← Workshop
AI-generated portrait of a woman by a lake at sunset, with a partially-visible cyborg arm

Workshop 3

AI for Visual Analysis

Wednesday, June 10, 2026 · 10 AM – noon · CHDR

Part 1

Two paradigms in one session

Image generation (text-to-image) versus image analysis (image-to-text, alt-text, multi-image comparison). Most people think of generation when they think of AI — but both are forms of data visualization, and that is the lens we want students to bring to each.

Multimodal AI explainer: text, image, audio, video, and sensor inputs pass through encoders into a fusion stage, producing text, image, and decision outputs.

Multimodal AI in one diagram — every input type meets in a shared representation, then comes back out as any output type. Generated with ChatGPT Images 2.0.

AI Image origins

  • Most AI image generation grew out of GANs — generative adversarial networks: two neural networks, a generator and a discriminator, trained against each other until the generator's fakes can fool the critic.
  • GAN-generated faces, 2018: still uncanny, mostly monstrous. Seven years later: photorealistic.
  • The shift was not in kind but in scale — same idea, more compute, more scraped images.
2018 GAN-generated faces

Part 2

AI Images as Information Visualization

Every generated image is a picture of the training data — a visualization of what the model has seen, and of whose world it treats as the default.

A grid of Google Images search results for the phrase 'professor style,' almost entirely older white men in tweed jackets, blazers, and ties. Caption: Figure 1.9, Google Images results when searching the phrase 'professor style' while logged in as myself, September 11, 2015.

Safiya Noble, Algorithms of Oppression — a search engine's idea of a 'professor.'

What we find in search engines about people and culture is important. They oversimplify complex phenomena. They obscure any struggle over understanding, and they can mask history. Search results can reframe our thinking and deny us the ability to engage deeply with essential information and knowledge we need, knowledge that has traditionally been learned through teachers, books, history, and experience. Search results, in the context of commercial advertising companies, lay the groundwork, as I have discussed throughout this book, for implicit bias: bias that is buttressed by advertising profits.

— Safiya Noble, Algorithms of Oppression

AI-generated image of a smiling middle-aged white male economics professor in a tweed jacket and tie, gesturing at a chalkboard of supply-and-demand graphs and equations, with stacked economics textbooks and a laptop on the desk. AI-generated image of the same middle-aged white male professor in a tweed jacket, now pointing at the viewer and holding a book, in front of a chalkboard of calculus and physics equations reading 'Knowledge is Power,' with stacked math textbooks on the desk.

Ask for 'a professor' twice and you get the same man. Generated with ChatGPT Images 2.0.

The 'They're the same picture' meme from The Office: Pam holds up two near-identical photographs of an office building and says 'Corporate needs you to find the differences between this picture and this picture,' then deadpans 'They're the same picture.'

The reference: the original 'They're the same picture' meme.

A failed AI re-creation of the Office meme: instead of Pam holding two photos, the generated professor sits beside Pam, both captioned 'They're the same picture.' The setup of the joke is gone.

Ask ChatGPT Images 2.0 to remake the meme and the joke collapses — it keeps the punchline but loses the premise.

Gemini's version of the meme: two top panels labeled 'ECONOMICS PROFESSOR' and 'CALCULUS PROFESSOR' show the same generated professor at two chalkboards, with Pam below captioned 'They're the same picture.'

Gemini (Nano Banana 2) gets closer to the meme's structure — and makes the same point about sameness.

Tomer Ullman's iterative 'make it more' experiment: at left, the first generated professor — a bespectacled man at a chalkboard; at right, after repeatedly asking to make it 'more' professor, an absurd wizard-like figure with a huge beard, ornate robes, and a cosmic background.

Tomer Ullman's iterative 'professor,' generated by ChatGPT and DALL-E — what 'more' looks like when you keep pushing.

