used in AI assisted decision making makes it difficult to assign responsibility.

If we take facial recognition technology as an example, there can be severe consequences when the system does not perform as intended or when black box models calculate outputs we take as infallible truth. In fact, it’s already happening. There are already people being implicated of crimes they did not commit, wrongful arrests against people of colour, and the possibility of wrongful death isn’t out of the realm of possibility with increasing integration of facial recognition technology in the military. It’s important to call out here that there are ethical consequences when facial recognition systems don’t work as intended but there are questionable ethics even when they do. The amount of biometric data required to train these powerful models is astronomical and even with the implementation of privacy preserving technology, we may never be able to mitigate the risks to privacy to an ideal or safe level. Asking people to risk their privacy requires a high level of justification and transparency which might vary depending on the context and intended use but should be especially high in societies that prioritise democratic values such as Australia and New Zealand.