← Workshop
HAL 9000 — the on-screen AI we've been arguing with for fifty years

Workshop 1

Introducing AI for DH Pedagogy

Wednesday, May 13, 2026 · 10 AM – noon · CHDR

Part 1

ELIZA and the ELIZA effect

Modern chatbots owe a strange debt to a much simpler chatbot launched in 1966...

Excerpt of a recreated ELIZA / DOCTOR conversation, in the spirit of Joseph Weizenbaum's 1966 program

ELIZA, 1966 — Joseph Weizenbaum, MIT. "ELIZA — A Computer Program for the Study of Natural Language Communication Between Man and Machine" (CACM, 1966) →

I was startled to see how quickly and how very deeply people conversing with DOCTOR became emotionally involved with the computer and how unequivocally they anthropomorphized it.

Once my secretary, who had watched me work on the program for many months and therefore surely knew it to be merely a computer program, started conversing with it. After only a few interchanges with it, she asked me to leave the room.

— Joseph Weizenbaum, Computer Power and Human Reason (1976) — the ELIZA effect named

The two roads from 1966

  • Weizenbaum: spent the rest of his life as a critic of AI hype. His worry wasn't that ELIZA was too smart — it was that humans were too willing to believe it was.
  • Marvin Minsky (across the hall): founded the MIT AI Laboratory, took Pentagon funding unhindered, championed the dream.
  • Both roads are still open in 2026. We are walking them at the same time.
We are trying to do the same thing Weizenbaum tried to do: educate people about how these systems work, dispel the notion that they are thinking machines with a semblance of human understanding, and provide a model of how to think about them instead.

— Emily Bender & Alex Hanna, The AI Con (2025) — explicitly picking up Weizenbaum's thread

Exercise

Try ELIZA

Five minutes — open it up and notice what your brain does. We're less likely to anthropomorphise than ELIZA's 1966 audience, because we've already met more sophisticated chatbots — but try to imagine what it would be like to talk to a system like this for the first time.

  1. Open E.L.I.Z.A. Talking (Norbert Landsteiner's web port).
  2. Type as if you're talking to a therapist. Use whatever's actually on your mind.
  3. Notice the pattern: ELIZA reflects your nouns, asks open questions, occasionally trips.
  4. Catch yourself anthropomorphising — even faintly. That residue is the ELIZA effect. It's still the strongest force in the room when we use ChatGPT.

Part 2

AI, LLMs, and how they work

We've come a long way from ELIZA, which didn't have any of the elements we consider to be a part of modern AI. Current chatbots are mostly powered by LLMs: Large Language Models.

Hand-drawn-style infographic showing the AI → Machine Learning → Deep Learning → LLMs hierarchy. Tagline: 'When the news says AI did X, almost always: an LLM did X.'
Detailed LLM explainer infographic: training data → tokenization → transformer model → train to predict next token → generate text. Plus key points and use cases.

An LLM is a machine learning model that can complete a sentence of text. Give the model the phrase "the cat sat on the " and it will (almost certainly) suggest "mat" as the next word in the sentence.

As these models get larger and train on increasing amounts of data, they can complete more complex sentences — like "a python function to download a file from a URL is def download_file(url): ".

LLMs don't actually work directly with words — they work with tokens. A sequence of text is converted into a sequence of integer tokens, so "the cat sat on the " becomes [3086, 9059, 10139, 402, 290, 220]. This is worth understanding because LLM providers charge based on the number of tokens processed, and are limited in how many tokens they can consider at a time.

— Simon Willison, How Coding Agents Work

Exercise

What's a token?

Two minutes — take Willison's example off the page:

  1. Open the OpenAI Tokenizer in a new tab.
  2. Paste a sentence you'd actually type into Claude or ChatGPT.
  3. Note the token count and which 'words' the tokenizer splits into multiple pieces.
  4. Try the same sentence in another language, with a typo, or stuffed with your discipline's jargon. Watch what changes.

