Schedule of Readings & Work

General Info

• Readings listed under each week are what you should read BEFORE SEMINAR
• Readings are listed in the order that I think works best
• ALL readings are available online or in the course Zotero library. See the syllabus home page for directions.

Part I: Expertise and Trust

Before looking at any technology, we need a vocabulary. Who counts as an expert, why do the rest of us defer to them, and what is trust anyway? Philosophers, sociologists, and historians of science have been arguing about this for decades, mostly without thinking about AI at all, which is exactly why they’re useful.

1: Unprecedented?

Almost everything written about AI calls it unprecedented. That word does a lot of work: if nothing like this has happened before, then nobody can be expected to know anything and every opinion is as good as any other. This week we start from the opposite assumption, that AI belongs to a long line of technologies that changed how people know things, and ask what that line looks like.

• Melvin Kranzberg, “Technology and History: ‘Kranzberg’s Laws’”, Technology and Culture 27.3 (1986): 544–560. Six short laws from a founder of the history of technology. The first one, “technology is neither good nor bad; nor is it neutral,” is the course’s starting position.
• Henry Farrell, Alison Gopnik, Cosma Shalizi, and James Evans, “Large AI Models Are Cultural and Social Technologies”, Science 387 (2025). Short, and the best single argument for treating AI as kin to writing, print, markets, and bureaucracies rather than as a mind.
• Fred Gibbs, Expertise and Trust in Humanities AI (2026). Slides from a fifteen-minute talk that sketches the argument of this course. Click through quickly.

Discussion

• What does calling something “unprecedented” let us avoid?
• What would count as a good precedent for AI? What would be a misleading one?
• Where in your own research do you already depend on something you can’t check?

2: Depending on Others

Every scholar depends on testimony: the archive catalog, the translation, the colleague who read the sources you never will. Philosophers call this epistemic dependence. This week introduces the four problems that run through the course: definition, identification, deference, and transfer.

• John Hardwig, “Epistemic Dependence”, Journal of Philosophy 82.7 (1985): 335–349. Short and foundational. Rational belief often means deferring to experts, and that isn’t a failure of rationality.
• Alvin Goldman, “Experts: Which Ones Should You Trust?”, Philosophy and Phenomenological Research 63.1 (2001): 85–110. The identification problem in its classic form: how can a novice choose between experts who disagree?
• Miranda Fricker, Epistemic Injustice: Power and the Ethics of Knowing (2007), ch. 1, “Testimonial Injustice”. Credibility is unevenly distributed, and not by competence.

Discussion

• How do you decide which historians to trust outside your own field? What cues do you actually use?
• How many of those cues are about competence?
• Is depending on an AI system a case of epistemic dependence, or something else?

3: Experts and Professions

Expertise is not just something a person has. It’s also something a profession claims and defends. This week looks at expertise as a social position, and at what happens when a new tool lets outsiders do (or appear to do) the experts’ work.

• Harry Collins and Robert Evans, “The Third Wave of Science Studies: Studies of Expertise and Experience”, Social Studies of Science 32.2 (2002): 235–296. Long; read the opening and the sections distinguishing contributory expertise (being able to do it) from interactional expertise (being able to talk about it fluently).
• Andrew Abbott, The System of Professions: An Essay on the Division of Expert Labor (1988), selections. Professions compete for jurisdiction over tasks, and new technologies reopen those contests.
• Gil Eyal, The Crisis of Expertise (2019), introduction.

Discussion

• Which kind of expertise does a language model have, contributory or interactional? Which kind does a historian writing outside their field have?
• What jurisdictions might historians lose to AI? Which might they gain?
• Is there a “crisis of expertise,” or is that a story experts tell?

4: What Trust Is

We use “trust” for people, institutions, and things alike: we trust a colleague, a journal, a rope. Are these the same attitude? This week separates trust from mere reliance, and asks whether transparency actually helps.

