Stanford researchers use AI to challenge thinking, not replace it

Cynthia Bailey introduces three Stanford researchers: Mark Algee-Hewitt from English, Megan Ma from the law school, and Markus Covert from bioengineering. Each uses computational methods or AI in a different research setting.

Algee-Hewitt describes his work as the meeting of data science and the humanities. His research center employs undergraduate interns, offering students opportunities to participate in computational humanities projects.

He calls the approach reading at scale: models can find patterns across thousands of texts or examine a single novel along dimensions a reader cannot easily observe. Related projects map geographical references and historical correspondence networks.

For his group, generative AI is a middle layer between a statistical foundation and human interpretation. People still have to interpret both the literary data and what a model is doing, he says.

His project “Style in the Age of AI” asks how a model’s writing style affects meaning. That matters particularly in humanities research, he says, because words are not just a vehicle for reporting results but part of how meaning is made.

Algee-Hewitt contrasts disciplines where results are often expressed through numbers with literary scholarship, where the wording itself participates in making the argument. He asks what changes when generative AI supplies that wording.

A pilot examined about 2,000 articles to ask whether disciplines have recognizable styles independently of their subjects. A larger study is expanding to 10,000 articles across 18 disciplines; topic modeling initially placed philosophy and computer science far apart.

Using grammar, sentence length and named-entity counts rather than subject matter, a classifier identified the discipline about 43% of the time. Algee-Hewitt compares that with roughly 5% chance performance and says the approach is still being improved.

The team then prompted a local language model to rewrite philosophy articles in a computer-science style. A style classifier assigned much of the rewritten material to computer science, though some remained in a long tail that the team is still examining.

A separate content-based topic model also moved the rewritten philosophy texts slightly toward computer science. Algee-Hewitt says the preliminary result suggests a computationally detectable relationship between style and meaning, not merely a change in surface grammar.

Megan Ma introduces Stanford’s Legal Innovation through Frontier Technology, or LIFT, Lab, which studies applied AI in legal practice. A central question is how to define the quality of legal work when the client cannot always directly judge what they have purchased.

Ma says legal quality is difficult to evaluate against a single ground truth because legal work is judgment-rich and clients build relationships with individual lawyers. Her lab studies evaluation, delivery of services, training and other new methods.

In the lab’s JudgmentBench work, preference rankings—choosing the better of two outputs—far outperformed rubrics, Ma reports. She says the result points to the value of lawyer intuition and asks whether a larger set of preferences could help derive a rubric for model training.

AI productivity also puts pressure on the billable-hour model. Ma uses the example of work that once took ten hours taking two with AI, and asks how firms should think about margins and exposure across different kinds of legal work.

Her group plans an observational study with legal-AI providers, examining editing histories and where humans intervene. It wants to distinguish work left untouched because an agent handled it well from work that receives little attention because it is low value.

Training is another concern: Ma says lawyers acquire judgment by doing practical work, yet some routine tasks traditionally given to junior lawyers may disappear. Her lab is exploring simulations to give them practice without the same stakes.

The lab built negotiation and deposition simulators; Ma says the deposition tool is open source. A user can upload a discovery document or witness statement to create an AI witness, then practice questioning as often as needed.

With input from senior litigators, the simulator offers 14 witness archetypes, including combative, evasive and forgetful, whose intensity can be adjusted. The goal is to practice steering a line of questioning despite difficult behavior.

Ma recalls an attorney who struggled in a real deposition because opposing counsel’s distracting laughter repeatedly broke their concentration. She wants simulations to rehearse not just ideal procedure but situations lawyers should be prepared to handle or avoid.

A separate research direction explores personalized legal personas: models that imitate an individual partner rather than a generic lawyer. Ma says models readily role-play but can slide into caricature; the challenge is to represent how a person’s behavior varies by context and degree.

Ma gives the example of an aggressive negotiator whose behavior changes with context. She asks whether numerical controls could preserve that nuance instead of producing an exaggerated stereotype.

Markus Covert introduces his bioengineering work by describing complex disease as the product of many genetic, epigenetic and environmental factors. He argues that understanding disease requires grappling with those interacting causes rather than treating it as a single-gene problem.

