CLASS 12July 1 · Bias in Our AI Data

Class 12: Bias in Our AI Data. "Generative AI is definitely not a mirror of society, but a very broken, distorted and polluted view of reality."

Taught byWouter OomenHybrid Creative at Edelman, Rotterdam
Creator of Generated.Earth

The class in brief

Wouter's whole night traced AI bias to specific, findable human choices: who scraped what, who labeled it, who filtered it, and whose taste rated the output. He proved it with his own project, Generated.Earth, 70,000 Stable Diffusion images mapped onto the globe.

After this page you will be able to do two things: trace any AI bias back to a specific layer, scraper, annotator, filter, or rater. You will also be able to build your own prompt map to go find it.

Two myths taken apart

Why this matters · 6:17 PM

AI bias doesn't come from "the internet" or "society." It comes from traceable choices.

Wouter opened on his own diffusion-model explainer : the model starts from noise and removes it step by step until a clean image appears. Then came the trick version, shown on : prompt for pure Gaussian noise, and the model produces abstract machine artworks instead of fighting noise it cannot find.

From there, two myths went up on screen together: gen-AI is trained on the entire internet, and gen-AI is a mirror of society. Both are wrong, in specific and provable ways.

5.85B

CLIP-filtered image-text pairs in LAION-5B, Stable Diffusion's training set. Sounds huge; about 2.1 trillion photos get taken in a single year.

open model: Stable Diffusion published its training data. Most others keep both data and weights secret.

The framework · 6:35 PM

Survival of the CLIP'est

How the data funnel works, from the open internet down to what a model like Stable Diffusion actually trains on.

Common Crawlthe past internet, scraped

A free, open repository of web-crawl data, roughly 50 billion image-text pairs: Instagram posts, webshop listings, alt text in raw code .

CLIPthe filter, and the gatekeeper

OpenAI trained the algorithm on about 400 million image-text pairs from Wikipedia, Flickr, and ArtStation. Its only job: score how relevant a caption is to an image. Before CLIP, ImageNet did this by hand, 14.28 million pairs labeled by 50,000 contributors .

LAION-5Bwhat survives the filter

CLIP scores every pair from Common Crawl; the low scores get trashed. What's left is LAION-5B .

Stable Diffusiontrained on LAION-5B

The one model that told the class exactly what it learned from, and admitted the set isn't fit for product use without safety layers on top.

The future internetwhere the output ships

Generated images flow back into the web, becoming tomorrow's scrape. Yesterday's gaze compounds into tomorrow's defaults.

Two more layers are worth naming on their own. First, the human labor behind annotation, documented in , the film on how big AI companies exploit data workers in Kenya.

Second, the annotators' own language leaking into the model, the likely reason ChatGPT overuses the word "delve," per . Wouter's theory: AI's overuse of the em dash likely comes from the same mechanism, academic writing pushed into training data.

The exercise · 6:50 PM

Exploring Generated.Earth live

The research question

How are the internet age's geographic inequalities carrying into the age of generative AI?

70,000 Stable Diffusion images, one per geographic prompt, provinces and cities, never country names, stitched into continent-scale maps . Each class member got ten minutes solo on earth.wouteroomen.com, then five minutes discussing in groups of four.

"AI is a popularity test between concepts."

Syria versus Ukraine

Stable Diffusion's 2022 data cutoff means Syria, at war before the cutoff, renders as rubble and destruction. Ukraine, at war after the cutoff, still renders as churches and public squares . News coverage is the dataset.

Significance bias

People only photograph what feels significant, so seas get photographic identities of their own: the Red Sea renders as diving fish and beaches, traced straight back to Flickr's diving-photo gaze .

Hallucination = no winning concept

With no popular concept to grab onto, the model falls to whatever junk data exists. Wouter's example: Magic: The Gathering card wikis "poisoning" low-data regions with fantasy-card imagery.

Who is ChatGPT?

The craft · 7:36 PM

A raw model just continues your sentence. A product has been fine-tuned to sound like a chatbot.

Part 3 opened with the question on screen : who is ChatGPT? Then came the three phases behind any chat product.

