The Alignment Problem by Amogh Joshi

The Alignment Problem

The Alignment Problem is a novel dedicated to exploring the ethical questions posed by

artificial intelligence, primarily the inherent bias which arises through traditional machine

learning algorithms. In a world dominated by machine learning algorithms, from complex

threat detection softwares or generative animation algorithms, to the curated feed right

here on Quora, this novel provides a detailed insight into the potentially disastrous problems

posed by machine learning algorithms––algorithms so effective it took years just to

determine these biases.

The novel takes a historical approach, tracing the development of various machine learning

algorithms over time, and exposing the underlying biases revealed within them. By exploring

such biases, including a famous example, the gender prejudice in Google’s word2vec

software, it paints a vivid picture of the various problems within existing machine learning

systems. Or, possibly the most commonly known example of machine learning bias, the

inability of facial recognition softwares to recognize people of color with similar precision.

In fact, while it’s likely that you haven’t heard of a portion of the algorithms mentioned in the

novel, they are vastly integrated into common applications that you’ve definitely used, and

that makes it all the more interesting and yet scary as the underlying biases are revealed.

One thing that made this novel, specifically, stand out, was its capability to approach the

problem from all sides. Ethical AI is a field that happens to be ethically difficult to approach

(no pun intended), and in turn it’s easy to simply expose common biases and squarely place

the blame on the field of artificial intelligence. In contrast, this novel not only provides

context as to the decisions of researchers, but covers the actual causes of the biases in

algorithms, such as patterns in the training data used. Furthermore, it delves into potential

solutions for the issues it points out, and the current proposed answers to the so-called

“alignment problem” by leading researchers.

Overall, the novel is extremely well-written and could even pass off as a general history of

artificial intelligence algorithms, as it covers all of the major machine-learning-related

breakthroughs of the last near-century and simultaneously provides insight as to the

underlying problems in current architectures, why these problems exist, and what is being

done to solve them. A must read for those interested in further exploration of the field of

machine learning.