Course Outline
Artificial Intelligence in the Summer
Artificial Intelligence in the Summer
| Professor | Dr K Darcy Otto |
| Title | Artificial Intelligence in the Summer |
| Code | Special |
| Credits | None |
| Term | Summer 2026 |
| Times | M 2–5, T–F 9–12 |
| Location | Commons 253 |
| Delivery | Fully in-person |
| Contact | |
| Office Hours | Anytime |
Artificial intelligence is a strange object: built from text and images, fluent in conversation, and yet nothing like a mind. This course asks what AI systems are, how they work, and how we should act with respect to them.
Our study will fall into three parts. First, framing: what is an AI model, and what does a new technology do to the people and institutions that adopt it? Second, machinery: how do models turn data into results? Third, ethics: how should we act in a world where AI is increasingly prominent, and who is responsible when things go wrong?
One question runs through the whole week: what should AI be used for, and who answers when it fails? You will write a provisional answer on the first afternoon and return to it on the last day.
Expect close reading, hands-on work with AI tools, and sustained argument. No technical background, or background in mathematics, is required.
A case document on a clinical AI failure will be distributed in class.
Bennington. Complete Summer Experience Schedule.
3Blue1Brown. But what is a neural network? | Deep Learning Chapter 1
3Blue1Brown. Transformers, the tech behind LLMs | Deep Learning Chapter 5
TuringTest. Can you pass the Turing Test?
SightEngine. Which image is AI?
Visual Culture Exercise. Instructions for Clark visit
Timing and emphasis may shift as the week develops.
Two short readings, completed prior to Monday afternoon:
Read Chiang first. He makes a specific and contestable claim about what a language model is; Franklin then supplies the wider framework for thinking about any technology at all. We return to Chiang on Wednesday, once you have the technical background to evaluate his claim again.
What does a technology do to the people who use it?
We open with the question that runs through the week, and you write a one-sentence answer to it before we begin. Then Franklin’s argument that a technology is not a device but a practice, a way of doing things that reshapes the people who take it up. Her distinction between holistic and prescriptive technologies applies to any technology, and we turn it on this one: we sort current AI deployments, from tutoring systems to radiology triage to hiring screens, along three axes. Is it holistic or prescriptive? Who benefits? Who bears the risk?
For the evening:
What is this actually good at, and how would I know?
The practical day. You present your domain claim and we test it against Mollick’s jagged frontier. Then we work through a task the system does startlingly well, a task it fails at with complete confidence, a naïve prompt set beside a considered one, and sustained refinement of a prompt of your choosing.
In the afternoon we practise the interpretative method the week’s project depends on, using the Exekias amphora showing Achilles and Ajax at their board game. Describe the materials, then the parts, shapes and composition, and only then infer meaning. The discipline lies in withholding the interpretative claim until the description has earned it.
For the evening:
What do these systems do, as against what they appear to do?
We answer those questions by building a neural network by hand on a spreadsheet: inputs, weights, bias, and a threshold. It learns a simple logical function. Then a single unit fails at one it provably cannot compute, and we see why adding a layer fixes it. From there we work up to how a large language model is structured: how words become numbers, how the system predicts what comes next, and how attention lets it weigh some parts of a sentence more heavily than others. We close by returning to Chiang and asking whether his metaphor holds.
The afternoon is at the Clark, where you select a single work and stay with it: structured looking, notes, blind tracing. One rule, stated plainly. No looking the work up and no AI until after Thursday’s presentations, please.
For the evening:
What does using this tool do to the person using it?
Aristotle first: what conditions must hold for an action to be genuinely one’s own, which kinds of ignorance excuse and which do not, and how character is formed by practice rather than instruction. Then Vallor on moral attention, and what it costs us to hand morally significant tasks to machines. If virtue is built by exercising a capacity, what becomes of capacities we give away?
The rest of the morning belongs to the interpretations, presented and discussed: the work in its human form, examined before any comparison exists.
For the evening:
When the system fails, where does responsibility land?
Elish argues that automated systems produce moral crumple zones: the human nearest the failure absorbs blame for a system they neither designed nor could fully see. We test this against a case in which a clinical AI tool fabricates a warning and a patient is harmed. You allocate a hundred points of responsibility across five parties (the nurse, the clinic, the vendor, the company that built the underlying model, and the regulatory environment) and defend that allocation to the others. Disagreement is the point.
The rest of the morning is the final presentations: your own interpretation, the model’s interpretation of the same work, and a reflection on what the differences reveal.