GEnerative AI - think Dall.E, ChatGPT-4 and many more - is all the rage. Its remarkable successes, and sometimes catastrophic failures, have sparked important debates about both the scope and dangers of advanced forms of artificial intelligence. But what does this work actually reveal about natural intelligences like ours?
I'm a philosopher and cognitive scientist who has spent his career trying to understand how the human mind works. Drawing on research spanning psychology, neuroscience, and artificial intelligence, my search has led me to a view of how natural minds work that is both interestingly similar to, but also deeply different from, the core principles of generative AIs. By exploring this contrast we can better understand them both.
The AIs learn a generative model (hence their name) that allows them to predict patterns in different types of data or signals. What generative means there is that they learn enough about the deep regularities in a given data set that they can create plausible new versions of that kind of data for themselves. In the case of ChatGPT, the data is text. By knowing the many weak and strong patterns in a vast library of texts, ChatGPT can, when prompted, produce plausible versions of that kind of data in interesting ways, when shaped by user prompts. For example, a user can ask for a story about a black cat written in the style of Ernest Hemingway. But there are also AIs that specialize in other types of data, such as images, allowing them to create new paintings in the style of Picasso, for example.
What does this have to do with the human mind? According to many contemporary theories, the human brain has learned a model to also predict certain types of data. But in this case, the data to be predicted are the various floods of sensory information recorded by sensors in our eyes, ears and other perceptual organs. Now comes the crucial difference. Natural brains must learn to predict these sensory flows in a very special kind of context: the context in which the sensory information is used to select actions that help us survive and thrive in our world. This means that among the many things our brains learn to predict, a core subset concerns the ways in which our own actions in the world will change what we next perceive. For example, my brain has learned that if I accidentally step on my cat's tail, the next sensory stimuli I receive often involve whining, wriggling, and occasionally feelings of pain due to a well-deserved retaliatory scratch.
The story continues
Read more: AI and the rise of mediocrityThis type of learning has special virtues. It helps us separate cause from simple correlation. Seeing my cat is strongly related to seeing the furniture in my apartment. But neither causes the other to act. Stepping on my cat's tail, on the other hand, causes subsequent whining and scratching. Knowing the difference is critical if you are a being who must act in your world to bring about desired (or undesirable) effects. In other words, the generative model that makes natural predictions is limited by a known and biologically critical goal: the selection of the right actions to perform at the right times. That means knowing how things currently are and (crucially) how things will change and change if we act and intervene in the world in certain ways.
What do ChatGPT and today's other AIs look like compared to this understanding of human brains and minds? Most obviously, today's AIs specialize in predicting fairly specific types of data - strings of words, in the case of ChatGPT. At first glance, this suggests that ChatGPT is better thought of as a model of our textual output rather than (like biological brains) models of the world we live in. That would be a very significant difference indeed. But that step may be a little too fast. As the wealth of major and minor literature shows, words already express all kinds of patterns-patterns of appearance, taste, and sound, for example. This gives the generative AIs a real window into our world. However, the crucial ingredient is still missing: action. At best, text-predictive AIs get a kind of verbal fossil trail of the effects of our actions on the world. That trace consists of verbal descriptions of actions ("Andy stepped on his cat's tail") along with verbally formulated information about their typical effects and consequences. Despite this, the AIs have no practical ability to intervene in the world - so no way to test, evaluate and improve their own world model, the one that makes the predictions.
More from TIME
This is an important practical limitation. Rather, it is as if someone had access to a vast library of data on the form and outcomes of all previous experiments, but was unable to do anything themselves. But it can also have a deeper meaning. It is likely that this only happens by poking, prodding and generally intervene on our worlds, so that biological minds anchor their knowledge in the world it is meant to describe. By learning what causes what, and how different actions will affect our future worlds in different ways, we build a solid foundation for our own later understanding. It is that basis for actions and their consequences that later allows us to truly understand sentences like "The cat scratched the person who stepped on his tail." Our generative models - unlike those of generative AIs - are formed in the heat of action.
Can future AIs also build anchored models in this way? Could they conduct experiments in which they send reactions out into the world to see what effects those reactions have? Something similar is already happening in the context of online advertising, political campaigns and social media manipulation, where algorithms can launch ads, messages and reports and tailor their future behavior to the specific effects on buyers, voters and others. If more powerful AIs were to close the action loop in this way, they would begin to transform their currently passive and "secondhand" window into the human world into something closer to the kind of hold that active beings like us have on our worlds.
But even then, there would be other things missing. Many of the predictions that structure human experience involve our own internal physiological states. For example, we experience thirst and hunger in a highly anticipatory way, allowing us to remedy impending shortcomings in advance to stay within the appropriate zone for bodily integrity and survival. This means that we live in a world where some of our brain's predictions matter in a very special way. They are important because they allow us to continue to exist as the embodied, energy-metabolizing beings that we are. We humans also benefit greatly from collective practices of culture, science, and the arts, which allow us to share our knowledge and explore and test our own best models of ourselves and our world.
Furthermore, we humans are what we might call 'knowing knowers' - we portray ourselves as people with knowledge and beliefs, and we have slowly designed the complex worlds of art, science and technology to test and improve our own knowledge and beliefs . . For example, we can write articles that make claims that are quickly challenged by others, and then conduct experiments to try to resolve the disagreements. In all these ways (even if we bracket obvious but currently intractable questions about "true conscious awareness") there seems to be a very large gulf separating our special forms of knowing and understanding from anything that has hitherto has been achieved by the AIs.
Could AIs one day become prediction machines with a survival instinct, making basic predictions that proactively try to create and maintain the conditions for their own existence? Can they therefore become increasingly autonomous, where they can protect their own hardware and production and extract energy where necessary? Can they form a community and invent some kind of culture? Could they begin to model themselves as beings with beliefs and opinions? There is nothing in their current situation that drives them in this familiar direction. But of course none of these dimensions are off limits. If changes were to occur along all or some of these important missing dimensions, we might yet glimpse the soul of a new machine.
Contact us at [email protected].
