Stephen Hawking might have seen things that are decades out when he warned of the perils of AI - the larger, grander view of the world that is not accessible to most average people. Those are words of caution to be heeded. Even in the most mundane of AI applications that I run into given my work, the perils are plainly evident. The simple business of sentiment analysis for example can be fraught with trouble that were never anticipated. The accuracy of many models depends on the the data used to train it. The more comprehensive the training set the better the quality of results. In a sentiment analysis problem, the biases of the subject matter experts assigning sentiment scores to free form text in the training set forms the guts of the model.
It is not very hard to sway results in specific ways to achieve desired outcomes. If the results of the sentiment analysis for example were to be associated with financial outcomes for the those requesting such analysis, it is easy to see how subtle biases could be introduced to maximize their gains. All actions following from there would be driven by AI and unlikely to deliver benefits in a democratic manner. This is very far from machines taking over the world, but it is a way for machines to scale and expand the pernicious impacts of our character flaws. Instead of having local and contained impact, we now have the ability to cause large scale damage by way of AI.
It is not very hard to sway results in specific ways to achieve desired outcomes. If the results of the sentiment analysis for example were to be associated with financial outcomes for the those requesting such analysis, it is easy to see how subtle biases could be introduced to maximize their gains. All actions following from there would be driven by AI and unlikely to deliver benefits in a democratic manner. This is very far from machines taking over the world, but it is a way for machines to scale and expand the pernicious impacts of our character flaws. Instead of having local and contained impact, we now have the ability to cause large scale damage by way of AI.
Comments