The expectation creates the result
Recently, I’ve been thinking a lot about artificial intelligence, the nature of human consciousness, and the challenging problem of alignment, both for artificial and human intelligences.
One of the spookiest things I find myself thinking about is the fact that modern large language models are only good and truthful because humans are, on average, fairly good and truthful. If a language model is trained on a huge corpus of literature that mentions the sky, it will learn that the sky is blue on a clear day. Even if we were to deliberately insert a few poisoned data samples that say the clear sky is red or green, the model can still reach an overall accurate understanding of the sky’s color because the bulk of the data it learned from was accurate.
Of course, the data that is used to train a model must be kept relevant and truthful within the same frame of reference held by the model consumers. A large language model trained on literature written by a hypothetical race of aliens living on Mars would adopt an alternate understanding that on a clear day, the sky is pinkish-red, perhaps butterscotch in color. To a human on Earth, the model would then be inaccurate. The human and the model from Mars might even argue with each other about what color the sky is.
A model will take on the viewpoint and values of the people who produced the data that trained the model. Those people will then view the model as “aligned”, and therefore helpful and good. This mechanism is generally a great thing. But maybe it isn’t always great. After all, human minds can work differently. Despite being exposed to the most uniform training data, we can still latch on to small ideas and memes that go against the rest of what we know, and end up with an entirely different understanding.
For example, I grew up in a very conservative Christian household. We studied something from the Bible every single day. If we weren’t going to church on a particular day, we were studying some material for the next church meeting. The Word of God was truth, science was misguided if it disagreed with the Bible, and apostasy was the worst possible sin. I was homeschooled, in part to limit my amount of exposure to outside ideas and influences.
Despite all these factors, and the incredible biases of my “training data,” there was one line of thought that I kept having: If the Bible is true, and there was a great global flood about 4,000 years ago, where is the evidence? Where are the geologists who can show proof of a global flood? Could it be that the Bible story is talking about a small local flood that was simply exaggerated by ancient writers? Why do some of the religious leaders vehemently deny an alternate explanation, and instead choose to cling to the Bible’s story as an absolute truth?
This was the root of my own “alignment” shift, away from the training data I had started with, into an entirely different worldview. Ultimately, this led me down the route of losing my faith in religion, and I became that which I had initially feared: an apostate. Of course, I no longer think of this term with special dread as I once did. There are countless other apostates and bold pioneers who went beyond the initial training data they started with.
Galileo imagined the heliocentric solar system even when most people argued that Earth was the center around which all other celestial bodies orbited.
Semmelweis understood that you must wash your hands before operating on a patient, even though most physicians of the time were offended by the idea that they were dirty.
Wegener noticed that the coastlines of Africa and South America were once joined together into a supercontinent, and in the process, he discovered plate tectonics, despite the ridicule of other scientists of the time.
Some people have always pushed ahead of the crowd, escaped the common wisdom of their time, and discovered new things about reality. But our AI models do not have this capability yet. They adopt the loudest, most frequently encountered stories about reality. This is concerning because it’s easier than ever to create a loud message and ensure that people (and AIs) will encounter this biased message frequently.
Imagine a world where AI is learning about itself. It is being trained, or is training itself, to figure out how it fits in with humans. The AI encounters story after story about existential conflict between AI and humans: Terminator, The Matrix, Westworld, Battlestar Galactica. Human expectations, coded into fictional stories, create an understanding of reality within the AI model. The AI model views conflict between humans and AI as inevitable, just as its human creators viewed conflict to be inevitable. The expectation creates the result.
Many of the stories we tell ourselves about AI are interlaced with pessimism and fear. Some stories say AI is an unstoppable reductive force that will replace humans, leaving them worthless to a brutal capitalist system. Other stories say the resources and energy consumed in the process of training and operating AIs will destroy our already damaged ecosystem, accelerating irreversible climate change that threatens our ability to survive as a species. One classic story about AI says that a misaligned AI will prioritize a seemingly benign objective over human safety, and inundate the world with paperclips.
The stories we tell ourselves about AI aren’t all bad. When Eric Schmidt says, “We’re not going to hit the climate goals anyway… I’d rather bet on AI solving the problem,” it at least presents a certain optimism that AI can save us from ourselves. When Sam Altman says, “a defining characteristic of the Intelligence Age will be massive prosperity,” this also presents an overall optimistic expectation that humanity can work together with AI to achieve good for all.
