Book Review: AI Snake Oil

This week's review is on AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference by Arvind Narayanan and Sayash Kapoor.

Arvind Narayanan & Sayash Kapoor approach the topic of AI with a background in computer science and analysis of what AI can actually do. This enables them to clarify rather than obstruct some of the fundamental issues with Artificial Intelligence. This in stark contrast to other approaches to the topic of the usage of AI and its implications, like Kate Crawford's Atlas of AI - a polemic lacking grounding - or Khan's Brave New Words - a sales pitch. AI Snake Oil remains grounded in its assessment of AI.

The book begins by defining itself and its field - splitting AI into two main categories: predictive - the use of AI to make future calls based on past data - and generative - the use of AI to create new words/text/image based on training data. This split enables a better analysis of the field. Later chapters also specifically look at content moderation - a third category of AI according to the authors and one where AI likewise fails to deliver (Kapoor worked for Facebook in content moderation).

Narayanan and Kapoor lay out their text fairly logically - moving through predictive AI and its issues, a research specialty of Narayanan, and highlight several specific issues. Their helpful highlighting of AI snake oil as "AI that not and cannot work as advertised" highlights the gold rush nature of the field. They further evaluate some prominent claims on an access of harm - the effect of the product on people and society - vs truth - whether the product can actually deliver.

An AI-generated meme supposed to show "where someone slaps AI on a product and charges more for it."

Chapter 4 marks a shift to generative AI. The explanation of the development of Generative AI in this chapter is one of the best written explorations and introductions to the technical history of AI and might make the purchase of the book worth it single-handedly. The chapter also discusses some of the ethical issues of GenAI - appropriation of creative labor, errors with AI classification, image generation, and automating "bullshit." This last point leads to the perfect definition of "speech that is intended to persuade without regard for the truth" (139).

Ch 5-8 are a little less linear, discussing the existential threat of AI, social media moderation, the hype cycle of AI, and possible futures/paths forward. I found Chapter 5 and 8 as twins - relatively weak - lacking in deeper analysis and seeming to be a bit surface-level with regards to both aspects. Chapter 6 certainly offers a powerful investigation of the use of AI to monitor social media posts and the issues inherent in this. Chapter 7 is an indictment of the media and capitalism's hyping of AI.

All-in-all the work answers most of the questions on the technical problems of AI and discusses the base ethical issues of AI with a technical and academic approach. A longer work might engage with the deeper philosophical and moral issues of AI, but the discussion here is a start. I'm not quite sure if the "how to tell the difference"past of the subtitle delivers, however, Narayanan and Kapoor deliver something near that target.

AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference by Arvind Narayanan and Sayash Kapoor.

Rating: 4/5 Stars

Good For: Those wanting a deeper investigation into what AI is and where the jagged frontiers of predictive AI in particular lie.

Best nugget: Looking at AI through the lenses of predictive and generative clarifies many of the issues of use and ethics around this field.

Please note: As an Amazon Associate I earn from qualifying purchases. However, I am not paid to provide reviews or use content.