By Diogo Costa, President,
Foundation for Economic Education
The public conversation has been inundated with predictions of impending doom from AI. I do not dismiss them. But I was reminded of how little we are actually able to see or imagine our technological futures when I came across a recent piece Matt Ridley wrote for the IEA.
Ridley describes a paradox that seems to accompany every major technological change:
Looking backwards, innovation looks inevitable and therefore predictable: once you invent the drone, warfare changes; once you invent the internet, social media arrives; once you invent electricity, the light bulb is unavoidable.
But go back to the moment when people were actually looking forward, and almost no one saw any of that. Even in our time, the pioneers of the internet were not predicting that billions of dollars would flow into search and social media.
Ridley remembers sitting on a House of Lords committee that produced a well-regarded report on artificial intelligence in 2018. Their vision of AI’s future was already so distant from what has been happening that the phrase “large language model” does not appear once in the report.
The same was true of landmark AI legislation I followed fairly closely in Brazil. There was no room in it for LLMs either. The only use of the word “language” was to require AI systems to use “simple language” so that people could understand them. What does that even mean in the age of Claude and ChatGPT?
As Ridley puts it, “there is no such thing as an expert on the future.” We have no clear idea what the AI future will look like five or 10 years from now. We do not know what business models will have appeared, what jobs will have been created, what institutions will have adapted, or what concepts and words will have entered our vocabulary. We will have to watch it unfold.
I do not blame anyone for trying to see ahead of their time. That is what our brains are constantly doing. Every time a disruptive technology arrives, we sketch scenarios for what we might do with it: the printing press, the steam engine, the electrical grid, the personal computer, the internet, and now AI. And AI is the master disruptor, so we are sketching faster than ever.
Extinction or abundance. Mass unemployment or radically augmented human capacity. Perhaps imagining these possibilities is part of how we take the future seriously.
But the exercise could learn something from Ridley’s provocation. There is a natural engineering mindset behind scenario building. We begin with an extrapolation of technological capabilities. We ask what machines will soon be able to do, then reason forward toward how society will react and what society will become.
What these scenarios usually lack is a theory of learning.
If we cannot successfully predict what is going to happen, then instead of asking only what will happen, we should also ask how we will find out. What is the process of social discovery through which we learn what a new technology is actually for?
That is a question economics has a great deal to say about. Economic thinking carries its own theory of learning. We call it entrepreneurship.
When we talk about learning, we are not describing some amorphous collective moving from mental state A to mental state B. Even when we speak of the “wisdom of crowds,” the crowd is not itself the unit of wisdom. Individuals are. Each one of us possesses bits and fragments of knowledge that others do not know. Much of that knowledge cannot even be written down. Learning happens when individuals act on what they know and are able to understand and bear the consequences.
Entrepreneurship is a learning process combined with a creative process. Entrepreneurs are constantly forming conjectures about a world that does not exist yet. They act with purpose toward futures that cannot be known in advance because it is their actions that help bring those futures into existence. Entrepreneurial futures are so difficult to predict because they are endogenous to the actions of the people trying to imagine it. The future is not merely hidden from us; it just doesn’t exist yet.
Entrepreneurship makes economics thicker, and it gives us something more interesting to say in the AI debate.
The key word in those debates is misalignment. And the arguments about AI misalignment turn out to be about two sets of problems. The first problem is imperfect specification. As much as you can tell an AI model what you want, or what its objectives are, you can never tell it completely and unambiguously. Those are hard limitations on what we can formulate. So the specification the model is able to follow will not fully correspond to the underlying intentionality. That is a hard engineering problem.
The second misalignment problem touches on the possible conflicts among objectives and parties. Perhaps the AI will have wants and objectives of its own. Or perhaps we will perfectly specify our objective to an AI model, but we will still have to deal with the question of whose objectives they are. A plural society disagrees about ends and values by definition. Individuals and firms, governments and communities can have different ends, as well as different conceptions of what counts as harm, justice, dignity, and how acceptable the tradeoffs are between them. Societies do not have a single social objective waiting to be translated into code. AI researchers are concerned that a superintelligence might pursue goals different from ours. I am more concerned that some people may acquire the power to decide what “our” goals are. That is a hard political problem.
The engineering problem begins from an objective and asks whether we can specify it correctly. The political problem asks whose objectives should govern when people disagree.
But there is a third problem, and I suspect it is the most neglected. It is an economic problem. We do not merely lack the technical ability to specify precisely what we want from AI, nor do we merely disagree with one another about what we want. In many cases, we do not yet know what we will want from AI.
We do not know which applications will prove transformative and which will prove useless. Which new problems will emerge and which old problems will disappear. Which capabilities people will value, which risks they will tolerate, and which institutions will come into being. Not even our preferences are fixed outside this process. They change as we encounter possibilities that previously did not exist.
That is not a problem any laboratory can solve in advance because much of the relevant knowledge does not yet exist. It will be generated through millions of encounters between technology and people with different problems, values, capabilities, and imaginations.
Social alignment, then, cannot be a one-time act of design. It has to be an adaptive process of conjecture and criticism, experimentation and error-correction. Entrepreneurship is the process through which ends, as well as means, are discovered.
When economists are invited into the AI conversation, they usually talk about labor markets: which jobs disappear, what happens to wages and productivity, etc. Those are important questions. But they are not the deepest economic problem either. The deeper problem is the discovery process around AI.
Most entrepreneurial conjectures will be wrong. Most conceivable uses of a transformative technology will go nowhere. Many business models will fail. Some applications will be socially rejected. Others will reveal risks that have to be constrained. But those failures around the discovery process are the actual discovery process. Action is not what we take after we have learned. Learning happens through action.
Markets are not simply information-processing systems. They are future-generating systems. The uses of AI that will matter most probably do not exist yet, which is why they cannot be widely predicted. They have to be invented by people acting on their own visions of the future, most of which will turn out to be wrong. That is what markets are for.
I do not dismiss the risks superintelligence can pose. But I am seeing many scenarios that end by centralizing a powerful technology in the hands of a few, and that lack a theory of learning. Centralization may reduce some forms of variation, but under radical uncertainty, variation is also how a society learns. Fewer conjectures might mean fewer mistakes. But fewer conjectures also mean fewer experiments. Since we do not know in advance which mistakes those are, a process with fewer experiments will produce less knowledge about which risks are real.
If we knew exactly where AI was going, perhaps we could simply choose the safest road, but we are going to have to discover it.