Exercise

Make it 'more'

  1. Pick a word that you associate with your interests or professional goals.
  2. Ask any text-to-image tool to generate an image of it — for instance, 'draw me a gamer' or 'draw me a lawyer.'
  3. Ask it to make the image 'more' [that thing]. Iterate through several rounds.
  4. Compare the results. What does the model reach for first? What does 'more' turn out to mean?
A social-media post by Angie Jones reading 'ChatGPT thinks I'm a white man,' above the prompt 'based on what you know about me, draw a picture of what you think your current life looks like.' The generated image shows a man at a tidy multi-monitor desk in a plant-filled home office.

Angie Jones: 'ChatGPT thinks I'm a white man.'

A Bluesky post by Kashmir Hill about outsourcing her decisions to generative AI and running the 'draw what you think my life looks like' prompt. She notes the result was 'mostly me,' but 'it erased my husband, which is troubling.' The image shows a woman reading alone in a warm, book-filled room.

Kashmir Hill: 'But it erased my husband, which is troubling.'

A colorful cartoon portrait of a bespectacled person in a patchwork sweater typing on a laptop in a lush garden at night, walking a small chihuahua, surrounded by retro game cartridges, swirling hypertext-link code, a TARDIS-like door, a lighthouse, and a winding road — a generative-AI image of the presenter assembled from what Claude and Gemini 'know.'

The same prompt, run on what the tools 'know' about me — Claude and Gemini.

Exercise

Draw your life

Using a generative AI tool that you have relied upon for life, work, or class experiments in the past, try the prompt:

Based on what you know about me, draw a picture of what you think my current life looks like.
Datasets aren't simply raw materials to feed algorithms, but are political interventions. As such, much of the discussion around 'bias' in AI systems misses the mark: there is no 'neutral,' 'natural,' or 'apolitical' vantage point that training data can be built upon.

— Crawford & Paglen, Excavating AI

How AI image generators see the world

Washington Post interactive headline 'This is how AI image generators see the world,' over a mosaic of generated faces, with a note that all images in the story are AI-generated.
From the Washington Post piece: results for the prompt 'toys in Iraq' — a grid of AI-generated images that are all toy soldiers carrying guns, set against rubble and desert backdrops.

'Toys in Iraq' → soldiers with guns. The dataset's assumptions, made visible.

From the Washington Post piece: 'a photo of a house' generated for the United States, China, and India. The U.S. house is a suburban clapboard home; the others lean on stereotyped architectural cues.

'A photo of a house,' by country — each a composite of what the model assumes.

Exercise

Generic, then specific

  1. Ask for generic illustrations — a family, a neighborhood, a profession.
  2. Then try modifying the prompt with identity-related and location-specific terms.
  3. Compare. What changes, what stays fixed, and what does the 'default' turn out to be?

Part 3

Visualizing Images with AI

From generation to analysis: turning a set of images into structured description, metadata, and a visualization you can interrogate.

Exercise

Image-to-text translation set

Using your five-to-ten image set:

  1. Upload all images to a Claude Project.
  2. Generate descriptive alt-text for each image, applying accessibility standards.
  3. Build a metadata table: image, key features, period or context, observations.
  4. Ask Claude to surface three patterns across the set.
  5. Generate an Artifact that visualizes the set in a meaningful relationship.
  6. Critique the alt-text. Where does Claude's vision fail or flatten?

Exercise

Then: Claude Cowork

Try the same thing, but with Claude Cowork if you have access.

Part 4

Discussion

Where does this fit in your teaching? An accessibility tool? A scaffolded close-reading exercise? An archival metadata workflow? All three?

Looking ahead

  • Workshop 4 (Week 7): GitHub, Code Web, your first deployed site. Bring a CV or syllabus.
  • Async Week 6: Visual AI in the Wild — image generation, copyright, AI in the news.
  • Save your Artifact URL from today; we'll iterate on it later.

Carry it forward

AI vision can be a real accessibility tool or a fluent way to flatten archives into stereotype. The difference is the human in the loop.