At a time of rapid iteration in artificial intelligence (AI) technology, the key determinant of economic impact is no longer the capability of the models themselves, but the depth of application and the pathways of diffusion. Neil Thompson, Director of MIT's FutureTech and a Principal Research Scientist at MIT's Computer Science & Artificial Intelligence Laboratory (CSAIL), stated that the current productivity gains from AI remain limited. The core bottlenecks lie in the costs of enterprise implementation, process redesign, and organizational capabilities. He believes AI is reshaping not "jobs" but the "structure of tasks," which will have a profound impact on employment patterns and skill requirements. From a global perspective, AI could create new development opportunities through the dissemination of knowledge, but it could also exacerbate inequality due to the concentration of capital and technology. As AI becomes further embedded in the economic system, a trend of income distribution tilting towards capital is emerging, which has become a core issue policymakers must address. Dr. Neil Thompson currently serves as a Principal Research Scientist at MIT's CSAIL, is a Fellow of the Initiative on the Digital Economy (IDE) at the MIT Sloan School of Management, and leads the FutureTech research group. Under his leadership, the FutureTech team focuses on studying cutting-edge trends in AI development, analyzing how these trends support scientific progress and economic growth, and producing methodologically rigorous research to inform policy and industry decisions.
Excerpts from the dialogue:
Question: When will AI truly translate into measurable productivity gains? What are the key bottlenecks currently hindering this process? Neil Thompson: What we are seeing now is that the practical application of AI is still quite limited. Although the technology itself has made astonishing progress, the proportion of genuine adoption within enterprises and the breadth of its use remain limited. This means the impact on productivity is still relatively small. However, we can expect this impact to gradually materialize as applications continue to expand. We are indeed seeing an acceleration in adoption. Of course, there is a lag between initial adoption and full integration into production processes.
Question: So where do you see the bottlenecks? And since they may differ by country, specifically for the United States, what are the bottlenecks in AI application and adoption? Neil Thompson: We don't have completely definitive answers yet. My lab and other institutions are conducting research on this, and we have some preliminary assessments. One reason is that building these systems is often very costly, and companies typically need to make significant adjustments to their existing processes to use AI effectively, which itself takes time. For example, some data from manufacturing indicates that when companies start adopting AI, productivity can initially decrease because they have a set of highly optimized processes that need to be reconfigured. During this adjustment phase, productivity drops before it can recover and potentially increase. Another reason is that management is still figuring out what these systems can actually do and how to manage them correctly. When I speak with executives, they say the first time they launch an AI product might take 5 to 10 times longer than the tenth time because they are still learning—how to ensure safety, guarantee quality, what testing is required. Building these capabilities also takes time.
Question: There are many policymakers at the IMF annual meetings who are very concerned about AI's impact on employment, such as whether AI will replace a large number of jobs. From your perspective, is AI more likely to reshape tasks rather than completely replace jobs? What does this imply for how companies design workflows? Neil Thompson: I think the key to thinking about this is to understand automation at the 'task level' rather than the 'occupation level,' because what gets automated are tasks. A job typically consists of multiple tasks. Some are very easy for AI to perform, but many are very difficult. So, in answer to your question, I think both things will happen. On one hand, in some areas AI will become very proficient, replacing a lot of work previously done by humans—customer service is a very clear example already. But more broadly, we will see a large number of jobs being 'reshaped,' meaning some tasks change and some do not. When tasks are removed, several things can happen. One, more optimistically, is that people find new tasks, realizing that with these tools they can do new things. Even without new tasks, job reshaping doesn't necessarily mean it's good or bad for wages or employment. For example, if 30% of the tasks in your job are automated, and if the automated tasks are the ones requiring your most specialized skills, your wage might fall because more people can do the job, increasing competition. But simultaneously, the number of people doing that job might actually increase. Take the taxi industry. Being a taxi driver used to be a highly specialized profession requiring knowledge of city routes. Now, with GPS, that's no longer necessary. As a result, over recent decades, taxi drivers' wages have decreased because the specialization decreased, but the number of drivers has increased significantly. This contrasts with the common intuition that automation leads to job loss and lower wages; here, wages fell but employment rose. The opposite can also occur: if the automated tasks are the most basic ones, your wage might actually increase. For instance, in proofreading, much of the work involved spelling and grammar checks, which are now automated. What remains requires more complex logical and expressive skills, making the role more specialized, leading to higher wages but potentially fewer such jobs. So, the future changes involve not only which tasks are performed but also the skill level required for jobs.
Question: Regarding the issue of uneven development caused by AI, many policymakers are discussing this, especially the gap between developed and emerging markets. Do you think AI will widen this gap or create opportunities for these markets to leapfrog? Neil Thompson: I think both are possible. One aspect that makes me most optimistic is the way knowledge can diffuse. In many parts of the world, there is a shortage of doctors and engineers. If this expertise is embedded in AI tools, people even in remote areas could access crucial medical information, significantly improving living standards. Education is similar; AI can provide numerous examples to help people understand knowledge, enabling high-quality, personalized learning, which has great potential globally. But there is also a counter-trend. The development cost for AI models is extremely high, and they tend to become platform-based, meaning not every company develops its own; instead, a few platforms provide the service. This characteristic of high fixed costs and low marginal costs can push markets towards monopoly or oligopoly structures. This means that early entrants—companies or countries—gain an advantage, which could be detrimental to other regions.
Question: Shifting perspective from the macro to the industry level, for example, Musk's upcoming XChat, which reportedly integrates chatting, communication, and payments into one app. What is your view on this trend? What does it mean for AI adoption? Neil Thompson: This kind of integration is likely to increase the number of AI users, with many people perhaps not even realizing they are using AI. For example, in the US, many people using Google Search are already using AI-generated answers. So, the adoption of AI will accelerate, which is beneficial for users. But if we are talking about productivity gains, the key still lies within enterprises. What truly matters is how companies embed AI into their own processes, focusing on the most valuable parts. Therefore, deeper, industry-specific applications and platforms are more important than mere consumer apps.
Question: You are participating in a roundtable on AI and the new economy at the IMF meetings. What is the core message you most want to convey to policymakers? Neil Thompson: The core message I want to convey is that we are facing enormous change, accompanied by significant risks. One key issue is that as AI development and application advance, an increasing share of income is likely to flow to capital owners—such as owners of computing equipment and infrastructure—rather than to labor. This poses societal challenges because capital owners will become wealthier. If this capital is predominantly concentrated in developed countries, it also becomes a problem for developing countries, as key models are developed in advanced economies but used globally, potentially leading to a flow of value back to those countries. In other words, a major question is how the global distribution of value will undergo new changes.
Comments