His lab has built a computational model of a bacterial cell that takes its gene content into account and predicts observable behavior. A simulation can show growth, DNA replication and division; he says one cell cycle generates about a gigabyte of data.

When the modeled environment changes, sensors, gene expression and metabolism respond together, altering growth and replication. Covert calls this work closer to artificial life than AI, while also studying how the two can interact.

His group began using deep learning more than a decade ago to identify and track individual cells in microscopy images. He recalls work around Google X and specialized hardware as a reminder that future computation can make previously inaccessible problems tractable.

Covert recalls conversations around the hardware behind AlphaGo as a formative example of a large change in computational capability. He says such advances can change what researchers consider possible.

Covert asks what ambitious human goals become conceivable when technologies such as AI, induced pluripotent stem cells or CRISPR emerge. The important question for him is not merely which small tasks improve, but which formerly infeasible problems can be approached.

Covert and teaching assistants built an AI helper using material from his own class and textbook. He says that, after reading an MIT report he understood to associate such tools with worse learning and well-being outcomes, the group voted not to release it. He frames the decision as avoiding possible harm, not a finding from their own trial.

He asks how new tools can support human flourishing, drawing on conversations about AI and faith traditions. He is interested in what different traditions consider worth protecting as technology changes.

He says religious and secular traditions may offer different perspectives on how to protect human beings and encourage flourishing. He wants those questions to be part of technology development.

Covert also returns to a concern that productivity tools can create more output while leaving less time to think. He asks how AI might instead give scientists time to reason, doctors time to care, students time to explore and people time to connect.

Bailey asks how researchers should balance present AI limitations against uncertain capabilities five years from now. The question matters when a tool or research program takes years to build.

Covert recalls being told at Google X to assume abundant computation and memory rather than design only around current constraints. He says that invitation helped him think bigger, while recognizing that the enabling conditions might arrive later.

Algee-Hewitt is skeptical that anyone can reliably forecast model capabilities five years ahead. He instead centers the human experience of interacting with models: people and their needs may change more slowly than the technology.

Algee-Hewitt says human interaction with models provides a more stable subject of inquiry than a specific prediction about the technology. Researchers can study those interactions even as the underlying systems change.

Ma says legal technology has often sought a standardized ideal document, while the profession depends heavily on individual judgment and relationships. The gap between tacit legal practice and how tools interface with lawyers is as much cultural as technical.

Ma describes the lab’s Socratic Mentor, which repeatedly questions a user about why they made a decision and helps break down research problems. In trials with senior attorneys, she says the probing helped them articulate how experience shaped their judgments.

Rather than focus only on AI executing legal work, Ma sees research value in helping attorneys understand and explain why they work as they do. Improving models can make that back-and-forth more productive, she says.

Bailey asks how people should handle mismatched expectations about responsible AI use: for example, a student may view limited AI editing as acceptable while an instructor does not. Similar gaps can arise wherever others expect a personal contribution.

Bailey gives an example from faith communities: leaders must consider whether a task is seen as administrative or as requiring a personal or spiritually meaningful contribution. Discovering a mismatch between expectations and practice can undermine trust.

Covert polls the class about AI-written recommendation letters. He says he has decided to write his own letters, though he may use AI to store and retrieve information such as a student’s internship history.

His aim is to improve the quality of a recommendation, not merely save time: the letter should convey his personal knowledge of the student and their relationship. Ma asks the class to compare this with a professor asking a student to draft the letter for a signature.

Covert asks whether the class would favor a professor writing with AI support or asking the student to draft their own letter. The question illustrates how actual alternatives complicate a simple prohibition.

Ma says some junior attorneys are reluctant to use AI because of the risk of publicized errors such as invented citations. She recounts discussing one incident in which, she says, seven attorneys had seen the brief before the problem escaped notice.

She says some law-firm partners document how they use AI to guide junior colleagues. Because models and work change quickly, Ma argues that experienced users and educators should explain responsible practice rather than expect individuals to rely only on their own intuition.

Ma asks whether educators should keep reusing the same exams and materials or rethink how they teach as tools change. She sees it as a responsibility of experienced practitioners to communicate their own boundaries and methods.

