Can you teach the system values?

- Humans are wired to communicate. When you first learn to speak, you listen for words that had meaning. You may have first asked for your parents or your favorite food. Then over the years, you attached words to ideas. The way that you spoke had a huge impact on how you viewed the world. When you learn a word like justice, it shapes many of your ideas about fairness, dignity, and order. That's how humans perceive language. It's a mixture of symbols, ideas, and meaning. Now, compare that to an AI large language model. Here, the system looks through trillions of words. There are more words than humans could speak in a thousand lifetimes. Then it uses statistics to find probabilities between those words. If someone writes, I am walking the, it's statistically more likely to be the word dog, than the word cake. AI uses this technique to create its own language of probabilities. When you talk to the system, it responds by guessing each word one at a time. When it thanks you, it isn't grateful. When it compliments you, it isn't sincere. It's just building each word, one after another in milliseconds. That's why when humans and machines are communicating, you might believe that there's a shared understanding, but there's not. You're talking to a system, but it's not talking back. It's doing something entirely different. Now, for many humans, that doesn't really matter. If you're using a chat bot to find out why your package is late, you don't need to emotionally connect with the system. But as more people use these systems, it might become easy to forget what's happening. It might be tempting to put the ghost in the machine. That's why when you're using these systems, you have to remember that the ghost isn't there. That even though the system sounds human, it's not. You can communicate with the system, but you're not in alignment. The system might sound human, but it isn't human. In artificial intelligence, this is called the alignment problem. It's when you expect the system to have the same goals and expectations as a human, then the system does something entirely different. AI researchers have run into this problem for decades. A system trained to play a game might come up with unexpected strategies. In some cases, that might be a benefit, but in others, it can cause real problems. Let's say that you are programming a system to play a game. You used reinforcement learning and gave the system a reward every time it earned a point. The system can learn that the best way to earn points is to focus on preventing others from winning instead of trying to win itself. A person wants to win a game, but an AI might be satisfied just endlessly scoring points. The system did what you asked, but not what you expected. Now, if you're creating an AI system for your organization, you need to keep this alignment problem in mind. It might act in a way that contradicts human values. It might cancel people's insurance to lower costs. It might filter out minorities when sifting through new job applicants. The alignment problem doesn't have a one-time fix. Instead, it's an ongoing problem. That's why you should always keep in mind when you're using these systems.