• Annette Baier, “Trust and Antitrust”, Ethics 96.2 (1986): 231–260.
• Onora O’Neill, A Question of Trust (BBC Reith Lectures, 2002), lectures 1 (“Spreading Suspicion”) and 4 (“Trust and Transparency”).
• C. Thi Nguyen, “Trust as an Unquestioning Attitude”, Oxford Studies in Epistemology 7 (2022).

Discussion

• Can you trust a thing, or only rely on it? Does the answer change for a thing that talks?
• O’Neill argues that transparency can make trust harder. How?
• What would it mean to trust a search bar the way Nguyen describes?

Part II: Precedents

Trust does not arrive with a technology. In each of these cases, a new way of making knowledge outran the existing ways of checking it, and people had to build trust deliberately, over decades, through institutions that didn’t yet exist. The point is not that AI is “just like” any of them. It’s that each one gives us better questions to ask.

5: Trusting Writing

The oldest complaint about AI is older than AI: Plato made it about writing. Medieval England shows how long it took for written records to be trusted at all.

• Plato, Phaedrus 274c–277a (the story of Theuth and Thamus). A few pages. Writing, Thamus warns, will give people the appearance of wisdom without the reality.
• M. T. Clanchy, From Memory to Written Record: England 1066–1307, chapter “Trusting Writing”. Charters didn’t earn trust on their own; people trusted witnesses, seals, and symbolic objects before they trusted the parchment.
• Walter Ong, Orality and Literacy (1982), ch. 4, “Writing Restructures Consciousness”, selections.

Discussion

• Plato’s complaint about writing is the most common complaint about AI. What’s right about it?
• What made a charter trustworthy, and how long did that take?
• What did writing actually do to expertise: concentrate it, spread it, or change what it was?

6: Print and Fixity

Did print make texts trustworthy, or did people have to make print trustworthy? Historians fought about this for decades, and the fight maps closely onto current arguments about AI.

• “AHR Forum: How Revolutionary Was the Print Revolution?”, American Historical Review 107.1 (2002): 84–128. Anthony Grafton’s introduction, Elizabeth Eisenstein’s and Adrian Johns’s essays, and Eisenstein’s reply.
• Ann Blair, Too Much to Know: Managing Scholarly Information before the Modern Age (2010), introduction. Information overload is old, and so are the tools scholars built to cope with it: compilations, indexes, and abridgments that let readers seem to have read books they hadn’t.

Discussion

• Johns argues that fixity is not a property of print but something people had to manufacture. Is accuracy a property of AI, or something people will have to manufacture?
• Is a language model a reference book? What would Blair’s compilers make of it?
• Galileo’s pirated woodcuts were reprinted for centuries, and the bad copy outcompeted the good one. What’s the AI equivalent?

7: Witnesses and Gentlemen

An experiment happens once, in one room, and everyone else has to take someone’s word for it. Seventeenth-century natural philosophers solved this problem socially rather than technically.

• Steven Shapin and Simon Schaffer, Leviathan and the Air-Pump: Hobbes, Boyle, and the Experimental Life (1985), ch. 2, “Seeing and Believing: The Experimental Production of Pneumatic Facts”.
• Steven Shapin, “The Invisible Technician”, American Scientist 77.6 (1989): 554–563. Who did the work, and who got to vouch for it.

Discussion

• Boyle’s solution to the trust problem was exclusionary, and it worked. What does that tell us?
• Who are AI’s witnesses? Who are its invisible technicians?
• Is a published benchmark score a kind of witnessing?

8: Footnotes and Referees

Historians have their own technologies of trust, and we rarely think of them as technologies. This week looks at two of them: the footnote, and peer review.

• Anthony Grafton, The Footnote: A Curious History (1997), ch. 1.
• Melinda Baldwin, “Scientific Autonomy, Public Accountability, and the Rise of ‘Peer Review’ in the Cold War United States”, Isis 109.3 (2018): 538–558. Peer review only became the badge of scholarly legitimacy in the late twentieth century.

Discussion

• A footnote is a promise that a reader can check your work. What happens to that promise when a machine can generate plausible footnotes?
• What does peer review actually certify? What did people trust before it?
• What should a footnote to an AI-assisted reading look like?