Algee-Hewitt reflects on the Turing test as a historically influential test based on a machine being mistaken for a person. His response to uncertainty about acceptable use is transparency: disclose how and where a tool was used so others can assess it and discuss expectations.

Bailey asks about practical disclosure norms in journals, conferences and public media. Algee-Hewitt argues that blanket bans can encourage people to use AI without admitting it, even though some uses are genuinely problematic.

Bailey asks what infrastructure could support disclosure, from research submissions to AI-generated political advertisements. She invites the panelists to consider community norms as well as formal labels.

In computational literary studies, he says, the writing itself is part of the human thought process, while researchers may use AI to help code. Some journals distinguish these uses: do not have AI write the paper, but disclose and credit its assistance with software.

Algee-Hewitt says his field values writing as a process of thought and a human-created object. Coding can be different for humanities researchers who need computational methods but did not train as computer scientists.

Ma describes LIFT Lab’s internal policy: researchers may use current tools, but must explain why they used them and the important design choices made along the way. She says projects that looked complete sometimes fell apart under questions about the research impact or architecture.

Ma says a completed output is not enough if a researcher cannot explain the choices behind it. Her lab expects people to understand the architecture and the effect of their decisions on the research.

Covert worries students could use AI to complete familiar coursework without doing the thinking, then face unfamiliar real-world research problems for which the tools are less reliable. He says students need both competence with AI and the underlying education they came to acquire.

In his lab, Covert asks researchers who save time with AI to give some of it back to checking their own and colleagues’ work. He offers an illustrative allocation of 20% for checking and another 20% for deep thought about important problems.

Covert says prolific generation can overwhelm readers and the publishing system. For him, transparency and quality require deliberate review rather than assuming that more pages mean more useful work.

He also wants researchers to spend time talking with other curious people. He recounts a museum discussion about intellectual revolutions and argues that social exchange can help new ideas emerge; to him, AI is valuable if it creates space for such thought and connection.

An audience member asks which abilities students should protect from erosion as AI becomes more capable. Algee-Hewitt names critical thinking: the ability to interpret using one’s own ideas and experience rather than outsourcing that faculty before it is developed.

Ma names judgment and the accumulation of a personal reference library of experiences. She encourages students to live beyond a race through school, so they can later draw on varied experiences to build analogies and tell meaningful stories.

Covert points to scientific taste: deciding which problems deserve attention. He also emphasizes intellectual courage, familiarity with more than one field, teamwork and sustained curiosity, because facts and tools can change rapidly.

Covert encourages students to combine knowledge across fields and tackle problems that appear unsolvable. He sees teamwork and habits of curiosity as more durable than any single fact learned today.

A student asks where the boundary lies between using AI to learn a new skill and simply asking it to complete the task. Covert compares letting AI do all one’s coursework to watching a robot lift weights: the learner misses the exercise that builds their own capacity.

Covert describes a writing experiment in which he drafted each paragraph himself and asked Claude for missing citations, the strongest counterarguments and analogies from other fields. He says he checked the suggested papers rather than trusting the model to identify them correctly.

For each sentence Covert asks for three potentially missing references, three arguments that could undermine it and analogies from other fields. He keeps a running list and checks the papers himself.

After two days he had four paragraphs that he considers among his best. He says he could have written them faster from his existing knowledge, but the dialogue led him to read and think more deeply without giving up his own voice.

Algee-Hewitt says the boundary depends on the purpose of the activity. In a proposed critical-gaming class, students would design and play games to examine narrative and gameplay, while AI could help implement simple code that the class is not intended to teach.

The game-design example reverses in a programming class, Algee-Hewitt says: coding might be the central skill instead. The relevant boundary follows from what the activity is actually intended to develop.

Ma says AI can help people acquire the basics of a neighboring field, but curiosity should not end with the first answer. She suggests using that starting point to speak with colleagues who have deeper expertise.

A final student asks which design choices in educational AI could avoid the negative outcomes Covert mentioned, and describes a course using an AGENTS.md file to place guardrails on AI assistance.

Covert does not identify a specific feature responsible for those outcomes. He emphasizes the chance to learn from researchers experimenting at the frontier, compares the current moment to earlier changes in web search, and encourages students to take classes where the field is actively being developed.