Most books written for general audiences about artificial intelligence have a problem: they either assume too much technical knowledge or too little intellectual curiosity. The ones that avoid both pitfalls tend not to be books about AI at all — they're books about intelligence, consciousness, and cognition that happen to illuminate exactly what AI systems can and cannot do. This list starts from those foundations and works outward.
The deepest book on intelligence I know that requires no technical background is Gödel, Escher, Bach: An Eternal Golden Braid by Douglas Hofstadter. It's long, it's strange, it uses dialogues between Achilles and a Tortoise to explore formal systems, and it remains the most illuminating account of self-reference, recursion, and consciousness that exists in the popular literature. The core question — how does meaning arise from symbol manipulation? — is precisely the question that matters most for understanding what large language models are doing and what they aren't. Hofstadter doesn't think current AI achieves genuine understanding, and he explains why in terms that are more precise than anything you'll read in a technology column. Start here if you want to think rather than just be informed.
The next book is Thinking, Fast and Slow by Daniel Kahneman. This is the definitive account of the two systems that drive human cognition: the fast, automatic, associative System 1, and the slow, deliberate, effortful System 2. AI systems — particularly the current generation of language models — are doing something that superficially resembles System 1 thinking: pattern-matching at enormous scale, producing outputs that feel intuitive and natural. What they lack is something like System 2: the capacity for deliberate reasoning that can override the fast pattern. Understanding this distinction — which Kahneman makes legible without any technical background required — gives you more insight into why AI sometimes gets things surprisingly right and sometimes confidently wrong than any amount of reading about neural network architecture.
Oliver Sacks's The Man Who Mistook His Wife for a Hat belongs on this list because it documents, with extraordinary precision, what happens when the brain's pattern-recognition systems are damaged or reorganized. His patients perceive the world through altered computational systems — they see the constituent features of a face but cannot assemble them into a face; they can describe a scene but cannot locate themselves within it. These case studies illuminate what perception and cognition are doing normally, which is to say: running a continuous inference about what reality is, based on incomplete data. Current AI systems do something structurally similar, and Sacks's case studies help you understand what it means for that inference to work or fail.
Richard Dawkins's The Selfish Gene is relevant here in a specific way. Dawkins's central argument is that evolution can be understood as an algorithm: genes that produce organisms that reproduce pass themselves forward; those that don't, don't. There is no designer, no intention, no understanding in the process — only selection operating on variation over time. This is the closest non-technical account I know of what machine learning does: optimization over enormous datasets producing systems with capabilities that were not explicitly programmed. The gene metaphor for the unit of selection maps interestingly onto the weight or parameter in a neural network, and Dawkins makes the algorithmic nature of the process vivid without requiring you to understand the math.
Finally, The Body Keeps the Score by Bessel van der Kolk contributes something the others don't: an account of intelligence that is fundamentally embodied and non-linguistic. Van der Kolk's patients have nervous systems that learned threat faster than language, and that knowledge is stored in the body, not in propositional memory. Current AI systems have no bodies, no nervous systems, no threat responses — they process language, and language is only a partial record of what humans think and know. Reading van der Kolk after the others on this list makes the limitation concrete: there is a vast category of human knowledge that cannot be captured in text, which means it cannot be learned from text, which means systems that learn only from text are missing something fundamental about human experience. That's not a criticism — it's a precise description of what AI currently is and isn't.
This reading order makes sense: Gödel Escher Bach for the foundational question of what intelligence is, Thinking Fast and Slow for the cognitive architecture, The Man Who Mistook His Wife for what happens when it breaks, The Selfish Gene for the algorithmic frame, The Body Keeps the Score for what's left out. The result is a picture of AI that's more accurate than most reporting provides — not because these books are about AI, but because they're about intelligence, which is the harder and more important question.