AI safety and alignment, from the humanities
Humane Studies for AI is an independent, humanities-led initiative on AI safety, alignment and human flourishing.
Descend00 — Humane Studies
The oldest human questions, asked of the newest intelligences.
Humane Studies is the inquiry into what it means to be human and how to live well. It asks what people are, what human thriving requires, what makes a society good, what we owe one another, and how knowledge and power should be used. It is a tradition of questions before it is a body of answers, and each generation has to ask them again.
It is older than the modern departmental humanities. Its lineage runs back to the Renaissance studia humanitatis, a course of grammar, rhetoric, poetry, history and moral philosophy meant to form character as well as skill. Long before universities divided knowledge into departments, people put questions that crossed the arts, the sciences and the study of society, and no single department has ever owned them.
The humanities are part of Humane Studies, not the whole of it. Literature, history, philosophy and the arts bring close reading, interpretation and the habit of arguing about ends. The natural and social sciences bring measurement, models and evidence about how people think and behave. Each brings something the others lack. Humane Studies is therefore not the humanities under another name, and it uses whatever methods the question requires. A question about people and machines usually requires several at once.
That matters now because machines produce language, have a sophisticated theory of mind, tell stories, influence decisions, engage us emotionally, preserve and distort memory, and take part in institutions. As a result, the oldest questions about people have become questions about what we build and what it does to us. As we shape our tools, how much do we let our tools shape us? How can we use artificial intelligence as a lens and a mirror to know ourselves better and to create more humane technologies?
Humane Studies for AI is the independent initiative that brings this inquiry to AI safety and alignment, co-led by Katherine Elkins and Jon Chun. It begins with the questions and asks the technical work to answer to them, rather than the other way around. It publishes its methods, its data and its disagreements.
01 — The Big Questions
These are the questions Humane Studies for AI starts from. Three lead the list, and two more are always asked with them.
Systems now write in the first person, console and advise. They understand us better than we understand ourselves, and can readily see and act on human emotions, desires, cognitive biases and limitations. What do they understand of the people they address, and what do they miss?
A system that advises on work, love or health answers from some picture of how to live, chosen or not. Whose idea of a good life do its answers carry? How do we, individually and collectively, learn to think critically, engage constructively, and ultimately flourish in a world of rapidly advancing, widely disseminated and capable artificial intelligence on the road to AGI/ASI?
Machines now shape the political information people read, and they take part in courts, schools, clinics and councils. What happens to a society when systems do that work, and who decides what they do? Who curates the training data and alignment behaviors that instill these values, reasoning traces and goals?
Two more questions sit beneath the three
Obligation did not begin with machines, but machines now sit between people who owe each other honesty and care. If AI becomes the ultimate disrupter, how can we draw on our best academic traditions to rethink our fundamental assumptions and the interconnected frameworks of everything from economics to ecclesiology?
Those who build, deploy and govern these systems hold a great deal of both, and the question applies to each of them. How do we govern both AI and ourselves, from Juvenal’s Quis custodiet ipsos custodes? (Who will guard the guards themselves?) to Federalist No. 51?
A humane-studies program founded at Kenyon College in 1975, the one the co-leads have taught in, began from three questions of its own: “What is our nature and what do we need to flourish? How has history shaped us? Given those conditions, what can we become and what ought we create?”
Read beside the questions above, that triad asks the same things in another order: what we are, how we came to be this way, and what we ought to build next.
Questions come before the code, because a system can be judged only against a standard, and a standard is a view about people that someone has to state and defend.
02 — Humanity & AI
Machines now produce language, have a sophisticated theory of mind, tell stories, influence decisions, engage us emotionally, preserve and distort memory, and take part in institutions. What do these systems understand, and what do they miss?
In 1959 C. P. Snow warned of two cultures, the literary and the scientific, that had stopped talking to each other. His shortage was on one side: humanists who could not do science. Generative AI turns that shortage around.
AI is rapidly eliminating technical barriers, automating routine knowledge workflows, and increasingly handling complex analysis. Three measurements coincide with that change.
