The Alignment Problem does this in a new, narrative way. For Christian, the alignment problem has been the driving question behind machine learning since the early 20th century. He introduces the core components almost unnoticeably, since the components–samples, labels, reward functions, optimization algorithms–are all discussed through historical examples that function as pericopes drawn out of a grander story. The effect is the feeling of an active, human project that we’re already a part of by virtue of being human–the feeling that our humanity precedes and is integral to the problems of machine learning.

Every machine learning question raised has analogous and related human questions, whether in learning theory or epistemology or otherwise. By the end of the book, the reader is equipped to enter the arena of AI ethics, not because they become technical experts, but because they have a human story for AI and can locate the alignment problem. It’s not in an arcane field of zeros and ones, but precisely in the thicket of human questions that characterize our lives.

Take, for example, one problem Christian introduces in the first movement of his book, called “Prophecy.” The computer-vision technology used to label our photos or unlock our phones has a history of not accurately identifying minority groups (relative to the US population) or failing to identify them as human at all. Christian’s investigation reveals that this is a classic problem of poor sampling techniques. People with darker skin tones were (and commonly still are) underrepresented in the training data used for these algorithms. One facet of the alignment problem, then, is the question How do we prevent AI from acting with and perpetuating social bias and inequity? As AI ethics hit mainstream conversations, questions like this one are likely near the top of the agenda. But Christian doesn’t let us miss that this question is intimately tied up in both a machine-learning question (What can go wrong when my data is poorly sampled?) but also, crucially, an unsettling human question (Do my limited experiences inhibit my judgment?). An effective approach to AI ethics demands that we parse them into both these machine-learning and human questions.

Christian’s second and third movements, “Agency” and “Normativity,” show how our relationship with the machines we train has evolved with our goals for machine intelligence. “Agency” traces how we began building our likeness into the machines we teach. We simulated human characteristics like curiosity, intuition, and expectation in machine learning in order to generalize machine intelligence, eventually enabling these models to beat us in chess and to explore virtual worlds. “Normativity” then shows how our dynamic with machines has changed as we’ve introduced more complex tasks. If we want these machines to learn tasks too complex for us to fully articulate–like driving a car–then we need them to learn to imitate our actions or anticipate our needs. As we know from children, animals, and new hires, imitation without correct inference of what exactly we’re trying to accomplish can be disastrous. Likewise, once a machine is granted license to interpret, misunderstandings and unexpected possibilities emerge. Controlling these possibilities requires new kinds of training, training that entails intimate interaction with machines. It’s sharing control of the wheel (literally) until it’s hard to tell who’s driving. It’s completing each other’s sentences.

Here the imagination runs wild. Most of our collective imagination about machine learning and AI has been formed by (often apocalyptic) sci-fi stories: 2001: A Space Odyssey, Total Recall, Minority Report, Her, Ex Machina, The Terminator series, etc. Each of these offers an answer to the question: What could go wrong with AI? But for any answer to that question, a follow-up question is warranted: How do we prevent that?

How, in this context, is inhospitably technical. Tempting as it might be to entrust these questions to our legislative body, that would be asking the same group of people who struggled to understand that Facebook made money by selling ads to now understand neural networks. And most of us don’t feel any better equipped to take on the math and magic behind AI. But we can’t avoid the alignment problem that Brian Christian took the time to investigate in painstaking, compelling narrative detail. We can’t avoid it because ChatGPT (with the help of South Park) has forced the alignment problem back into public conversation.

ChatGPT isn’t alone, either. Google has since released its competitor, Bard. Facebook released Llama. (Yes, “Llama”.) A stampede of these large language models (LLMs) are emerging as if from some recently beached ark. And the open-source community is now able to apply, adapt, and connect these LLMs to other applications. We can train them for particular use cases–to answer questions about our company’s training documentation or to summarize a Slack conversation. AI is moving with dizzying velocity. But the field is split on how to proceed, especially with the alignment problem looming and unresolved.