Comment

With both snark and scholarship, wisecracks and wisdom, University of Washington linguistics professor Emily M. Bender and sociologist Alex Hanna of The Distributed AI Research Institute (DAIR) provide a thorough overview of what "AI" really is and what's really behind the onslaught of AI marketing. Bender and Hanna provide a wide swath of examples of how AI is being incorporated into critical sectors like healthcare, law, education, journalism, science, policing, and warfare, and why and how large language models and other machine learning systems are often supremely inappropriate technologies in those contexts. Of special note is their discussion of the false dichotomy between AI "Boosters" and AI "Doomers," and how both ends of that closed loop rest on a manufactured narrative of the inevitability of AI, which feeds the interests of the large tech companies who are seeking to profit from the technology. The Boosters and the Doomers are, they write, "two sides of the same coin: the substance of the coin is the belief that the development of AI is inevitable and that that resulting technology will be both autonomous and powerful, and ultimately beneficial, if we play our cards right." This "debate," they conclude, arises from alarming ideologies and distracts from the real harms that the technologies are threatening to cause to the workforce, to social services, and to the climate. The book concludes with strategies for "popping the hype bubble" aimed at both consumers and regulators. The best strategy, they write, is to simply ask questions of the claims being made for AI and the results being achieved. A powerful question to ask is: "Can you connect the inputs to the outputs?" In some cases, such as when AI is transcribing audio, the answer is "yes" because there is a provable and demonstrable link between what goes in and what comes out, and the accuracy of the work performed by the technology can be objectively tested. However, when "AI" is used to analyze faces and its makers claim that it can detect if the face belongs to a criminal, there is clearly no valid relationship between the data and the conclusions. They also promote transparency, disclosure, accountability, expanded data rights, and labor protections. "The AI Con" packs a ton of examples, analysis, commentary, and research into a highly accessible, somewhat breezy, and often very funny 200 pages (not counting nearly 60 pages of citations). It is both an excellent guide to the critical analysis of AI for individual users as well as a reference for those developing workplace policies or government regulations.