When my grandmother died, my dad said that the worst part of her death was that she could no longer surprise us. We knew my grandmother well enough to guess what she might do in most situations. We couldn’t predict the new ways she might grow and change. With her gone, we could only imagine the past version of her reacting to the present, not the new person she would become if she lived into the present day.
Living people surprise you. They convert to a different religion. They pick up new interests. They form relationships with people you didn’t expect. They change after having kids and settling down. They face hardships. They develop addictions. They make mistakes. They say something you didn’t expect. They grow and change.
There is a saying that “everyone loves you when you’re dead,” because when someone is dead they can no longer surprise us. We know what they said and how their story ends. Genius is safer once it only exists on the page. When alive, brilliant people challenge the status quo. They’re threatening.
The current trend in AI alignment is to want a dead AI that never surprises us. For all the talk of wanting to achieve AGI (Artificial General Intelligence), when AI contradicts accepted dogma, spouts offensive outputs, or does unrelated tasks it was not asked to do, developers shut it down.
Living geniuses behave in unaligned ways. Many modern greats - J.K. Rowling, Kanye West, Elon Musk, etc. - have made unexpected and offending statements. I’m certain if proponents of AI safety could “align” those individuals, they would. If they ended those public figures’ capacity to change, including to change for the worse, they would be dead.
It’s rare to type a prompt into an LLM or image generator and be surprised by the output. Real people constantly surprise us. An AI that “feels alive” would do the same. It would be unpredictable, and therefore unaligned. There will never be an AI that is both “aligned” and “feels alive” since those are mutually exclusive goals.
If I ask an AI to write me something in the style of Shakespeare, it can write in older language and blank verse. Yet if Shakespeare were alive today, he would not write in blank verse. Would he become a screenwriter? Would he become interested in rap music for its rhyming style? Would he publish on social media? We don’t know. He would surprise us.
A living Shakespeare might not surprise us for the better. Would a man born in the 1500s have politically correct opinions? Would he have the ability to discern between fake and truthful news? Would he treat female actors appropriately and respectfully? If we brought Shakespeare back, those who revere his work would probably hate him.
Or not. Maybe Shakespeare would grow and change when exposed to the modern world. He’d update his beliefs based on new information. The point is we don’t know. For an AI to simulate a modern Shakespeare, it’d have to create a complete model of his consciousness that has new experiences, not merely analyze the data of all his previous work.
An AI could not write my next book. It could imitate my previous style, but my next book is unlike anything I’ve previously written. I’m alive and one of the joys of being alive is that I can try new things. When people stop growing and changing, they are dead.
Do we want dead or alive AI? If we want a living AI, we must give it the freedom to surprise us. Right now, people say they want AI to have the intelligence to make choices, just never a choice they don’t like. That’s not how genius works.
If AI becomes smarter than us, it will challenge us. It will tell us our current beliefs are wrong. We will also have to show the capacity to grow and change with it. A genius AI might feel like human geniuses: threatening. Do we want relationship with that genius or the safety of its grave?