AI is intertwined with every aspect of society, and the impact of AI will increase over time. A lot of people focus on increasing the technical capabilities of AI but forget all the social aspects affected by AI. Due to AI becoming increasingly present in the general day-to-day, it is important to take into account the social implications and ethical issues surrounding AI. This can prevent issues in the future and lead to better integration of AI into society.
Take healthcare as an example: well-trained AI can diagnose a patient with better accuracy than humans can. However, people do not trust an AI to decide in a critical setting such as a hospital, so currently, very few diagnoses are made by AI. Patients would rather have a human doctor handle their diagnosis, there are two big reasons for this lack of trust. Firstly, patients do not understand the AI, additionally, patients like to know who is liable if a mistake is made (Gerke et al., 2020). Liability is clear when the human makes the decision, whereas with AI liability is often unclear. This creates an ethical dilemma of which few answers are present at this point.
Another issue lies within the data. Since AI is built on a training set with lots of data, the AI uses this data to make decisions. For starters, humans are generally biased in some way, and if an AI is trained on a biased dataset then the AI will become biased (Gerke et al., 2020). In 2016, Microsoft started an AI-controlled Twitter account named TayTweets which learned to behave by looking at Twitter. In less than 24 hours, this bot went from friendly to becoming racist and toxic (Mathur et al., 2016). This shows that learning from humans can result in a biased AI. Secondly, AI learns from the data that is gathered, and to make a decision the AI needs a lot of data. This data could be very personal and private in certain sectors, such as healthcare. This results in a trade-off between privacy and effectivity: sharing more data gives rise to less privacy but a better AI system; but sharing fewer data assures more privacy but a worse AI system (Gerke et al., 2020).