3 Questions: What is the best path forward for AI in academia?

As artificial intelligence capabilities continue to advance, universities are grappling with complex questions about how to navigate these developments. In a recent essay, MIT Statistics and Data Science Center Director Sasha Rakhlin synthesizes discussions about AI's impact on academic research, particularly in mathematics, statistics, machine learning, and engineering. Here are three key questions Rakhlin addresses and potential solutions.

1. What is changing in research, and why is it happening so quickly?

Rakhlin notes that AI models have reached human-level performance in mathematics, producing new research results and solving previously unsolvable problems. The speed and reliability of verification processes play a crucial role in AI progress within a discipline. Faster and more reliable verification allows AI systems to generate candidates, learn from outcomes, and improve, leading to a compounding process that accelerates progress.

2. How should departments rethink academic credit and graduate training?

Rakhlin argues that polished papers are becoming a weaker signal of individual expertise as AI capabilities grow. Departments should reconsider what they reward, focusing on asking good questions, replication, synthesis of ideas, informative negative results, and shared datasets. Evaluation should also take into account what a researcher contributed and their intellectual responsibility, including when substantial parts of the work were performed by AI.

This approach should guide hiring, promotion, and funding decisions, as well as be made explicit for current and incoming PhD students.

3. What should MIT and other universities build now?

Rakhlin sees an opportunity for universities to build a strategic asset in their accumulated knowledge and experience. Universities should be able to pursue long-term questions, share results openly, and evaluate claims independently. Partnerships with industry are essential, but universities should not rely on commercial priorities alone.

To foster technological independence, universities must invest in creating shared AI research infrastructure. This could involve a future where MIT's laboratories function as a single scientific organism, connected through shared AI research tools and workflows. These workflows would capture hypotheses, interventions, outcomes, failures, and interpretations, allowing AI agents to recognize relevant advances from other laboratories, connect researchers, propose benchmarks, and help iterate.

Establishing a system that traces this process can preserve the lineage of ideas and make contributions.