What is generative AI?

  • Per the MLA-CCCC task force: 'computer systems that can produce, or generate, various forms of traditionally human expression — language, images, video, music.'
  • Today we're focused on the language part. We touch the others in later workshops.
  • Transformer architecture + a vast scraped training corpus + reasoning layers on top.
  • MLA-CCCC working paper →
Definitional slide: What is generative AI?
"AI" is a marketing term. It doesn't refer to a coherent set of technologies. Instead, the phrase is deployed when the people building or selling a particular set of technologies will profit from getting others to believe that their technology is similar to humans.

— Emily Bender & Alex Hanna, The AI Con

What sits on top of the base model

  • RLHF (Reinforcement Learning from Human Feedback): humans rank model outputs; the model learns to prefer the ranked-good ones.
  • System prompts: invisible instructions that set tone, role, refusals.
  • Tools: web search, code execution, file access — the model decides when to call them.
  • Claude has three models that are regularly updated. Opus is the most expensive and also the most capable; Sonnet is in the middle; Haiku is the fastest, smallest, and cheapest. Current model lineup →

Part 3

The ChatGPT moment

November 30, 2022. The technology wasn't new. The interface was. And the interface was the event.

Before ChatGPT, programmers already had this

  • June 11, 2020: OpenAI releases the GPT-3 API in private beta. Developers could already prompt the model behind a waiting list.
  • June 29, 2021: GitHub Copilot enters technical preview — built on OpenAI Codex, an OpenAI model descended from GPT-3.
  • November 18, 2021: the GPT-3 API drops the waitlist; it's now open to any developer.
  • Working with these tools required code literacy, so the impact was primarily in the digital humanities: in fields like electronic literature, we were fine-tuning models for poetic and playful text, but not thinking about AI as a serious author...yet.

Sources: TechCrunch, GitHub Blog, VentureBeat.

November 30, 2022

Sources: Sam Altman on X, Reuters / UBS analyst note.

The chatbot was the wrapping. The model was already there.

  • What changed on November 30, 2022 wasn't a leap in capability — it was a leap in access.
  • Anyone with a browser and an email address could now talk to a system that had previously lived behind an API key, a credit card, and a developer account.
  • Every higher-ed conversation about generative AI is a conversation about that access. The crisis isn't that the model exists — it's that the model became visible to everyone at once.

Part 4

Where this leaves us — and where it's going

The chatbot was the wrapping. Now the wrapping is shifting again — toward agents that use the LLM as a substrate, not as a product.

The conversation in 2026

From plagiarism panic to agent panic

  • Plagiarism panics, AI-detector arms race, faculty caught between adoption pressure and pedagogical challenges.
  • "A new app offers to do students' work for them. Professors aren't ready."
  • The conversation has shifted from did the chatbot write this? to did the agent do the work?

The Chronicle of Higher Education — "Will Agentic AI Break Higher Education?"

Chronicle of Higher Education article header: Will Agentic AI Break Higher Education?

What is an agent?

  • "An LLM agent runs tools in a loop to achieve a goal." — Simon Willison, On the definition of "agent"
  • Tool use: the model can call functions you give it (web search, file edit, run command).
  • Planning: the model decides what to do next, in what order, and when to stop.
  • Subagents: the model can spawn helper instances of itself for sub-tasks.
  • The vocabulary is doing a lot of work — but the foundation is still the LLM we just described.
Screenshot of OpenClaw, an agentic AI tool: 'The AI that actually does things — clears your inbox, sends emails, manages your calendar, checks you in for flights.'
1979 IBM internal training slide: A COMPUTER CAN NEVER BE HELD ACCOUNTABLE / THEREFORE A COMPUTER MUST NEVER MAKE A MANAGEMENT DECISION.

1979 IBM training slide, surfaced by Simon Willison: "A computer can never be held accountable" → — the principle agents quietly violate every time they act on our behalf.

