The Alignment Problem: Machine Learning and Human Values

Brian Christian’s The Alignment Problem focuses not only on the development of AI technologies but, most importantly—as the title of the book suggests—on how to align machine learning and AI use in general with human values, ethics, and behaviors. The thematic significance of the intersection of human and machine learning remains central throughout the book, highlighting both the potential benefits and pitfalls of these technologies as they become more integrated into daily human life.

Christian explores the intersection of human and machine learning through various lenses, including the need for representative training data, the risks of AI systems perpetuating existing biases, and the challenges of designing AI that can interpret and respond to complex human behaviors and intentions.

One of the critical aspects Christian discusses is the concept of training data. He illustrates how AI systems are only as good as the data from which they learn. For instance, face recognition technologies have improved significantly as datasets have become more inclusive. However, these advancements also raise concerns about surveillance, especially as the legislation for regulating the use of AI is missing in many countries around the world. This discussion underscores the importance of careful consideration in data collection to ensure AI systems do not reinforce existing inequalities.

Christian also tackles the application of AI in the criminal justice system, specifically focusing on risk assessment tools like COMPAS. These tools often use problematic proxies such as rearrest and reconviction rates, which might not accurately reflect an individual’s likelihood to re-offend. By relying on these flawed models, we risk the perpetuation of these biases, ensuring that certain demographics are unfairly targeted. These risks highlight a broader issue in AI development: the challenge of ensuring that machine learning models are aligned with up-to-date concepts of fairness and justice, rather than merely replicating past patterns.

Christian dedicates significant discussion to the concept of reinforcement learning and reward shaping. He uses examples from games like Atari and Go to demonstrate how AI systems can learn from mistakes and correct them. However, he notes that in real-world scenarios, mistakes can have irreversible consequences, stressing the need for AI systems capable of understanding the long-term implications of their actions and adapting their goals accordingly. The discussion reflects on the necessity for more sophisticated AI models that can recognize their impact on the environment and interact with it in more meaningful and ethical ways. However, citing various researchers, Christian also notes that AI submission to human will is not desirable:

In a follow-up study, led by fellow Berkeley PhD student Smitha Milli, the group dug further into the question ‘Should robots be obedient?’ Maybe, they wrote, people really are sometimes wrong about what they want, or do make bad choices for themselves. In that case, even the human ought to want the system to be ‘disobedient’—because it really might know better than you yourself. As Milli notes, ‘There are some times when you don’t actually want the system to be obedient to you. Like if you’ve just made a mistake—you know, I’m in my self-driving car and I accidentally hit manual driving mode. I do not want the car to turn off if I’m not paying attention’ (289-90).

While the autonomy of AI systems is a complex matter, which specialists are continuing to discuss and develop, Christian emphasizes the delicate balance required in designing these technologies. On one hand, AI must be sophisticated enough to recognize when it may need to override incorrect human commands for safety, as Milli suggests. On the other hand, it also needs to respect human decisions and ethical considerations, integrating into societal norms and legal frameworks. This balance poses significant challenges in AI development and Christian, along with researchers in the field, emphasize the need for systems that not only learn and adapt but also consider the broader implications of their actions in real-world contexts. The debate over AI autonomy encapsulates a broader conversation about how much control machines should have and under what circumstances.

Christian’s discussion of the intersection between human and machine learning brings to the forefront critical questions about trust, responsibility, and the limits of machine intervention in human affairs. As AI systems become more integrated into everyday life, their design and functioning must evolve to ensure they enhance human capabilities without undermining human autonomy.