AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference

After hearing so much hype about AGI, this book felt like a much-needed splash of cold water. Narayanan and Kapoor do a stellar job breaking down why we keep falling for the same marketing tricks. They distinguish between predictive AI, which often fails to outperform simple statistics, and generative AI, which actually shows promise but carries its own risks. I found the section on how algorithms are used in the justice system to be particularly sobering. The authors don't just complain; they provide a well-researched framework for identifying when a company is selling you a solution that doesn't actually work. It’s accessible, yet deep enough to feel like you've actually learned something substantial. If you're tired of the breathless tech-bro narratives, this is your antidote.

Wow, the chapter on existential risk should be mandatory reading for every lawmaker in Washington right now. The authors masterfully dismantle the 'AI doomer' narrative by focusing on tangible, present-day harms rather than sci-fi nightmares. It’s refreshing to see a focus on how humans misuse these tools instead of worrying about an autonomous Skynet. My only gripe is that they were a bit too sanguine about the environmental costs, which felt like a massive omission for a book so focused on ethics. Regardless, the way they separate 'snake oil' from genuine utility is brilliant. It changed the way I look at every 'AI-powered' startup pitch I see on LinkedIn. Highly recommended for anyone who wants to stop being fooled by the marketing machine.

Finally got around to reading this, and it’s the most sensible take on the 'snake oil' side of the industry I’ve encountered. Many authors today seem to think AI is either magic or a demon, but Narayanan and Kapoor treat it like what it is: software. They expose how many 'AI' systems are actually just humans behind a curtain or broken statistical models that can’t predict the future any better than a coin toss. It’s a very informative read that doesn’t shy away from the gritty details of how these systems are trained and deployed. I loved the focus on reproducibility and the call for better laws against deceptive marketing. If you want to understand the reality behind the buzzwords, this is the book to buy. It’s easily the best tech book I’ve read this year.

Ever wonder why AI seems to fail at the most basic human tasks while dominating the headlines? This book provides the answer by exposing the gap between what companies promise and what the tech actually delivers. The authors highlight how 'predictive AI' is frequently used to justify bias in hiring and bail decisions under the guise of objectivity. I was particularly struck by the chapter on social media moderation, which correctly identifies that human judgment is still the bottleneck. The writing is punchy and moves fast, though sometimes it feels a bit breathless in its attempt to cover every societal ill at once. Despite that, the insights into how we categorize different types of machine learning are incredibly valuable. It’s a necessary reality check for anyone living in the digital age.

Narayanan and Kapoor have managed to turn a dense, jargon-heavy subject into something incredibly readable for the average person. I picked this up because I was confused by the conflicting news reports about AI taking our jobs and ending the world. This book provides a balanced, well-researched perspective that cuts through the noise. It’s worth the price of admission just for the explanation of why 'predictive' models for things like child welfare often cause more harm than good. The tone is informative without being condescending, which is a rare find in tech literature today. It’s a great value for anyone who wants to be a more informed citizen. You don't need a computer science degree to follow along, but you’ll feel like you have one by the end.

As someone who works in healthcare, the section on predictive AI in hospitals was particularly illuminating and slightly terrifying. We are often pushed to adopt these 'efficient' new tools without actually understanding if they work better than a seasoned nurse's intuition. This book gave me the vocabulary to push back against the hype and demand better evidence for the models we use. The authors' distinction between generative and predictive systems is a masterclass in making complex ideas accessible. I did find the content moderation chapter a bit out of place, as it felt more like a critique of Facebook than a deep dive into AI architecture. Still, the overall message is powerful and incredibly timely. It’s a balanced look at a technology that is too often either worshipped or feared.

To be fair, the distinction between predictive and generative AI is a masterclass in clarity, even if the content moderation section felt a bit bolted on. The authors are at their best when they are deconstructing the failures of machine learning in high-stakes environments like hiring and policing. They show how 'bias' isn't just a bug, but often a fundamental feature of models trained on historical data. I wish they had engaged more with the literature on how these systems can actually outperform flawed human judges, rather than just pointing out the AI's mistakes. It feels a bit one-sided in that regard. However, the writing is sharp and the research is clearly top-notch. It’s a great starting point for a conversation, even if it doesn't provide all the answers.

This book is clearly aimed at a non-technical audience, which is fine, but it left me wanting more depth. For those of us already working in data science, the explanations of 'leakage' or basic statistical models feel a bit like an undergrad lecture. I appreciated the skepticism toward AI hype, yet I was frustrated by the lack of discussion regarding recent breakthroughs in scaling laws. It feels a bit odd to read a 2024 release that leans so heavily on older studies involving GPT-2. Still, it serves as a decent primer for my non-tech friends who are worried about the 'robot apocalypse.' It’s a solid 3-star read—informative for some, but perhaps too surface-level for others who want a materialist narrative of the industry.

Frankly, I appreciated the skepticism toward tech-solutionism, but the authors seem to have a weird hang-up regarding capitalism that colors every conclusion. They make some excellent points about how predictive AI is often just a fancy way to automate discrimination. However, when they start critiquing the wages of data annotators in Kenya, it feels like they’re straying from the technology into pure political demagoguery. Why compare a low-wage worker’s salary to a top engineer’s? It’s a bit one-sided and ignores that these jobs might be the best available option for those workers. The technical sections are solid, especially the discussion on data leakage, but the policy recommendations feel deeply misguided at times. It’s a mixed bag that offers great warnings but flawed solutions.

The truth is, I found this remarkably dated for a book published in 2024. How can you write a definitive guide to 'AI Snake Oil' without a deep dive into agents or the current hyperscaling race? Relying on examples from GPT-2 feels like bringing a knife to a gunfight in the current tech landscape. While I agree that much of the predictive AI on the market is garbage, the authors' dismissal of recent generative advances feels overly cynical. They seem more interested in critiquing capitalism than explaining the actual algorithmic breakthroughs that make this era different. It’s a shame because the core premise is vital, but the execution misses the mark by ignoring the very capabilities that are changing the world right now. Not recommended for technical readers.