9: Trust in Numbers

In the nineteenth century, scientists and bureaucrats began to prefer procedures over persons: mechanical images over drawings, numbers over judgment. This “mechanical objectivity” promised knowledge without a knower.

• Lorraine Daston and Peter Galison, “The Image of Objectivity”, Representations 40 (1992): 81–128.
• Theodore Porter, Trust in Numbers: The Pursuit of Objectivity in Science and Public Life (1995), introduction. Quantification flourishes where experts are distrusted; numbers are a way of not having to trust a person.

Discussion

• Mechanical objectivity promised judgment without a judge. Is that what people want from AI?
• What’s lost when you trust the procedure instead of the person? What’s gained?
• When does a number become evidence? What about a confidence score?

10: When Computers Were People

Before computers were machines, they were people, often women, doing calculations by hand in carefully divided workflows. The history of mechanized calculation is a history of deskilling some workers and upskilling others.

• Lorraine Daston, “Calculation and the Division of Labor, 1750–1950”, Bulletin of the German Historical Institute 62 (2018): 9–30.
• Jennifer S. Light, “When Computers Were Women”, Technology and Culture 40.3 (1999): 455–483.
• Lisanne Bainbridge, “Ironies of Automation”, Automatica 19.6 (1983): 775–779. Five pages from engineering psychology. Automation leaves humans the hardest tasks, with the least practice at them.

Discussion

• Who was deskilled by mechanical calculation, and who was upskilled?
• Bainbridge’s irony is that automation makes human expertise more important, not less. Does that hold for AI-assisted research?
• Whose labor becomes invisible when work is divided this way?

Part III: The Machine as Expert

With a vocabulary and a set of precedents in hand, we turn to AI itself. No technical background is assumed. The goal is to understand enough about how these systems work to ask good historical and epistemological questions about them, and to test them against expertise you actually have.

11: Search, Archives, Silences

Historians have already lived through one disruption of this kind: keyword search changed how history is researched, and mostly nobody said so in the footnotes. This week connects that experience to older questions about archives and to newer ones about training data.

• Lara Putnam, “The Transnational and the Text-Searchable: Digitized Sources and the Shadows They Cast”, American Historical Review 121.2 (2016): 377–402.
• Michel-Rolph Trouillot, Silencing the Past: Power and the Production of History (1995), ch. 1, “The Power in the Story”.
• Kate Crawford and Trevor Paglen, “Excavating AI: The Politics of Images in Machine Learning Training Sets” (2019).

Discussion

• Trouillot identifies moments when silences enter history: the making of sources, archives, narratives, and history itself. Where does a training dataset fit?
• How did keyword search change your own research? Did you ever cite it?
• Whose knowledge is compressed into a language model, and whose is left out?

12: Encoding Expertise

In the 1980s, AI researchers tried to build “expert systems” by interviewing experts and writing down their knowledge as rules. It went badly, and the reasons why are the transfer problem in its sharpest form.

• Michael Polanyi, The Tacit Dimension (1966), ch. 1, “Tacit Knowing”. “We can know more than we can tell.”
• Diana E. Forsythe, “Engineering Knowledge: The Construction of Knowledge in Artificial Intelligence”, Social Studies of Science 23.3 (1993): 445–477. An anthropologist among the knowledge engineers: what they thought knowledge was, and what got lost when they wrote it down.

Discussion

• What part of your own expertise is tacit? Could you write it down if you had to?
• Forsythe’s engineers deleted what they couldn’t formalize. What do language models delete?
• Is writing down your method a loss of expertise, or a way of making it checkable?

13: What Language Models Do

What is a language model actually doing when it answers a question? These readings offer competing metaphors, all of them readable without technical background.

• Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?”, Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (2021): 610–623.
• Ted Chiang, “ChatGPT Is a Blurry JPEG of the Web”, New Yorker (February 9, 2023).
• Michael Townsen Hicks, James Humphries, and Joe Slater, “ChatGPT Is Bullshit”, Ethics and Information Technology 26 (2024).
• OPTIONAL: Murray Shanahan, “Talking About Large Language Models”, Communications of the ACM 67.2 (2024).

