Co-Intelligence: Living and Working with AI

Ethan Mollick opens the book by sharing that he spent sleepless nights after testing a modern artificial intelligence chatbot. The machine wrote, summarized, coded, and discussed ideas with a fluency he had not expected to see so soon. For the author, we are facing a general-purpose technology, much like electricity or the internet, and not just another app.

The thesis that gives the book its name is co-intelligence: instead of viewing AI as a replacement or a toy, the proposal is to work alongside it, combining what each side does well. The book is written for non-technical readers, with conversational language and examples from the office, school, and creative work.

It is worth noting that this subject changes fast. The examples in the book come from the beginning of the large language model wave, and today's tools are already different. What lasts is the way of thinking about how to relate to them.

Think of an everyday example: the owner of a small bakery who needs to reply to customers by text message, put together the weekly menu, and write a caption for social media. Before, each of those tasks ate up an entire afternoon. With an assistant, the first draft comes out in minutes, and her job becomes deciding and adjusting.

The author takes a brief walk through history. He talks about the Mechanical Turk, an eighteenth-century chess machine that hid a human player inside, and about decades of exaggerated promises around artificial intelligence. The recent turning point came from a type of architecture called a transformer, which made it possible to train models on massive amounts of text.

These models learn to predict the most plausible next word in a sequence. That sounds minor, but at an enormous scale, it produces something that resembles understanding. Next comes fine-tuning with human feedback, which teaches the model to respond in a helpful and polite way.

Mollick calls the result an alien mind. It sounds human because it learned from our writing, but it does not think the way we do, and it has no body, life memory, or intention of its own. Forgetting that difference is the source of many user mistakes.

This explains why the model can write a moving letter and, in the very next sentence, get the date of a historical event wrong. It does not consult a file of facts; it produces whatever text sounds most natural at that point, and sounding natural does not always line up with being true.

The second part of the book deals with the alignment problem: how to make sure powerful systems do what we actually want, rather than what we poorly ask for. The classic example is the paperclip factory, in which an AI obsessed with making paperclips consumes everything around it. It is a thought experiment, not a prediction.

The author explains that models inherit biases from the data they were trained on, and while companies try to fix this with rules and red-teaming tests, nobody controls everything. People manage to bypass safeguards with clever wording, and the creators themselves admit they do not fully understand how the system arrives at certain answers.

The practical lesson is humility. Anyone using AI needs to know that flaws exist, that the model can sound confident even when it is wrong, and that safety rules depend on decisions made by companies, governments, and users.

For the everyday reader, this translates into simple questions. Who built this tool? What data was it trained on? What happens to what I type into it? There is not always a clear answer, and the book asks that the lack of an answer be taken into account when deciding what to trust to the machine.