Emily M. Bender and Alex Hanna, The AI Con: How to Fight Big Tech’s Hype and Create the Future We Want (Harper 2025). Available from HarperCollins.

In an era saturated with AI hype, a truly objective and well-researched analysis of its potential and pitfalls is not just welcome—it’s sorely needed. Unfortunately, The AI Con falls short.

This is not to say The AI Con is totally without merit. It has some important strengths. Let’s start with the positive.

Strengths

The authors identify multiple real problems of widespread AI adoption, including but not limited to:

• Dangerously high levels of data center energy consumption.
• The risk that AI may harm our future by lessening levels of in-person mentoring, thus limiting the development of human expertise.
• The risk that the availability of AI will tempt governments and NGOs to underfund social services.
• The tendency of AI developers to move the goalposts when developing benchmarks.
• The near certainty that the AI user experience will deteriorate as vendors try to squeeze more money from users by letting advertisers pay to skew results, for example, as happened with Facebook and Google.

All these problems are significant, and the authors explain some of them reasonably well.They do a great job of highlighting personal and organizational biases that lead to AI hype. My background is in law, with a focus on the practical use of legal technology, rather than business, so many of their examples were new to me.

They display a gift for major league snark. They deploy it quite well to skewer a deserving target: AI boosters who claim that AI has the potential to destroy humanity, while at the same time insisting that there be no limits on developing or regulating it.

Three Key Weaknesses: Little New Value, Relative Lack of Technical Sophistication and Imbalanced Perspective

There are too many problems with the book to explain all in a 2000 word book review. I’ll start with three of the worst:

The authors’ contribution to the existing literature is modest, especially when compared to AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference, by Arvind Narayanan and Sayash Kapoor. While sharing a similar theme, AI Snake Oil preceded ‘The AI Con’ by approximately seven months.

While AI Snake Oil is far from perfect, its analysis and recommendations are more convincing than those in The AI Con. This may be a consequence of the authors of AI Snake Oil having stronger technical backgrounds and more hands-on experience with generative AI. As a lawyer and liberal arts major myself, I’m the last person to say that a linguist and sociologist can’t have valuable insights about AI, and might even be able to give helpful advice to a company like Google on some AI-related topics (like co-author Alex Hanna has done), but for a project like this, having degrees in computer science is a significant advantage.

AI Snake Oil’s higher level of technical sophistication is a plus, but its more balanced perspective is even more important. The AI Snake Oil authors acknowledge up front that, even in its infancy (having been “born” with the release of the revolutionary version of ChatGPT less than four years ago), Generative AI is already valuable and shows potential for even more practical benefits. The AI Con has no similar acknowledgement. The relentless negativity becomes grating far before the book’s conclusion.

Consequences of the Book’s Odd Genesis (Using MST3K as a Model)

The authors explain in the Preface that the inspiration for the book came from their fondness for the cult TV show Mystery Science Theater 3000, in which the characters Tom Servo and Crow T. Robot watch terrible sci-fi movies and make them enjoyable by adding running commentary. This inspired the authors and their friends to broadcast similar comments about AI on the streaming platform Twitch. They refer to this as “Mystery AI Hype Theater 3000.”

Their MST3K-inspired style often privileges snark over substance. While satire can have its value, the book sometimes sacrifices analytical rigor for rhetorical effect.

Name-calling can be effective for political rhetoric (remember Little Marco and Crooked Hillary?), but the book’s use of derogatory terms like “mathy maths” to describe AI is unfortunate. The phrase sounds like an epithet created by a frustrated and STEM-challenged junior high school student. It detracts from the serious analytical discourse the book presumably aims for.

The use of made-up pejoratively intended descriptors such as “synthetic text extruding machine” and “stochastic parrot” rather than generally accepted terminology (large language models and neural networks) also undermines the book’s credibility. Notably, the term stochastic parrot, coined in an article co-authored by one of The AI Con authors, is more objectively and nuancedly explained in a short Wikipedia article. When a general encyclopedia provides a better explanation of a key concept than a dedicated book, it is not a good sign.

The MST3K inspiration leads to a related problem: The authors are eager to find ways to make fun of all things AI. They are less diligent about citing sources that don’t support their biases. Their handling of Dr. Ethan Mollick’s writing, as described below, is a particularly egregious example.

Selected Other Problems with The AI Con

There are way too many other problems to explain here, but here are short summaries of a few:

Conflation of what we might call “legacy” forms of AI, like predictive AI and machine learning, with the revolutionary new AI apps based on large language models and neural networks, often referred to as Generative AI, or Gen AI. The problems with legacy AI applications, such as decision-making in contexts like credit scoring and facial recognition, have been recognized for years. Has developing self-driving cars been difficult? Definitely, but that’s not particularly relevant to the new advanced AI apps.

Large language models and neural networks possess distinct sets of strengths and weaknesses. Failing to consistently distinguish legacy AI apps and modern advanced AI apps will cause unsophisticated readers to lump them all together. Big mistake.

Confusion about the distinction between process and outcome. The authors’ disdain for the way modern AI works makes it difficult for them to acknowledge that in many fields, AI can generate beneficial, otherwise hard-to-match outcomes. Richard Susskind explains the process/outcome fallacy very well in his newest book, “How to Think About AI: A Guide for the Perplexed.”

The authors don’t like the process that Gen AI uses, so they won’t admit that it can have benefits. As part of my research for this review, I watched a promotional video, an interview the authors gave to YouTube podcast host Alex Kantrowitz It provided a striking illustration of the authors’ thinking. Kantrowitz warned the authors at the beginning that he disagreed with their maximalist anti-AI views. After giving them free rein to express their views for 26 minutes, he proffered a modest personal experience that supported his opinion that Gen AI could sometimes provide practical benefits: ChatGPT had let him get more benefit from the limited time he had to explore Paris on his vacation.

