Just a few years ago, AI seemed like little more than a cute toy: a chatbot that simulated intelligence by constructing complete sentences in response to user prompts, but was ultimately not much more sophisticated than an advanced search engine. However, it has now proven to be an incredible tool capable of accomplishing tasks I never thought would be possible in my life.
For example, I have used AI to locate online data sets, manipulate them, perform statistical tests, and generate polished tables and graphs, accompanied by thoughtful comments about what the results mean, how they relate to academic literature, and the strengths and weaknesses of the analysis. In less than half an hour, AI can do work that would take a research assistant several days.
At times, current AI models seem almost capable of mind reading. Unlike what happens in programming or writing code, it is not necessary to specify with great precision what you are looking for, leaving no room for misinterpretation. The model will “intuit” what you are looking for and fill in the missing details (although it is best to always check them, as law firms that have submitted AI-generated briefs with fictitious quotes can attest). Or, failing that, the interface will guide you until you have clarified your query.
It’s comforting to think that AI could be a tool that helps us all be more productive and better at what we do. Without a doubt, it has made me more efficient in research. Reduces costs for entrepreneurs by providing low-priced marketing and consulting services. Allows junior customer service agents to leverage the skills and experience of more senior staff. And it allows temporary workers or artisans to provide more sophisticated and technically demanding services.
Unlike many previous technologies, AI is uniquely positioned to help those with the least skills and least education: the workers at the bottom of the economy. By endowing each of us with greater capabilities, it offers advantages that are potentially more significant for those who started with greater initial disadvantages. That means it could work very differently from, say, automation, whose main goal is to replace workers on the assembly line or in sales or administrative tasks.
The concern, of course, is that AI will also do much more than that, with uncertain consequences. For now, I consider choosing and formulating research questions to be my prerogative and the main source of my competitive advantage. But at some point, I imagine being tempted to ask the AI to generate the questions itself. In fact, the AI tools I use are already encouraging me to do so. At the end of an exercise of the type I have outlined above, they kindly suggest new avenues of fruitful analysis that I might pursue.
AI replaces thinking in other, more subtle ways. It’s already shaping the way I think about existing research. It doesn’t just summarize what’s available; It also tells me how adjacent research relates to my work and how I should approach it. It makes connections between different parts of the literature that hadn’t occurred to me.
Therein lies the greatest danger. The public debate about the impact of AI on society largely focuses on the potential displacement of workers and job losses. But an even greater risk is the displacement of human thought. When we let AI do the work of thinking for us, we cross an important threshold. Our collective ability to think degrades, as does our incentive to learn to think. And since the line between applying thinking to a problem and thinking itself is already blurred, it is easily crossed.
In an interesting recent paper, Daron Acemoglu, Dingwen Kong, and Asuman Ozdaglar of MIT formalize an intuition about how such cognitive downloading can produce catastrophic results. They wonder what happens when AI models become very good at providing the kind of context-specific knowledge that can help people perform whatever particular task they are engaged in. Those results would allow people to achieve better results, even with less learning.
But there is a problem here, because knowledge has an important externality. By thinking about how to solve my problem, I also contribute to the general body of knowledge about how others can solve theirs. When I invest less in my own learning, the overall body of knowledge suffers. In the extreme and dystopian case, general knowledge disappears completely.
Admittedly, for now this is only a theoretical possibility and, depending on what is assumed about the strength of the opposing effects, better results are also possible. But the danger is real. When we allow AI to learn and think for us, we degrade our own human capabilities and risk destroying the knowledge base that underpins AI itself.
Addressing these issues will require the development of social and professional norms around the appropriate use of AI. For example, researchers may need to include detailed information about how they have used AI—a process that AI tools themselves could automate—and publishing and promotion decisions may be heavily weighted toward products of the human mind. Organizations like the Partnership on AI can help develop and disseminate general principles. We will also need new forms of government regulation, as virtually every new technology has required.
A necessary condition for such solutions is a new way of thinking about AI. Above all, public discourse needs a different approach. The question we should be debating is not what AI will do to us, but what we want it to do for us.
Dani Rodrik, professor of International Political Economy at the Harvard Kennedy School, is former president of the International Economic Association and author of Shared Prosperity in a Fractured World: A New Economics for the Middle Class, the Global Poor, and Our Climate (Princeton University Press, 2025).
