How AI will change scientific research?
Ajay K. Agrawal, John McHale, and Alexander Oettl examine whether AI can raise scientific productivity by helping researchers search enormous combinatorial spaces of questions, ideas, designs, and experiments, while recognizing that human judgment may remain essential for interpretation, causal reasoning, contextual understanding, and ethical trade-offs. The paper is primarily a theoretical and conceptual analysis, rather than an empirical evaluation of a new dataset or AI system. The authors’ methodology has four main elements:
- It models scientific production as a multi-stage knowledge-production process comprising question generation, idea generation, design generation, and testing.
- It incorporates AI into a Romer-style knowledge-production function, allowing AI to affect the productivity of each stage differently.
- It uses examples from fields such as biology, materials science, physics, mathematics, computer science, medicine, and economics to illustrate where AI is likely to be more or less effective.
- It develops task-based models of AI augmentation to analyze how AI expertise, worker skills, bottlenecks, and organizational design affect productivity and inequality.
The authors describe AI as a general-purpose meta-technology: a technology that can help develop other technologies by improving search, prediction, design, and experimentation. They use the idea of a “jagged frontier” to emphasize that AI’s capabilities are uneven across tasks and scientific domains.
Findings
- AI can increase scientific productivity by searching large and complex combinatorial spaces that are difficult for humans to explore, particularly in data-rich areas such as molecular design, genomics, and protein folding.
- Human judgment remains important for generating meaningful questions, interpreting anomalies, assessing causal explanations, resolving ethical trade-offs, and deciding whether an AI-generated result represents genuine understanding rather than an artefact.
- AI’s effects will differ across disciplines. Data-rich fields such as biology may benefit strongly, whereas fields such as theoretical physics may remain more dependent on conceptual reasoning and analogical insight.
- Bottlenecks matter. Improving one stage of research may have limited effects if another stage, such as physical experimentation, clinical trials, or causal validation, remains slow.
- AI expertise is a major complement. The productivity gains from AI depend not only on the quality of the technology but also on how many researchers possess the skills to use it effectively.
- Productivity and inequality may move in different directions. AI can raise overall research productivity while concentrating rewards among researchers with scarce domain knowledge or advanced AI skills.
- Complementary investments are essential, including training, workflow redesign, specialized research tools, autonomous experimentation, and improvements in the underlying AI technology.
AI in Science
Authors: Ajay K. Agrawal, John McHale, Alexander Oettl
From: University of Toronto, University of Galway, Georgia Institute of Technology
AI and the erosion of collective knowledge
Daron Acemoglu, Dingwen Kong, and Asuman Ozdaglar examine whether highly capable generative and agentic AI improve individual decisions while simultaneously weakening the human learning and collective knowledge needed for good decisions in the long run. The authors develop a dynamic theoretical model of individual decision-making and collective learning. The model distinguishes between:
- General knowledge, which is shared across a community. For example, knowledge of financial instruments, medical conditions, or economic institutions.
- Context-specific knowledge, which concerns an individual’s particular situation. For example, risk tolerance, symptoms, or personal constraints.
- Individuals make better decisions when they combine both types of knowledge. Their costly learning effort generates a private signal about their own circumstances and a public signal that contributes to the community’s stock of general knowledge. The model then introduces agentic AI as a provider of context-specific recommendations with varying levels of accuracy.
Findings
- More accurate AI recommendations improve the information directly available to individuals. However, because these recommendations substitute for part of the information people would otherwise acquire themselves, they reduce incentives for human learning.
- Lower human learning effort reduces the production of shared general knowledge, which future users also need in order to interpret context-specific information effectively.
- Under some conditions, the economy can enter a knowledge-collapse equilibrium in which general knowledge gradually vanishes, even though AI continues to provide high-quality personalized recommendations.
- The welfare effect of AI accuracy is therefore generally non-monotonic: moderate accuracy can improve welfare, but excessive accuracy can reduce welfare by eroding the knowledge base and learning incentives.
- Better aggregation and sharing of human-generated knowledge increase welfare and make knowledge collapse less likely.
- The model supports information-design policies that limit the effective precision of AI recommendations by, for example, deliberately adding some friction or “garbling” information to preserve incentives for human learning.
AI, Human Cognition and Knowledge Collapse
Authors: Daron Acemoglu, Dingwen Kong, Asuman Ozdaglar
From: MIT