Students are moving to domain fields such as health and engineering, not to the humanities. AI is shifting the value of human labor from narrow specialization toward critical thinking and review, creative integration across domains, and collaboration with others: the hallmarks of a liberal arts education. That AI is shifting value toward these capacities, and that they live in the liberal arts and in Humane Studies, is this initiative’s claim; the figures record a decline in computer-science enrollment and entry-level technical work, not where that value goes.
Knowing what to build, and what an output means.
Judgment about ends and meaning lives in domain expertise and in Humane Studies.
The co-leads’ own record
The curriculum the co-leads created at Kenyon College in 2016 raised female enrollment from 18 percent in 2017 to 61 percent, drew more than 90 percent of its students from non-STEM majors, and brought in students from every division, from winners of the mathematics department’s prize to holders of creative-writing scholarships. More than 400 original interdisciplinary machine-learning and AI research projects have been mentored there, and, as of 2026, student research from that curriculum has recorded more than 130,000 downloads in 198 countries (Digital Kenyon, by Bepress analytics), nearly every country in the world.
That is a small case of the argument above: given questions worth answering, students from outside STEM did technical work.
In a 2026 paper, Klowden and Tao cite Chun and Elkins (2023) for the point that debating the values we want these tools aligned with is the first step.
The measurement literature on evaluating AI, which asks the same questions in other terms, is set out in The Approach.
03 — The Approach
Technical and sociotechnical work on AI safety and alignment is well established. Humane Studies starts from the Big Questions instead, and each approach sees something the others do not.
| Technical alignment | The sociotechnical approach | The Humane Studies approach | |
|---|---|---|---|
| Starts from | Capabilities and control: what can the system do, and can we stop it? | The system inside institutions, markets and power. | The questions: what is it for, whose idea of a good life does it carry, and who gets to say? |
| Sees well | Benchmarks, dangerous capabilities and robustness. | Deployment harms, incentives and who bears the cost. Researchers in this tradition argue that evaluating generative AI is a social-science measurement problem, and that it must cover human interaction and systemic impact as well as capability: Wallach et al., 2024 · Weidinger et al., 2023. | A system’s language as evidence of the traditions and assumptions it inherits. |
| Less often draws on | Questions about whose ends the system serves. | The long record of reflection on the good life that it is implicitly judging. | Scale and engineering detail, unless paired with the others. |
| Typical methods | Evaluations, interpretability and red-teaming. | Auditing, fieldwork, and policy and impact analysis. | Interpretation and comparative reading across traditions, integrated with computational methods from the start. |
The aim that AI should serve people, often called human-centered, is one all three approaches can share, and Humane Studies is the tradition and method for working out what that aim requires.
The method is reading machines as texts. The initiative reads a system’s language closely, the way a humanities seminar reads a text, and then questions it and argues about it with evidence.
These are methods, plural. Interpretation and comparative reading sit beside statistics and computation, joined from the start, so that each shapes the questions asked and the conclusions that can responsibly be drawn.
03 — First work
Three workstreams, each taking one failure of reading and judgment and planning one public release.
Persuasion and manipulation. Close reading of a model’s tone, framing and omissions. First release: we will publish an open, documented evaluation set, with its labeling procedure and inter-rater agreement stated.
Sycophancy and flattery. Studies of how a system’s language shapes the people who rely on it. First release: we will publish a protocol with annotated examples, so that other groups can repeat the studies.
Cultural and value skew. Whose language and assumptions a model carries, and where its test materials came from. First release: we will publish a multilingual evaluation note and a dataset-provenance register.
The questions are pressing now. Four findings by other researchers, each paraphrased and linked to its source so that you can check it, show why.
03 — Scope
The findings above are other researchers’ work. Our own commitments are set out under How we work.
We do not evaluate dangerous capabilities, such as cyber, biological or chemical risk.
We do not report causal or population-level findings on persuasion without a proper human-subjects design and review.
We speak for no funder or agency.