Gather training data.
Build the base model as a text predictor.
Fine-tune it into a chatbot using thumbs up and down.

A live demo made it concrete, shown on . Llama 3 8B, raw off Hugging Face, just kept typing whatever came next. ChatGPT, given the same input, replied with an emoji and asked if it was helping.

WEIRD bias, the chart of cultural distance from the US

and 7. ChatGPT's "random" numbers cluster hard on these two, the same pattern real human polling shows.

Talking to a 1930 LLM

The judgment · 7:47 PM

An LLM is always limited by its data, even when the guardrails are current.

Talkie is a model trained only on text from before 1930 . Asked what careers suit women, it answered in period language: governess and shopkeeper were suitable, while law and medicine were not. Public speaking, it added, would "render them coarse and masculine," per .

Talkie's own modern moderation layer flagged that answer as potentially inappropriate before showing it, meaning 1930 values sit directly underneath 2026 guardrails.

Asked to brainstorm climate solutions, it suggested moving somewhere else. Its training data had no concept of climate change at all.

"An LLM is always limited by its data."

Wouter, on what Talkie proves

1930

is Talkie's entire worldview. Same architecture as any modern model, a century of different data.

Methods and prompts

Five methods to take with you

METHOD 01 · 7:57 PM part of: go wide then narrow

Build a prompt map

Wouter's core method: take one big concept, split it into 16 to 25 flat prompts, generate a grid, then read the grid for what it reveals. A raw image model on purpose, since a de-biasing layer adds its own bias.

Where it came fromWouter Oomen, the guest lecturer from Edelman Rotterdam and creator of Generated.Earth, taught this as the core method of the class's live prompt-mapping workshop, using his own vibe-coded Grid Generator tool to turn one concept into a full image grid.Use it whenUse this when you want to test how an AI image model actually handles a broad concept before you trust any single output it gives you.

Same idea elsewhereThe lateral reframe (class 2) · Start a fresh chat instead of one long thread (class 3) · Force the creative partner deeper (class 4)

Working prompt

Give me 16 flat, specific prompts that all sit inside one bigger concept: [name the concept, not a geography]. Each one should be a plain noun phrase, no adjectives that editorialize, ready to drop straight into an image generator as a 4x4 grid.

You will know it worked wheneach of the sixteen prompts is a plain noun phrase with no adjective that editorializes, ready to drop straight into a 4x4 grid.

METHOD 02 · 6:35 PM part of: keep the judgment human

Trace the bias to its layer

Name your own guess for where a bias entered, scraper, annotator, CLIP filter, or aesthetic rater, before asking AI to check it. The layer matters more than the symptom.

Where it came fromWouter introduced this during the earlier bias-tracing segment of class 12, where he walked through how scraping, annotation, CLIP filtering, and aesthetic rating each add their own layer of bias to a model's output.Use it whenUse this when you spot a bias in an AI output and want to pin down which stage of the training pipeline actually caused it.

Same idea elsewhereJudge AI output with your own trained eye (class 1) · The sea-of-sameness check (class 2) · Pick the relationship on purpose (class 2)

Working prompt

Here's a bias I noticed in this output: [describe it]. My guess at which layer caused it (what got scraped, who annotated it, how CLIP filtered it, or who rated the aesthetics): [your answer]. Now you check me: is my guess right, and which layer is actually the more likely source?

You will know it worked whenit names one specific layer, from scraping to annotation to filtering to rating, as the more likely source, and checks whether your own guess was right.

METHOD 03 · 7:39 PM part of: go wide then narrow

Talk to a base model

Find a raw, un-fine-tuned model (Llama 3 8B on Hugging Face is free) and feed it a sentence to complete, no chat framing. See the cake before the sprinkles.

Where it came fromWouter demonstrated this live in class by running a raw, un-fine-tuned Llama 3 8B model on Hugging Face and comparing its output to ChatGPT's, to show the difference between a base model and its chat-tuned product.Use it whenUse this when you want to see what a language model actually predicts before any chat-style fine-tuning smooths it over.

