Cutting Through the AI Hype: Predictive vs. Decision-Making AI ============================================================== Sam and Sophie unpack the hype around AI, separating what actually works from what's just snake oil. They dive into the dangers of using AI for high-stakes decisions like hiring and criminal justice, and why we need to ask better questions before trusting the algorithms. ---------------------------------------- SAM: Hey there, and welcome back to 7 Minute Books. I'm Sam, and today we're diving into a book that's all about cutting through the hype: 'AI Snake Oil' by Arvind Narayanan and Sayash Kapoor. So, Sophie, you read this one too, right? SOPHIE: I did, and honestly, it's one of the most grounding books on AI I've come across. It's not a technical manual or a doomsday prophecy. It's a practical guide for anyone who wants to understand what AI can and can't do, so we can make smarter choices about it in our own lives. SAM: Absolutely. And the core argument is so simple, but it kind of flips the whole conversation on its head. The authors say the most hyped AI applications are often the ones that work the least well. SOPHIE: Right. They split AI into two categories. There's predictive AI, which is all about finding patterns in data to forecast outcomes. That's behind things like recommendation systems and language models. Then there's decision-making AI, which is supposed to replace human judgment in high-stakes situations, like hiring, policing, and or criminal sentencing. SAM: And that's where the snake oil comes in. The predictive stuff works reasonably well, but the decision-making stuff? It's largely a mirage. SOPHIE: Exactly. Because predictive AI can tell you what's likely to happen, but it can't tell you why, and it definitely can't tell you what you should do about it. When you ask AI to make decisions about people, you're asking it to understand causation, context, and values. And it's completely blind to those. SAM: It's like asking a calculator to judge a pie-eating contest. It might count the pies, but it has no idea which one tastes better. SOPHIE: That's a great analogy. And the book is packed with real-world examples that show exactly why that fails. Take automated hiring systems, for instance. They're supposed to remove human bias, but they actually bake it in. SAM: Right, because if a company historically hired mostly men for engineering roles, the AI learns that maleness is a predictor of success. So it systematically downgrades female candidates. It's not being sexist in a conscious way, but the bias is in the training data, and the AI just amplifies it. SOPHIE: And it's the same with predictive policing. Train an algorithm on historical crime data, and it'll over-police minority neighborhoods because that's where past arrests happened. Then those arrests confirm the pattern, and you get a vicious cycle. SAM: It's like the algorithm is creating the future it's predicting. The book has this great term for that, feedback loops. SOPHIE: Yes, and that's one of the key reasons why AI fails in high-stakes contexts. There are also issues like distribution shift, where the future doesn't look like the past, and proxy variables, where AI uses crude stand-ins for what it's really trying to measure. SAM: Like using the prestige of your university as a proxy for competence, even though plenty of brilliant people go to less famous schools. SOPHIE: And then there's the biggest problem of all, values. AI doesn't have values. When it decides who gets hired or paroled or approved for a loan, it's making trade-offs between competing values, like efficiency versus equity. But it has no way to weigh those choices. SAM: So we're handing over these deeply human decisions to a system that's essentially a parrot on a calculator. It can recite patterns, but it's not actually thinking. SOPHIE: Exactly. And that's why the authors are so adamant that we need to ask better questions before trusting AI. They give us a checklist, What exactly is it predicting? How accurate is it really? What data was it trained on, and does that data represent the people it'll be used on? SAM: And who's accountable when it makes a mistake? That's a big one, especially when you see these systems used in criminal justice. There are risk assessment tools that are supposed to predict recidivism, but studies show they're no more accurate than random people, and they're biased against Black defendants. SOPHIE: It's terrifying, because judges get a false sense of certainty from these algorithms. They might defer to the machine even when their own judgment would be better. SAM: So what about the generative AI boom? The book covers ChatGPT and those models too, right? SOPHIE: It does, and the authors are cautiously optimistic. They're useful for brainstorming and drafting, but they have no real understanding. They're just generating text that's statistically similar to what they were trained on. They can argue for anything, regardless of whether it's true. SAM: So they're like a supercharged parrot. They can combine words in new ways, but there's no meaning behind it. SOPHIE: Exactly. And that makes them great for some tasks and dangerous for others, like giving medical or legal advice. The key is to not anthropomorphize them. SAM: So what's the takeaway here? It's not 'AI is bad' or 'AI is good.' It's more like, 'AI is a tool, and it can be used well or poorly.' SOPHIE: Yes, and the book is a call to action. We can't just blindly trust AI claims. We need to demand evidence, transparency, and accountability. And we need to remember that these are political decisions, not just technical ones. We have the power to shape how AI is used. SAM: And that's what I'm taking away from this book. We don't have to be helpless against the hype. We can ask the right questions and make informed choices. SOPHIE: Absolutely. And if you want to dig into these ideas even more, the whole library is over at 7minutebooks.com/app, with over 6,000 fiction and nonfiction titles you can read or listen to in any language. It's just $2.99 a month, $9.99 a year, or $19.99 for lifetime access. SOPHIE: AI Snake Oil reminds us that the real danger isn't the technology itself, but the belief that it can solve problems that are fundamentally human. We'll see you in the next one.