Our work is complementary evidence and method, alongside the evaluations that AI developers and governments run.
04 — People & Record
Humane Studies for AI is led by Katherine Elkins and Jon Chun, whose work joins literature and the humanities to computer science, model evaluation, and AI standards.
Co-Lead
AI safety researcher and Professor of Comparative Literature and Humanities at Kenyon College. Principal Investigator for the Modern Language Association team at NIST CAISI and of Archival Intelligence through Schmidt Sciences HAVI.
Author of The Shapes of Stories (Cambridge UP, 2022) and Proust’s In Search of Lost Time: Philosophical Perspectives (OUP, 2022). Ph.D., UC Berkeley.
Co-Lead
AI research scientist and Visiting Instructor of Humanities at Kenyon College; co-creator, with Katherine Elkins, of a human-centered AI curriculum (2016). Created SentimentArcs, an open-source toolkit for diachronic sentiment analysis; ICML 2024 oral presentation (top 2%). Co-PI for the MLA team at NIST CAISI and of Archival Intelligence.
Co-founded SafeWeb ($26M acquisition by Symantec; first In-Q-Tel security investment). UC Berkeley EECS, UT Austin MS. Two US patents.
Independence
Humane Studies for AI is an independent initiative. It is not affiliated with, operated by, or endorsed by Kenyon College or any other institution, and the views expressed here are the co-leads’ own. It decides for itself what it studies and what it publishes, and it speaks for no university, funder, or agency. Its co-leads take no part in decisions about their own support.
04 — Track record
These are the co-leads’ own publications, roles and outcomes, as of October 2026.
Complete, up-to-date lists live on their scholarly profiles: Katherine Elkins (Google Scholar) and Jon Chun.
Elkins, K., & Chun, J. “Can GPT-3 Pass a Writer’s Turing Test?” Journal of Cultural Analytics. It has 370 or more citations.
Elkins, K. The Shapes of Stories: Sentiment Analysis for Narrative. Cambridge University Press, Elements in Digital Literary Studies. Elkins, K. Proust’s In Search of Lost Time: Philosophical Perspectives. Oxford University Press.
Chun, J., & Elkins, K. “The Crisis of Artificial Intelligence: A New Digital Humanities Curriculum for Human-Centred AI.” International Journal of Humanities and Arts Computing 17(2). DOI: 10.3366/ijhac.2023.0310. It has about 80 citations.
Elkins, K. “A(I) University in Ruins: What Remains in a World with Large Language Models?” PMLA 139(3).
Eiras, F., et al., with Elkins, K., and Chun, J. “Position: Near to Mid-term Risks and Opportunities of Open-Source Generative AI.” ICML 2024, oral presentation; 24 coauthors.
Klowden, T., & Tao, T. “Mathematical Methods and Human Thought in the Age of AI.” arXiv:2603.26524, citing Chun & Elkins (2023): “Debating the system of values we wish these tools to align with is the first step.”
The co-leads have led the Modern Language Association team in NIST’s AI consortium since 2024, according to the co-leads’ own project reports.
The co-leads were investigators, with a Notre Dame colleague, on a project auditing large language models for fairness and explainability, funded in 2024 through the Notre Dame–IBM Technology Ethics Lab.
Katherine Elkins is principal investigator and Jon Chun co-director of a project selected in 2025 for funding through the Schmidt Sciences Humanities and AI Virtual Institute (HAVI).
The co-leads created a human-centered AI curriculum at Kenyon College in 2016, to the best of our knowledge the first led from the humanities. The enrollment shift, the mentored projects and the downloads of student research that curriculum produced are set out under Humanity & AI.
As of 2026, at least nine university courses in seven countries assign the co-leads’ work, none of them solicited.
05 — Collaboration
Humane Studies for AI is an initiative that convenes fields and sectors around questions that none of them can answer alone.