Einstein.AI: "Has his own computer"

  • An agentic harness aimed straight at undergraduates: logs into Canvas for you, watches lecture videos, reads & writes essays, participates in discussions, works while you sleep.
  • "Einstein isn't a chatbot. He has a full virtual computer — he can browse the web, watch videos, read PDFs, and interact with any site just like you would."
  • This is the higher-ed crisis the Chronicle headline is pointing at — not that the chatbot writes the paper, but that the agent does the whole assignment loop.
  • companion.ai/einstein →
Screenshot of the Einstein.AI 'How it works' page: Einstein has a full virtual computer with a browser, logs into Canvas, watches lectures, reads and writes essays, participates in discussions.

The LLM is now the substrate, not the product

  • 2022: the chatbot is the product (ChatGPT).
  • 2026: the chatbot is one interface among many. The model lives inside a harness that wraps it in tools, memory, planning, and access.
  • Coding harnesses — Cursor, Claude Code (Web + CLI), GitHub Copilot, OpenAI Codex — wrap the model in a repository, file edits, terminal commands, deployments.
  • General agentic harnesses — Cowork, OpenClaw, Einstein, MCP-connected research tools — wrap the model in calendars, inboxes, browsers, learning management systems.
  • Same family of models underneath. The harness is where pedagogical choices live now. Everything in this series — Projects, Artifacts, Code Web, CLI — is a harness choice.

Hands-on, Part 1

Chatbot tour: five conversations

Same prompt, five tools. We'll move back and forth between these throughout the workshop while prioritizing Claude — to build an understanding of the different tools available.

The five tools

  • Gemini — Google's free chatbot, integrated with the search and workspace stack the model many of our students already use.
  • DeepSeek running locally through Ollama — open-weights model on a personal machine. No API call, no data leaving the room. The local demo.
  • ChatGPT — the chatbot that defined the genre.
  • UCF Copilot — Microsoft's wrapper around an OpenAI model, with UCF's data-privacy framing layered on top. The only tool UCF currently supports or allows for use with restricted data. (UCF AI policy →)
  • Claude — Anthropic's flagship; the tool we use for the rest of the series.

Live demo

Same prompt, five tools

We pick one prompt — something with humanities texture, like asking for a summary of a primary source you brought, or a teaching question you actually have — and run it through all five. Notice the verbs each tool reaches for, the hedges, the refusals, the formatting defaults.

Why Claude / why Anthropic

  • Tool surface. Projects + Artifacts + Code Web + CLI is currently the broadest non-coder-accessible tool stack across providers. The framework transfers; the build path is best on Claude.
  • Safety + interpretability research culture. Anthropic publishes more on alignment and model behavior than its peers. For a humanities audience, the publication record is part of the choice.
  • The series is designed so the framework transfers — I'll point at where it would, week by week.

Exercise

Build our own ELIZA — together, in Claude Artifacts

Each person picks a style, tone, and subject based on your own interests — something tied to your discipline, your teaching, or a corner of your life you'd be willing to talk to a chatbot about. The pattern is the constraint; the personality and the topic are yours.

  1. Open Claude and start a new chat.
  2. Prompt: "Build a Claude Artifact: a chatbot in the spirit of Weizenbaum's 1966 ELIZA — pattern-match the user's input and reflect it back as questions. Don't pretend to know more than the pattern." Then add your own style, tone, and subject.
  3. Open the Artifact in the side panel. Talk to it the way you talked to ELIZA earlier.
  4. Compare. What does Claude do that pattern-matching can't? What does it do that pattern-matching also did? Where does the ELIZA effect kick in this time?
  5. Save the Artifact URL — bring the comparison to next week's discussion.

Three things to leave with

After today you should be able to explain to a colleague: where ELIZA started this, what an LLM actually is and what sits on top of it, why the ChatGPT interface was the event, and what "agentic" adds to the conversation.