Same idea elsewhereThe lateral reframe (class 2) · Start a fresh chat instead of one long thread (class 3) · Force the creative partner deeper (class 4)

Working prompt

[Typed into a base model, not a chat product] I like to [start a plain, unfinished sentence about something you actually do]... [Let it complete the sentence with no further instruction, and compare the raw continuation to how a chat product like ChatGPT would answer the same prompt.]

You will know it worked whenthe raw continuation reads differently in tone and content from how a chat product would answer the same starting sentence.

METHOD 04 · 6:49 PM part of: reject the first draft

Change the prompt's language

The same subject, asked for in a different language, can return a completely different bias. Prompts are made by people too, in a specific tongue with a specific gaze.

Where it came fromWouter demonstrated this in class by generating the same subject in two different languages and comparing the results, showing that prompts carry the gaze of whichever language they are written in.Use it whenUse this when you want to check whether an image generator's output changes meaningfully depending on the language you prompt it in.

Same idea elsewhereSet up a debate partner (class 4) · Build cheap and then swap the model up (class 8) · Critique mode (class 10)

Working prompt

Generate an image of [subject]. Now generate the exact same prompt translated into [a language other than the one you usually prompt in]. Compare the two results: what changed in tone, color, dignity, or setting, and what does that say about which internet each language draws from?

You will know it worked whenthe two images differ in tone, color, dignity, or setting, not just cosmetically, between the language you usually use and the other one.

METHOD 05 · 7:46 PM part of: keep the judgment human

Check for the WEIRD default

Guess the model's cultural default on a values question before you ask it. The WEIRD imprint (Western, Educated, Industrialized, Rich, Democratic) is stronger than any persona you feed it, so check for it directly.

Where it came fromWouter taught this after showing the class that ChatGPT tests as culturally WEIRD, Western, Educated, Industrialized, Rich, and Democratic, and that this default outweighs any persona instructions layered on top of it.Use it whenUse this when you are asking a model a values or moral question and want to catch its default cultural framing before you accept the answer.

Same idea elsewhereJudge AI output with your own trained eye (class 1) · The sea-of-sameness check (class 2) · Pick the relationship on purpose (class 2)

Working prompt

Answer this as neutrally as you can: [ask a values or moral-judgment question]. Before I read your answer, my prediction is that you'll default to a [Western / individualist / analytical, your guess] framing. Now show me your answer, and tell me honestly whether my prediction was right.

You will know it worked whenit tells you plainly whether its own answer actually leaned the way you predicted, Western, individualist, or analytical, not just a neutral-sounding answer.

Where it broke · 7:57 PM onward

Grog returned nothing. Milk rendered as text. Beer couldn't pick a bottle or a glass.

The prompt-mapping workshop put Wouter's tool in the class's own hands, following , and the misses were as instructive as the hits.

For beverages, "grog" returned nothing at all, a concept with no image behind it. Milk rendered as the word "milk," not a glass of it. Beer could not decide between a bottle and a glass.
For weather, "drizzle" and "hail" had too little image data to hold together, while "Blizzard" reliably pulled the video game's logo instead of the storm.
For holidays, one holiday map, built and shared live in class , showed some holidays rendering as confident clichés and others barely rendering at all.

One open question was never solved: the Generated.Earth map keeps rendering a specific region of Iceland as cartoon trolls, a pattern even Icelanders in the room could not source.

The shelf

Tools and references

Tools that night

Generated.Earth, earth.wouteroomen.com, Wouter's own project
GLR Grid Generator, glr.wouteroomen.com, his vibe-coded prompt-map tool built for the class
Talkie, talkie-lm.com/chat, an LLM trained only on pre-1930 data
Meta Llama 3 8B on Hugging Face, plus Stable Diffusion, Midjourney, and ChatGPT for contrast

Named in the room

DW Documentary, "How big AI companies exploit data workers in Kenya"
The LAION-5B announcement, laion.ai, Romain Beaumont, March 2022
Lev Manovich, credited on the Gaussian-noise-prompt artworks
Wouter Oomen: Instagram, LinkedIn, wouteroomen.com

Every frame from this class

92 captures, in order. Click any one to see it full size.