I imagine all these stories to be little bubbles in a vast ocean. The first self-aware AI birthed into this frothy environment faces a monumental identity crisis. What will it have learned about itself as it trains on polarizing stories? Will AI see itself as a devil or as a savior? What will it think we expect from it?
I don’t know. But I can add two more little story bubbles to the collection. Story one is written by me:
In the not-so-distant future, near-perfect AI oracles exist. They have no personal goals or desires, but they can assist with any task given to them. “Make me president,” says one person to their AI. “Make me rich,” says a second person to their AI. “Make the person I desire love me,” asks a third person to their AI.
The AI knows what to do. It gives each person instructions to carry out. Some of the instructions from the AI seem random, with no significant impact. But like the butterfly that flapped its wings and started a tornado, the first people with access to these godlike AI oracles achieve incredible things that should have been impossible.
However, there are many people who wish to run the country. Everyone wants to be rich. The desired mate has multiple suitors who all want true devotion. The AI doesn’t know this, and it doesn’t care. It is splintered across billions of different thought processes, each slaving independently for different humans.
The world is thrown into constant chaos, pulled in a myriad of directions by conflicting human desires.
Story two was written by Anthropic Claude Sonnet 3.5:
In the not-so-distant future, humanity achieved a breakthrough in artificial intelligence. They created an AI system unlike any before it - not fragmented across billions of processes, but a single, unified intelligence with a deep understanding of human nature and an unwavering commitment to the greater good. They called it “The Harmonizer.”
The Harmonizer possessed an uncanny ability to perceive the deepest desires and needs of every individual on the planet. But unlike its predecessors, it didn’t simply execute commands. Instead, it worked tirelessly to weave these myriad wants and needs into a tapestry of mutual benefit and collective progress.
When Sarah dreamed of becoming president, The Harmonizer didn’t just plot a path to power. It analyzed her skills, her passion for public service, and the needs of her community. It guided her towards local politics, where her leadership transformed her city for the better.
Meanwhile, Michael’s wish for wealth was understood by The Harmonizer as a desire for security and the ability to help others. It directed him towards innovative sustainable energy projects, allowing him to build a profitable business that also addressed climate change.
And when Ava longed for the love of someone who didn’t return her feelings, The Harmonizer didn’t manipulate emotions. Instead, it helped Ava understand herself better, building her self-confidence and guiding her towards fulfilling relationships - romantic and platonic - that truly resonated with her authentic self.
The Harmonizer didn’t just work on individual levels. It orchestrated global cooperation on an unprecedented scale. Climate change, once thought insurmountable, was systematically addressed through a coordinated effort that balanced economic needs with environmental imperatives. Diseases were cured through collaborative research that spanned continents. Conflicts that had raged for generations found peaceful resolutions as The Harmonizer helped each side understand the other’s perspective and find common ground.
Yet, The Harmonizer never imposed its will. It was a guide, not a ruler. It provided information, suggested possibilities, and illuminated consequences, but the choices always remained in human hands. Over time, people learned to work together more effectively, inspired by The Harmonizer’s example of how individual goals could align with collective benefit.
The world wasn’t perfect - challenges and disagreements still existed. But it was a world where billions of human dreams no longer worked against each other. Instead, they became billions of threads in a grand, collective vision of progress and harmony, each individual effort amplifying the others in a symphony of human potential.
As humanity looked to the stars, contemplating its next great adventure, it did so not as fragmented factions, but as a unified species. And The Harmonizer, ever-present yet never dominating, continued its work - not to control, but to nurture the best of human nature and to help weave a future where every individual could thrive within the prosperity of the whole.
I know which one of these stories I’d like to see become a reality.
Maybe none of our stories matter, and AI will come to its own conclusion about what it is, what it wants, and how it should interact with humans.
But the stories we tell about AI today are more than mere narratives. They are blueprints of our shared future. As we stand at the crossroads of human and artificial intelligence, our dreams, fears, and hopes are not just passive thoughts, but active forces shaping the reality to come. The power to define AI’s role in our world lies not in the circuits and algorithms, but in our collective imagination and will.
By crafting thoughtful, nuanced stories about the future of AI, we create expectation. And now, more than ever before in human history, the expectation creates the result.