1. Role of AI in Research - 1

Summary

The discussion explores how generative AI can support research while emphasizing that researchers must retain expertise, judgment, and responsibility. The central recommendation is to use AI to extend human capabilities—not as a substitute for learning or independent thinking.

Main ideas and guidance

• AI is more than “autocomplete,” but it is not an independent researcher. Modern AI can help break complex problems into smaller tasks and test many possible approaches. Its strongest contributions are often automating repetitive work and expanding the range of ideas a researcher can examine. Humans still need to frame the problem, judge which results matter, and decide what to pursue.
• Learn the subject before relying on AI to augment your work. The guest’s guiding principle is “build expertise before augmentation.” For example, early-career researchers should learn how to conduct a literature review—how to search, identify relevant terms, evaluate papers, and organize findings—before using AI to accelerate that work. Otherwise, AI can become a crutch, and researchers may miss the learning that comes from doing the work themselves.
• Research is about learning how to solve problems, not just producing answers. Tackling a difficult question develops research skills: finding collaborators, choosing methods, learning experiments, and drawing ideas from other fields. Outsourcing the entire process risks losing these opportunities to build expertise.
• Use AI, but verify its output. The guest repeatedly stresses: “Never trust; always verify.” AI can produce plausible but incorrect claims, code, or references. The interview mentions reports of fabricated citations in published work; the exact statistic quoted is presented as an example, not independently verified in the discussion.
  • Check whether claims are actually supported by the cited papers.
  • Look for the relevant passage, figure, or data point in the source.
  • Test AI-generated code, including on different datasets or against expected results.
  • Apply basic sanity checks to calculations, hypotheses, and analyses.
  • Reject claims that cannot be substantiated.
• Researchers remain accountable for their work. AI is not a co-author who can take responsibility for a paper. Authors remain responsible for every claim, figure, equation, and word, even if AI assisted with them. The guest compares this with other high-stakes settings, such as medical decisions, where responsibility and the role of AI remain difficult questions.
• AI can help across research tasks, but use should depend on the task and the researcher’s training. Potential uses include literature searching, exploring or challenging hypotheses, initial data analysis, generating code, reviewing a draft for weaknesses, and proofreading. The more a task depends on domain judgment or helps build essential skills, the more cautious researchers should be about handing it over.

A suggested approach for researchers

1. Build AI literacy. Understand, at a basic level, how AI works and what its limitations and risks are, including bias and errors.
2. Learn core research skills without depending on AI. Practice literature reviews, problem-solving, analysis, and other essential methods.
3. Use AI to augment work you understand. Apply it to save time, explore alternatives, or automate routine tasks.
4. Check the output. Verify sources, reasoning, code, and results against primary references and your own expertise.
5. Keep responsibility and authorship human. Do not present AI-generated work as your own without appropriate acknowledgment, and do not rely on AI where you cannot evaluate its correctness.
6. Protect essential skills. Identify which abilities are non-negotiable in your field. A chemist’s chemistry expertise, for instance, cannot be replaced by AI; some programming tasks might be delegated if the researcher can still test and assess the results.
7. Preserve your own voice and judgment. The guest recommends writing the first draft yourself, then asking AI for feedback or improvements rather than having it turn bullet points into a manuscript.

Institutional policies: red, amber, and green uses

The guest describes a draft policy approach developed at IIT Madras:

• Red — prohibited: Examples include putting confidential or peer-reviewed manuscripts into AI systems, entering confidential data, or submitting an AI-written essay as one’s own without acknowledgment.
• Green — generally acceptable: Asking AI to explain a concept, teach a topic, or generate practice problems that the researcher then solves independently.
• Amber — requires judgment and verification: Examples include using AI for literature reviews, improving code, or debating a hypothesis. These uses may be helpful, but they can create an “illusion of competence” if the researcher cannot assess the output.

Risks of overreliance

The guest warns that routinely outsourcing thinking may weaken problem-solving and analytical skills, foster dependence, and make researchers’ writing and ideas less distinctive. AI-assisted writing may also lead to a more uniform style. The recommendation is not to avoid AI altogether, but to decide which skills should be strengthened through practice rather than routinely delegated.

Access and the future of research

AI may raise the baseline level of research support by giving more people access to tools that were previously unavailable. However, the guest distinguishes improved access from true democratization: access to computing power, advanced models, reliable internet, and infrastructure will remain unequal, and geopolitical restrictions may affect availability. Developing domestic AI systems is presented as one way to reduce dependence on systems built elsewhere.

The guest expects AI to have a growing role in research and describes human–AI work as collaborative. Examples discussed include AI contributing to mathematical and algorithmic work that humans then guide or improve. The overall analogy is a crane operator: AI supplies greater capacity, but the human decides what to move, where it should go, and when to stop.

Speakers and sources featured

Speakers

• Interviewer/host: Unnamed presenter of the research-methodology episode.
• Professor Karthik: Guest from the Department of Data Science and Artificial Intelligence at IIT Madras; no surname is given in the subtitles.

People and sources cited in the discussion

• Arvind Narayanan (auto-transcribed in the subtitles as “Arvin Naran”), in connection with an ICML keynote.
• Arjun Krishna, a collaborator whose article on AI use is discussed.
• Donald Knuth (auto-transcribed as “Kanut”), in connection with Claude’s work on a mathematical theorem.
• Anil Ananthaswamy, discussed in relation to retaining a personal commitment to writing.
• DeepMind, in the example of an AI-assisted matrix-multiplication algorithm later improved by mathematicians.
• AlphaFold and Google Co-Scientist, cited as examples of AI’s impact and potential in scientific research.
• An unidentified Quanta magazine article about human–AI collaboration.
• A CBSE committee and AI-literacy framework for schools, mentioned by the guest in connection with curriculum development.