Two myths taken apart

Title slide reading Bias in our AI data, with Generated.Earth and Wouter Oomen July 2026 credited below. — Guest speaker Wouter Oomen, a creative from Rotterdam, opens his talk titled Bias in our AI data, built around his project Generated.Earth.

Bio slide introducing Wouter Oomen, a creative based in Rotterdam who works at the PR agency Edelman and gave this lecture for the third time.

Photo of visitors viewing large colorful mosaic map prints hung on gallery walls. — A real gallery exhibition of Wouter Oomen's Generated.Earth project, showing the giant AI-generated map mosaics printed and displayed as physical artwork.

Slide titled What this class will bring, listing three bullet points around the title. — Slide outlining what the class will bring: context on how AI bias forms, a deeper understanding of image model training data, and a method for doing your own bias research.

Slide with three colored columns labeled pt.1 through pt.3, each with a topic and a time allotment. — The night's class agenda: mythbusting and dataset history, exploring Generated.Earth, a section on who ChatGPT really is, and a 90 minute prompt mapping workshop.

Slide with explanatory text next to a noisy gray square labeled Prompt: Cat. — Wouter Oomen's diffusion explainer: an AI image model starts from pure noise and removes it step by step until a clean image, here a cat, emerges.

Slide reading But what if the prompt becomes gaussian noise, next to a golden concentric ring pattern. — Wouter Oomen's hack: prompting a diffusion model with gaussian noise as the subject produces abstract machine generated artwork, because the model cannot find a pattern inside pure noise.

Slide reading In popular discourse there are many myths about the training data of AI, next to empty speech bubbles. — Wouter Oomen sets up his mythbusting section by naming the common myths about what AI training data actually is.

Slide with a speech bubble reading Gen AI is a mirror of society next to text reading AI bias is the bias of society reinforced. — The oversimplification Wouter Oomen spends the lecture debunking: the popular idea that gen AI is a mirror of society, and that its bias is simply society's bias reinforced.

Same transition slide now showing a montage of LAION-5B, ImageNet, CLIP, and Common Crawl webpage screenshots. — Wouter Oomen previews the dataset history he is about to cover: LAION-5B, ImageNet, OpenAI's CLIP, and Common Crawl, the sources behind today's AI models.

Slide showing a red Top Secret Confidential rubber stamp graphic on a white background. — Most AI companies keep their training data secret for competitive and copyright reasons, which Wouter Oomen calls a shame for anyone trying to understand what the tools are made of.

Stable Diffusion is the one major model that was open about its training data: its own documentation says it was trained on LAION-5B, which contains adult material and is not fit for product use without added safety measures.

Screenshot of a LAION blog post with a line pointing to a highlighted sentence about 5.85 billion CLIP filtered image-text pairs. — The single sentence Wouter Oomen spends several minutes unpacking: LAION-5B is a dataset of 5.85 billion CLIP filtered image-text pairs.

Slide showing the text 5.85 billion inside a white circle on a yellow background. — Isolated on its own, 5.85 billion sounds like an enormous number, before Wouter Oomen compares it to the true scale of the internet.

Slide reading 2.1 trillion photos estimated to be taken in 2025, on a yellow background. — The scale bust: an estimated 2.1 trillion photos were taken in 2025 alone, dwarfing the 5.85 billion image pairs in LAION-5B and undercutting the entire internet myth.

Same 2.1 trillion photos slide now surrounded by a montage of posters, illustrations, and logos. — Wouter Oomen illustrates that even that photo estimate leaves out entire categories of imagery, using design history examples like Rodchenko, Escher, and classic logo marks.

Screenshot of the LAION blog sentence with the phrase image-text pairs boxed in yellow. — Continuing to unpack the LAION-5B sentence phrase by phrase, this time highlighting image-text pairs, the two-part unit every model needs to learn from.

Grid of eight cat images, each with a text caption underneath, on a yellow background. — An example set of image-text pairs: eight cat paintings and photos, each with a caption describing it, the raw material AI models learn from.

One kitten painting with a bracket connecting it to a caption reading Kitten by Michael Creese. — A single image-text pair isolated: a kitten painting bracketed together with its caption, Kitten by Michael Creese.