The authors condescendingly advised him to forget his concern about efficiency. Rather than use AI, he should instead have looked at French newspapers and asked people in Paris to give him ad hoc tips. Kantrowitz does not speak French, and Parisians are not exactly famed for generosity in advising American tourists, but not to worry about such mundane details. The human touch is paramount for the authors, efficiency be damned. Viewer comments on their presentation were not kind. Commenter @hazasaleemi247 lamented, “I wish I used AI to summarize this video so I could get back the time I spent listening to these two guests speak.” I felt her pain.

Failure to distinguish genuine AI flaws from errors caused by unsophisticated users. AI, like any powerful tool, demands skilled operation. Drawing an analogy to the progression from riding a bicycle to driving a car, proficient AI users can achieve extraordinary outcomes, whereas those lacking proper training will frequently encounter suboptimal results. We do not attribute driving accidents to the car itself but rather to untrained drivers; similarly, blaming AI for user errors is illogical. The issue lies not with the tool, but with its improper application.

Does AI hype lead unqualified people to over-extend themselves using AI? Absolutely, and it’s entirely appropriate to make this point. However, there is also a responsibility to make sure readers understand the need for caution and education, rather than make the risible claim that AI is always worthless.

Occasional bizarre digressions. Yes, it’s true that many of the key movers behind the development of “intelligence” tests a hundred years ago—and some even more recently—had dubious views about race. It’s also true that some contemporary AI boosters (most prominently Elon Musk) espouse strikingly odd views about eugenics and reproduction. It’s a mystery why the authors devote pages to discussing these red herrings, which have little to no apparent relevance to contemporary discussions of practical AI use.

Dubious Research and Citation Practices

It’s worth taking a little time to examine an aspect of The AI Con that I found particularly frustrating: the mismatch between the authors’ self-congratulatory tone and their actual achievements. The Preface states:

[Checking] the sources for all of the hype-tastic [sic] claims often gives us a good vista on the house of cards (that is, thin research methods, shoddy argumentation, and questionable citation practice) supporting the flashy façade. In the same spirit, it is important to us to cite our sources: we care about both the provenance of the information we are sharing and about giving credit where credit is due.

Admirable goals indeed. Unfortunately, the authors fell short of their goals in at least one critical section. On page 97, they criticize an unnamed Wharton School of Business professor’s statement that when it comes to education, AI could present “the biggest equity opportunity we’ve ever had.” The authors disagree with this. In their opinion:

If we don’t get off the hype train, a privileged set of students may benefit from these tools, used in conjunction with close supervision by attentive, less-burdened human instructors. Meanwhile, most students will find themselves in classrooms led by harried, precariously employed adjunct faculty, who the academic administration expects to handle overfull classes by using the automated tools in lieu of actual interaction.

This sounds like reasonable speculation. How do the authors support their conclusion? They rely on the ipse dixit approach: “[W]e know from a long history of ed tech that this will not be the case.”

I don’t know this at all. Before attending law school, I taught for two years in a crowded middle school classroom in McDowell County, West Virginia, which the New York Times described as “the poorest in West Virginia, emblematic of entrenched American poverty for more than a half-century.” Access to AI would have made me a more effective teacher, benefiting my disadvantaged students immeasurably.

As a college student, I worked as a teaching assistant, reviewing essays written by freshman English students. I could have helped them much more if I’d had access to tools like ChatGPT. They would do a better job on some tasks, giving me more time for personal interaction.

Ethan Mollick had an advantage I lacked: the ability to use AI in the classroom. Dr. Mollick is a leading AI expert and the author of the invaluable “One Useful Thing” blog. Chapters 1 and 7 of his 2024 book, Co-Intelligence: Living and Working with AI, contain a detailed and convincing account of his lived experience with AI in education, along with a thorough analysis of the strengths and weaknesses of this powerful new tool.

The authors of The AI Con quoted Dr. Mollick, but did not mention his name nor his book, nor his invaluable blog, One Useful Thing, widely considered a leading resource on practical uses of AI.

The authors state that they wrote their book “in late 2024.” Co-Intelligence debuted on The New York Times bestseller list on April 3, 2024, its first week of publication, and remained there for weeks.

Even if the authors never saw a copy of Co-Intelligence before reading their book’s galleys, there’s no excuse for quoting Dr. Mollick in their manuscript and criticizing his ideas without including at least his name and a link to his highly respected blog, where he has been writing about Generative AI since November 2022. The omission of these basics is not the standard one would expect from authors who boast about their thorough research and citation practices.

The Bottom Line

The AI Con entertains and occasionally enlightens, but ultimately underdelivers. It’s a polemic disguised as analysis—more effective at scoring rhetorical points than helping readers navigate the real challenges of AI. For those looking to understand both the perils and promise of generative AI, better options abound. Here are a few, all referenced above, with links to facilitate online purchases:

Co-Intelligence: Living and Working with AI, by Ethan Mollick. Mollick acknowledges that the current generation of AI has significant flaws but also explains the current and potential future benefits.

How to Think About AI: A Guide for the Perplexed by Richard Susskind (reviewed recently at LLRX.com). He also has a balanced view of AI strengths and weaknesses.

AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference, by Arvind Narayanan and Sayash Kapoor (reviewed recently at LLRX.com). The prose style is not as smooth as that in The AI Con, but the analysis and conclusions are much better.