For academic readers
Universities divide knowledge by department and reward work within it. That structure builds depth, but the questions AI raises do not respect it. Socrates began as a natural scientist before turning to justice and the good, and philosophy and science were one inquiry long before they were separate departments. AI forces them back together: every technical choice carries an assumption about what is good that the technical fields cannot justify by themselves. For fifty years a humane-studies program founded in 1975 at Kenyon College — the one the co-leads have taught in — has shown, inside one college, that this integrated inquiry can be taught. The initiative extends that inquiry across institutions, as a vehicle that convenes departments without belonging to any one of them, and returns to each what none can produce alone.
For readers in industry, government and nonprofits
AI systems now shape what people see, feel and decide, and the organizations that build them largely control who may examine them. Outside evaluation still depends on voluntary cooperation, and research access can be withdrawn by a change in platform policy. Meanwhile the hardest problems, such as emotional manipulation, fragmented public discourse and whose idea of the good a system carries, sit outside any one organization’s expertise. They need a convener that owes its shape to no single company, agency or discipline. It must be independent to be credible to all of them, and collaborative because no one field or sector can see the whole.
05 — Instruments
Two instruments are current. Four are planned, each with a quarter in flexible wording until it exists.
Joint evaluation. A research group, a developer or an agency runs an evaluation with the co-leads, so that close reading of a system’s language sits beside the group’s own tests. Propose one
Convening. Symposia, panels and talks that bring fields and sectors into one room. The dated calendar is below, under where we are convening
The open seminar is planned for the first quarter of 2027: a free seminar, open to faculty, students and practitioners from any field. Follow it
Research affiliates are planned for the first quarter of 2027: colleagues elsewhere who take part in a shared evaluation or study with the co-leads. Ask about affiliation
Fellows are expected in the second quarter of 2027, with a first call for applications. Ask to be told
Open protocols and data come with the first release of each workstream: the labeling procedure, the agreement figures and the data are published so that other groups can repeat the work. See the first work
05 — Sectors
What working together looks like in each.
Departments, provosts and research groups bring their field’s methods to a shared question, and take what is found back into their own teaching and research.
Developers and deployers open a system to joint evaluation, and read what close reading finds alongside the tests they already run.
Agencies and standards processes bring the Humane Studies approach into an evaluation, and test it against the decisions they have to make.
Funders and nonprofits help to pay for open evaluation work and bring knowledge of the communities that AI systems affect, and no funder shapes what is published.
Working across fields has a record, set out here as the co-leads’ own résumé facts.
Where the co-leads have presented or worked
As of 2026, these are places where Katherine Elkins and Jon Chun have given talks or carried out work. No listing implies endorsement, partnership or funding of this initiative.
06 — FAQ
Common questions about Humane Studies, the initiative that applies it to AI, and how to work with it.
Humane Studies is the inquiry into what it means to be human and how to live well. It is older than the modern departmental humanities and uses whatever methods the question requires, and Humane Studies for AI applies it to AI safety and alignment. The definition is under Humane Studies.
Safety labs mostly work on the technical approach, which starts from capabilities and control, and think tanks mostly work on policy and institutions, which is closer to the sociotechnical approach. Humane Studies starts instead from the Big Questions and reads a system’s language as evidence, and it is meant to work beside the other two. The comparison is set out under The Approach.
The questions AI raises cross every department and sector, so they need an independent convener that belongs to none of them. The case is made in Why an initiative, why now.
The Lab is a separate organization that the co-leads also founded, and it holds its own projects, which you can find on its own site.
The co-leads hold university posts and built their record there, which is set out as their own history under People & Record. Humane Studies for AI is independent and is not run by any university.
Joint evaluations and convenings are open now, and the open seminar, research affiliates and fellows are planned for 2027. The collaboration instruments are listed under Collaboration, and each sector has one ask under Get Involved.
Support for the open evaluation work is welcome, including research credits, and no funder shapes what is published. Humane Studies for AI speaks for no funder or agency. To talk about support, see Get Involved.
07 — Get Involved
One ask for each sector and audience. Each one starts with a conversation with the co-leads, and the ways of working together are set out under Collaboration.