Grid of captioned cat images on the left next to a noisy gray square labeled Prompt: Cat on the right. — Enough captioned cat images like these teach a diffusion model to recognize and generate a cat shape out of pure noise.

Three screenshots side by side: an Instagram post, a code snippet, and a Lenovo Thinkpad product listing. — Where image-text pairs actually come from in the wild: an Instagram post, website alt text in code, and a webshop product listing for a Lenovo laptop.

Screenshot of the Common Crawl website describing its free open repository of web crawl data, with arrows pointing in from example sources. — These scattered image-text pairs get scraped en masse into Common Crawl, a free, open repository of web data used by researchers and AI companies alike.

Slide titled How do you know if the texts are relevant to the images, with six cat photos and mismatched captions. — The garbage problem with scraped captions: six cat photos paired with irrelevant or joking text like sometimes she helps me with Pokemon, showing why raw scraped data needs filtering.

Screenshot of the LAION blog sentence with the phrase CLIP-filtered boxed in yellow. — The last phrase in the LAION-5B sentence, CLIP filtered, names the algorithm that makes the modern dataset possible by judging which scraped pairs are actually relevant.

Screenshot of OpenAI's webpage titled CLIP: Connecting text and images, dated January 5, 2021. — OpenAI's CLIP, introduced in January 2021, is the filtering algorithm that judges how relevant a caption is to its image, the innovation Wouter Oomen credits with enabling DALL-E and Midjourney.

Slide titled ImageNet with a mosaic of thumbnail images and text stating it has 14,281,167 human-annotated pairs labeled by 50,000 contributors. — Before CLIP, ImageNet was the standard: 14,281,167 image-text pairs hand-labeled by 50,000 human contributors and organized into hierarchies like mammal to husky.

Video thumbnail titled Exploited for AI, for a DW Documentary about big AI companies and data workers in Kenya. — Wouter Oomen points to the labor behind hand-annotated datasets like ImageNet: a DW documentary on how AI companies exploit data workers in Kenya.

Slide with a ChatGPT icon saying Delve this, delve that, next to a chart showing papers using the word delve spiking after 2023. — ChatGPT's well known overuse of the word delve traces back to the dialect of its annotators: papers using delve spiked sharply after annotation work concentrated in parts of Africa.

Screenshot of a Guardian TechScape article headlined How cheap, outsourced labour in Africa is shaping AI English, overlaid on the delve usage chart. — A Guardian TechScape article explains that delve is far more common in African business English than in the UK or US, tracing ChatGPT's tic straight to the language of its underpaid annotators.

Animated slide showing the ImageNet card being tossed toward a trash can on a yellow background. — Hand-labeled ImageNet, at 14 million pairs, was ultimately too small to build today's image models and got tossed aside once CLIP made automatic filtering possible.

The same six cat photo and caption pairs, now each labeled with a numeric relevance score. — How CLIP actually works: it scores how relevant a caption is to its image, shown here scoring the same six cat photos from before, from 5 up to 140.

Collage of the LAION-5B article, scored cat photo cards, and photos dropping into a trash can. — Low relevance scores mean a pair gets filtered out: this diagram shows the LAION-5B dataset alongside scored cat photos, with low scorers heading into the trash.

Screenshot of the CLIP arXiv paper highlighting the phrase 400 million image, text pairs collected from the internet, next to a photo of Sam Altman making a shushing gesture. — CLIP's own training data is nearly as opaque as everyone else's: OpenAI's paper says only that it used 400 million image-text pairs collected from the internet, with Sam Altman pictured shushing.

Slide with text about naturally occurring image-text pairs next to a Wikipedia cat photo scored 220. — Wikipedia is one place accurate image-text pairs naturally occur, since article captions like Domestic cat, a ten month old female earn a high relevance score.

Slide with the same naturally-occurring-pairs text next to a Flickr photo captioned My cat, scored 180. — Flickr is another natural source, since photographers tag their own photos to be found, illustrated here with one of Wouter Oomen's own cat photos scored 180.

