Book review: Co-intelligence

Regular readers of this blog will know that I have posted reasonably frequently about generative AI and its current and future impacts on education (as I see them). It would be hard not to write about that topic, even if I didn't have a keen interest in how we can use generative AI for the good of student learning. Most of my reading on generative AI has been online (Substack, blogs, etc.), and most of my learning has been hands-on. As a change of pace though, I decided to pick up and read Ethan Mollick's 2024 book Co-Intelligence.

Mollick has been one of the key thought leaders on the use of generative AI over the last few years, and for that reason he should be taken seriously. This book collates his thoughts up to the end of 2023, and provides an excellent primer for those who have had little exposure to generative AI. Readers who are very familiar with generative AI, and especially those who follow Mollick on X and/or those who have read his research, will find little new in the book. However, for other readers, there is lots to digest.

The book first explains some of the basics of generative AI. This is not a book on the computer science or mathematic underpinnings of generative AI though. It is more of an overview of how AI models work and what to expect (and not to expect) from them. That section finishes with four very helpful principles or 'rules' for using generative AI: (1) Always invite AI to the table; (2) Be the human in the loop; (3) Treat AI like a person (but tell it what kind of person it is); and (4) Assume this is the worst AI you will ever use. Those four principles are important to keep in mind whenever we use generative AI. The book then pivots, with the rest of the book outlining a number of broad use cases for generative AI.

Although the book has a generally positive tone, it is not a rave about the value of generative AI. Mollick has realistic expectations about what generative AI can and cannot do, and happily shares examples of where generative AI has definite room to improve. The use cases are clear, and helpful, and highlight some of the limits of the technology. In this way, Mollick does his best to ensure that readers are never overawed by generative AI. For example:

And you can't figure out why an AI is generating a hallucination by asking it. It is not conscious of its own processes. So if you ask it to explain itself, the AI will appear to give you the right answer, but it will have nothing to do with the process that generated the original result. The system has no way of explaining its decisions, or even knowing what those decisions were. Instead it is (you guessed it) merely generating text that it thinks will make you happy in response to your query. LLMs are not generally optimized to say "I don't know" when they don't have enough information. Instead, they will give you an answer, expressing confidence.

It is the passages like that one that are provide the greatest value from this book for the general reader. For me, though, I most enjoyed the sections on generative AI in education ('AI as a Tutor', and 'AI as a Coach'). As a professor at Wharton, regularly using generative AI in teaching, Mollick has good insights into how generative AI will likely impact education in the future. He writes that:

...the ways in which AI will impact education in the near future are likely to be counterintuitive. They won't replace teachers but will make classrooms more necessary. They may force us to learn more facts, not fewer, in school. And they will destroy the way we teach before they improve it.

For a teacher, that is equal parts reassuring and terrifying. However, Mollick clearly sees the value in education, even in the face of generative AI. That may be self-serving, but it doesn't seem to be a blinkered view. In particular, I loved this bit about the 'paradox of knowledge acquisition':

Large Language Models seem to have accumulated and mastered a lot of collective human knowledge. This vast and tappable storehouse of knowledge is now at everyone's fingertips. So it might seem logical that teaching basic facts has become obsolete. Yet it turns out the exact opposite is true.

This is the paradox of knowledge acquisition in the age of AI; we may think we don't need to work to memorize and amass basic skills, or build up a storehouse of fundamental knowledge - after all, this is what the AI is good at. Foundational skills, always tedious to learn, seem to be obsolete. And they might be, if there was a shortcut to being an expert. But the path to expertise requires a grounding in facts...

The issue is that in order to learn to think critically, problem-solve, understand abstract concepts, reason through novel problems, and evaluate the AI's output, we need subject matter expertise... We need expert humans in the loop.

I wish that more educators (and their managers) would read this book carefully, especially those sections above. Simply outsourcing tasks to generative AI is not helping students to learn and become future experts. They need a grounding in foundational knowledge - basic concepts, models, and intuitions - before they can use that knowledge in critical thinking tasks.

Notice that the last part of the last quote links the example back to the second of Mollick's four principles. These links to the principles are sprinkled throughout the book, although at times I wish they had been made a bit more explicit. Nevertheless, Mollick writes in a style that is easy to read, and the book is suffused with interesting anecdotes and references to recent and relevant research (by Mollick, and by other researchers). Those aspects of the book will likely become out of date in due course, but not yet. Although I read the book more than a year after its release, it still seemed mostly contemporary, in a space where the technology is moving incredibly fast. The book will likely age well, and the four principles will remain relevant for some time to come. 

This book was a genuine pleasure to read, and is essential reading right now. I strongly recommend it to anyone looking for a basic grounding in generative AI.