Using AI in Research | Stanford CS193T Guest Panel with Mark Algee-Hewitt, Megan Ma & Markus Covert
Stanford Online
Humanities at scale 0:05
Mark Algee-Hewitt describes digital humanities as the meeting of data science and literary study. He works with large corpora, using models to find patterns across thousands of texts, but also to ask questions about a single novel. He says artificial intelligence belongs in the pipeline as a middle layer between statistics and human interpretation, not as a replacement for reading.
Law practice and training 9:00
Megan Ma says the Lift Lab studies applied AI in private legal practice. The lab looks at evaluation, service delivery, and training. Its research suggests preference rankings can outperform rubrics in judgment-rich legal work. It also studies billable hours, how much lawyers edit AI output, and uses simulators for negotiation and deposition practice.
Advanced legal personas 16:01
Advanced methods for law include partner personas, or legal personas: personalized language models that behave like a specific person. The aim is not broad role-play but finer control, so a model can reflect how one negotiator leans rather than turning into a caricature. The work asks whether numerical steering can make a system match an individual with more care and less stereotype.
Cells and human questions 18:00
A bioengineering view treats disease as multi-factorial, with genetic, epigenetic, and environmental causes working together. The lab builds whole-cell simulations and a first computer model that uses all gene content to predict phenotype. One cell run can produce about a gig of data. The speaker then turns to larger questions about what new technology makes possible, why AI tools can hurt learning and mood, and how to give scientists, doctors, students, and humans more time to think, care, explore, and connect.
AI and legal work 32:00
AI was described as less useful for simply replacing legal work than for helping attorneys understand why they work the way they do. The point was not speed alone. It was about using the tool to ask what shaped a lawyer’s judgment and how work quality connects to past experience.
Transparency and norms 35:30
The discussion turned to recommendation letters, student trust, and the mismatch between personal judgment and outside expectations. One speaker said he would not use AI to write letters, only to recall facts and keep his voice intact. The group kept returning to transparency, because hidden use creates the real ethical trouble. They also agreed that fields need clear rules, not blanket bans, so people can disclose what AI did and what they themselves did.
Keep learning and checking 49:31
AI can help, but it does not replace the rest of what you came to learn. The speaker says you still need to build critical thinking, judgment, and your own experience. Saving time with AI should not turn into careless work. It should make room for careful checking, deeper thought, and time spent talking with smart people about the hardest questions.
Use AI as a foil 1:00:06
The strongest use of AI is not to let it do the task for you. It is to push your own thinking harder. The speaker describes writing with Claude by asking for missing citations, weak points, and analogies, then reading more and revising. The result was slower, but it produced stronger paragraphs and better ideas.
AI expands learning 1:06:01
Used carefully, AI can widen what a class or research project can do, not shrink it. One speaker says offloading mechanical game building let the class focus on what it really wanted to test. Another says curiosity matters most, and AI often works best as a first step that leads you to colleagues and deeper learning.
Guardrails and Stanford 1:08:01
A student asks how educational AI should be designed so students still do the work themselves. The answer praises Stanford teachers for shaping AI use in education and compares the moment to web search before Google, when many tools vanished. The advice is to take classes where people are building the field now.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.

