Alignment problem

The alignment problem represents, to a certain extent, a paradox: since the discipline’s inception in the mid-20th century, mathematicians have sought to build artificial intelligence “in the image of humans.” However, this challenge has proven to be extremely complex, as was revealed in the field of language by the early days of machine translation in the 1980s.

Even today, the alignment problem persists, particularly for companies seeking to develop so-called “general” or “strong” artificial intelligence, which is involved in complex tasks such as autonomous vehicle driving, or has ethical implications, such as in natural language generation (issues of disinformation or fake news, in particular). Numerous approaches exist—evolutionary monitoring, adversarial models—that seek to limit the divergence that arises between the specification designed by the development team and the emergent behavior of the developed model. One of the most widely used approaches today is reinforcement learning, which involves—where feasible—having a human evaluate the results of the AI model so that it can be corrected through a sort of “reward system” over a series of trials and errors.

More generally, the alignment problem relates to the issue of controlling, understanding, and predicting the behavior of artificial intelligence, involving both engineering solutions and ethical considerations.