Discussion

• Parrot, blurry JPEG, bullshitter: which metaphor best fits what you saw in seminar?
• Hicks and colleagues argue that “hallucination” is the wrong word. What’s at stake in the name?
• Where was the system reliable? Could you have told the difference without already knowing the answer?

14: Performing Authority

In 1966 Joseph Weizenbaum built ELIZA, a simple program that imitated a therapist, and was alarmed at how readily people confided in it. Today’s chat interfaces are far more capable, and they’re designed to be trusted.

• Joseph Weizenbaum, “ELIZA—A Computer Program for the Study of Natural Language Communication Between Man and Machine”, Communications of the ACM 9.1 (1966): 36–45. Skim the technical middle; read the opening and the closing.
• Joseph Weizenbaum, Computer Power and Human Reason: From Judgment to Calculation (1976), introduction.

Discussion

• Which of the trust technologies we’ve studied (witnesses, seals, footnotes, numbers) do chat interfaces imitate? Which do they leave out?
• Weizenbaum’s subtitle is “From Judgment to Calculation.” Was he right about the direction?
• What would an interface that earned trust, rather than performed it, look like?

15: Expertise after AI

The last week turns the course’s question back on scholars. If AI is already reliable within tasks a skilled user has scoped, the open question is no longer whether to trust the machine. It’s how anyone knows when to trust the people who use it.

• Inkeri Koskinen, “We Have No Satisfactory Social Epistemology of AI-Based Science”, Social Epistemology (2023). Trust requires accountable agents; AI applications aren’t agents. So what are scientists doing when they rely on them?
• Mason Majszak and Lorenzo Sartori, “AI, Experts, and Epistemic Authority” (2026, preprint). AI may be a new category that is both scientific instrument and epistemic authority.
• Thomas Metcalf, “Is AI-Produced Humanities Scholarship a Case of Research Misconduct?”, Journal of Academic Ethics (2026). Argues that publishing AI-generated scholarship under your own name (“persona authorship”) doesn’t clearly violate research norms, at least not more than traditional authorship often does. A good argument to disagree with.
• Skim the professional guidance: American Historical Association, Guiding Principles for Artificial Intelligence in History Education (2025), and the MLA’s AI and the Humanities: A Framework for Language and Literary Scholarship (2026).

Discussion

• I have history credentials. What would AI credentials look like?
• Is Metcalf right? If not, what norm does persona authorship actually violate?
• What should disclosure of AI use say, so that a reader can decide how far to trust the work?

Finals Week: Research Presentations

Present your final project to the seminar.

Other Possible Weeks

Topics and readings that could replace or extend any of the weeks above.

• Expert witnesses in court: Tal Golan, Laws of Men and Laws of Nature: The History of Scientific Expert Testimony in England and America (2004).
• Lay expertise: Steven Epstein, “The Construction of Lay Expertise: AIDS Activism and the Forging of Credibility in the Reform of Clinical Trials”, Science, Technology, & Human Values 20.4 (1995): 408–437.
• Wikipedia and open authority: Roy Rosenzweig, “Can History Be Open Source? Wikipedia and the Future of the Past”, Journal of American History 93.1 (2006): 117–146.
• New media and their experts: Carolyn Marvin, When Old Technologies Were New: Thinking about Electric Communication in the Late Nineteenth Century (1988).
• Technology in use, not invention: David Edgerton, The Shock of the Old: Technology and Global History since 1900 (2006).
• Do artifacts have politics? Langdon Winner, “Do Artifacts Have Politics?”, Daedalus 109.1 (1980): 121–136.
• Echo chambers: C. Thi Nguyen, “Echo Chambers and Epistemic Bubbles”, Episteme 17.2 (2020): 141–161.
• The labor behind AI: Mary L. Gray and Siddharth Suri, Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass (2019).
• Data and extraction: Kate Crawford, Atlas of AI (2021), ch. 3, “Data”.
• Institutional trust in opaque systems: “An Institutional Approach to Epistemic Trust in Opaque AI Systems”, AI and Ethics (2025). Author to confirm before assigning.