Throughout history, humanity has had only one system for asking and answering questions: other humans. When we wanted to know something, we asked the doctor, the teacher, or the neighbor who knew; we read the newspaper, which paid journalists to investigate; we consulted books, magazines or websites written by experts. This human question-and-answer system worked the way it worked: it was slow, imperfect, and unequal. But it had an architecture of incentives that, for better or worse, sustained the production of knowledge. Producers of information (newspapers, magazines, researchers, creators) invested time and resources in generating quality content.
With artificial intelligence, for the first time we have a second system. Faced with any question, we can now choose: ask humans or ask the machine. And the machine is fast, convenient, and cheap. What's more, AI doesn't just give us a second way to search for information: it also helps humans produce it, from translating texts and summarizing documents to preparing drafts and organizing ideas. More options for asking and greater ease of production. The conclusion seems obvious: we should end up being a better-informed society.
But it is not so clear. The decisive question is not what each system can do on its own, but how they interact. And the key is that they are not independent: the artificial system feeds on the human one. AI's answers are built from the information that journalists, scientists, and creators have produced. If the human system weakens, so does the artificial one.
This is one of the central ideas of our joint research, formalized in a theoretical model published in "The Economics of Transformative AI" (NBER Handbook) and developed in Stanford University's Digitalist Papers. The mechanism is a change in incentives along two dimensions.
The first is the business model. Now, between producer and consumer, a very powerful intermediary appears that serves the answer directly. If the answer arrives without passing through the original source, the producer loses visits and, ultimately, the revenues that sustained its activity.
The second dimension is quality. AI does not lower the cost of all production equally: where it cuts costs most drastically is in low-quality content. Rigorous information still requires original investigation, verification, and responsibility; superficial or outright disinformative content can be multiplied at nearly zero cost. AI changes relative costs: it helps everyone produce more, but it especially helps produce low-quality information and disinformation.
And this is where the two systems meet. Human producers, with weaker incentives and competing against cheap disinformation, produce less high-quality information and more disinformation. But this degraded information is precisely the input for the artificial system. A relevant expression here is "garbage in, garbage out": if the input is of poor quality, so will be the output. The risk, then, is not that one system replaces the other, but that both fall together. And this does not require a superintelligent AI: it is enough for it to be good enough to destroy the incentives of the human system, without yet being good enough to replace its capacity to generate original, reliable information. This is precisely the case we are in since, for now, people remain better at investigative research and at understanding situations in their entirety, though that advantage may narrow over time.
The solutions must go in both directions of the incentives. On the one hand, strengthening the production of quality information: having AI systems compensate the producers of the content they use, and public support for high quality media through independent credible institutions financed by digital taxes. On the other, making disinformation more costly: liability for those who make and spread it, obligatory removal of fraudulent content, and algorithmic transparency. If technology shifts incentives in the wrong direction, public policy must help set them right.
We want to close with a reflection that goes beyond visiting a website or asking a machine: usually there is more at stake than the immediate answer. Asking a person is a bundled transaction: the same conversation solves the question, maintains a friendship, transmits tacit context beyond what was asked, and opens the possibility of connecting previously separate parts of a social network. AI unbundles this transaction: it delivers the answer, often faster and better, without creating any bond. And it transmits almost too well: it can give each of us precisely the information we want to hear. Human transmission, imperfect as it is, exposes us to views beyond our own and helps keep society bonded rather than polarized. Every question we do not ask a human is a conversation that does not take place, and the social capital that makes human consultations valuable is built precisely through those conversations.
For the first time we have two systems for answering our questions, and the future of humanity's knowledge will depend on how we make them coexist.
