2. Role of AI in Research - 2

Summary

The discussion examines how AI can support research, what researchers must continue to do themselves, and how education and research practices may need to adapt.

AI changes how research can be done

• The guest argues that AI is more than a conventional tool for speeding up an existing task. By lowering the cost of activities such as formal verification, it can make new kinds of work feasible.
• Researchers can use AI to explore prior work and identify open questions. Its conversational ability is especially useful: researchers can ask what lies between two established results and whether the gap is worth investigating.
• AI can generate many possible ideas quickly. The researcher’s role is to judge which ideas are significant, feasible, and aligned with their goals.

Human judgment remains essential

• AI can be creative and productive, but it can also be overly agreeable. Because it is inclined to please users, it may praise weak ideas or endorse a researcher’s proposed direction.
• Researchers should not ask AI to verify its own suggestions or outsource their judgment of what makes an idea worthwhile. Developing “taste” for good research questions takes practice and remains a central human responsibility.
• If many researchers rely on the same models and ask similar questions, their ideas may converge. This could reduce the diversity of perspectives that helps research produce unexpected advances.
• The guest does not regard current AI as thinking in the same way as a human expert. He describes it as effective at making connections across a broad body of knowledge and using external tools, but not as possessing original thought or an independent sense of what is valuable.

AI in programming languages and formal verification

• The guest works on programming languages and on reasoning about whether programs satisfy specifications. Formal verification is particularly important in safety-critical systems, such as aircraft fly-by-wire systems.
• Formal verification has a long history, but its use has been limited by the expertise and expense required. Verification tools can provide precise results, yet their feedback may be cryptic and difficult to act on.
• AI can generate hypotheses and possible proof strategies, while formal tools check those proposals rigorously. This combination can substantially reduce the time needed for some verification work.
• The guest describes a research project in which AI-assisted work appeared to produce extensions much faster than the team had expected. He emphasizes, however, that researchers still need to decide which problems and results matter.

Guidance for young researchers

• Learn the fundamentals before relying on AI. Beginners should first work through core skills themselves—for example, writing proofs by hand before using AI with a formal-verification system. Skipping the difficult learning stage can mean missing the underlying concepts.
• Use AI to support, not replace, your reasoning. It can help explore possibilities or carry out work once you can define the problem and assess the result.
• Use it to improve communication. A researcher can provide rough or fragmented sentences and ask AI to shape them into clearer prose. The researcher remains responsible for the claims and final text.
• Use it for suitable technical and analytical tasks. AI can assist with data preparation, analysis, coding, and producing graphs, but the researcher must interpret the findings and evaluate their significance.
• Retain ownership of research choices. Selecting questions, weighing competing directions, and deciding what counts as a meaningful contribution require motivation and judgment that the guest does not recommend handing over to AI.
• Reflect on AI’s performance. The guest recommends considering where AI helped and where it did not, particularly in open-ended research work.

Collaboration, authorship, and responsibility

• The guest uses AI interactively—as something to test ideas against and, in his field, to help work through formal proofs when paired with the right tools.
• He cautions that calling AI a “collaborator” does not make it an author. Authorship entails responsibility for the work, including accountability when results are wrong. Researchers cannot shift that responsibility to AI.

Education and research policy

• The guest says policies about AI use are evolving and can be difficult to enforce, especially in settings such as peer review where AI use may not be transparent.
• He advises against relying on AI-detection tools to determine whether writing was generated by AI, describing such tools as unreliable and noting that the boundary between AI assistance and ordinary writing aids can be unclear.
• In his teaching, he is moving more assessment into in-person settings. One approach assigns much of the grade to work done at home, then tests students in a lab or viva-style setting—without internet—on a straightforward extension of that work.
• In another course, students are expected to complete some work without LLMs and must use them for group projects. The projects have higher expectations, and students reflect on how AI helped or failed to help.
• His broader position is that education should continue to teach underlying skills without AI, just as students still learn arithmetic by hand even though calculators exist.

Possible effects on future research

• By lowering the cost of certain tasks, AI may enable researchers to attempt larger and more ambitious questions than were previously practical.
• AI may also help researchers divide problems into smaller parts and collaborate across a wider community.
• The guest expects AI to accelerate research, while emphasizing that the volume of unresolved problems is not likely to disappear. As AI raises research capacity, expectations for researchers may rise too.

Speakers and sources featured

• Host/interviewer: An unnamed presenter leading the research-methodology discussion and asking questions for young researchers.
• Professor K.C. Sivaramakrishnan: Guest from the Computer Science Department at IIT Madras, discussing AI and research, especially programming languages and formal verification.
• Other references mentioned: Anthropic and OpenAI; Lean, a system used for formal verification; and mathematician Terrence Tao, whose AI-assisted mathematical collaboration was cited as an example.