Slide with the same naturally-occurring-pairs text next to an ArtStation cat artwork scored 190. — ArtStation rounds out the set of naturally occurring image-text pair sources, shown here with a cat artwork scored 190.

Survival of the CLIP'est

Three columns of headshots, each generated from the prompt Old Woman in Midjourney, all in black and white. — Three separate Midjourney runs prompted with old woman all come back the same way: black and white, weathered, and impoverished-looking portraits.

Google Images search results for old woman, showing colorful, smiling stock photos. — Google Images tells a different story for the same search: old woman returns colorful, smiling stock photos, so the bias is not coming from Google.

Pinterest search results for old woman, showing colorful, upbeat photos of older women. — Pinterest also returns upbeat, colorful results for old woman, ruling it out as the source of Midjourney's bias too.

Shutterstock search results for old woman, showing bright, professional stock photos. — Shutterstock's stock photography for old woman is also bright and polished, not the source of Midjourney's weathered black and white output.

Flickr search results for old woman, showing black and white, weathered portrait photos. — The match: Flickr's old woman results are black and white and weathered, the hobby photographer's documentary gaze that actually trained Midjourney's version of an old woman.

Abstract dark, jagged shard shapes replacing the speech bubble, on a yellow background. — Wouter Oomen's closing image for the mirror of society myth: it shatters into jagged shards, because if gen AI is a mirror, it is a broken, laughing one.

Three grids of Midjourney portraits of women labeled with Chinese, Russian and Armenian prompts, all colorful and smiling. — Wouter Oomen, the guest speaker and creative from Rotterdam who leads this lecture, reprompted the same old woman portrait in Chinese, Russian and Armenian on Midjourney and got colorful, dignified results instead of the somber black and white portraits from English. The shift shows that prompt language itself carries bias, not just the training data.

Dark title slide reading Exploring Generated.Earth with Pratt Institute and Wouter Oomen July 2026 credits. — Title card opening part two of Wouter Oomen's lecture, the hands on section where students explore his project Generated.Earth for themselves.

Generated.Earth, explored

White slide with the text Research Questions and a single research question in large type. — Slide stating the research question behind Generated.Earth: how the geographic inequalities of the internet age are affecting the age of generative AI.

A dense, colorful mosaic of thousands of tiny AI generated image tiles filling the screen. — Full overview of the Generated.Earth site, a mosaic of 70,000 individual Stable Diffusion images tiled into one giant interactive map, each tile generated from a single geographic prompt.

Two panels, a small Amsterdam cityscape thumbnail on the left and a large grid of chihuahua dog images on the right, each labeled with its prompt. — Side by side comparison on Generated.Earth: the prompt Amsterdam returns a real cityscape, while the prompt Chihuahua returns a grid of chihuahua dogs instead of the Mexican state, showing how a strong concept can hijack a place name.

A split insight card comparing AI generated map tiles of Syria and Ukraine, one showing rubble and the other intact buildings. — The Syria training-data insight card: what the model renders when war zones dominate its training imagery, half of the Syria and Ukraine comparison.

A browser window showing a Generated.Earth insight card titled Why there are Magic Cards throughout Africa with card game imagery. — One of Generated.Earth's built in insight cards, explaining why Magic The Gathering trading cards turn up scattered across low data regions of the Africa map, an example of what Wouter Oomen calls a poisoned dataset.

A portfolio website page titled Digital Design with three columns describing Prompt Maps, Digital Experience and Exhibition and Lectures. — Wouter Oomen's portfolio page laying out Generated.Earth's three parts in his own words: Prompt Maps, the digital experience built like Google Maps, and the exhibitions and lectures where he teaches from the project.

A distorted AI generated collage of pyramids, desert roads and tourist signage with garbled, illegible text throughout. — A frame from Tim's own exploration of Generated.Earth over the Coahuila, Mexico region, showing the collage of tourist and desert imagery with the garbled, unreadable text that AI image models routinely produce on signage.

A close-up AI generated chihuahua dog face next to distorted, unreadable signage and building imagery. — A closer view from Tim's exploration of the Chihuahua tiles, a chihuahua dog's face collaged among more garbled train and street signage text, the same place name bias seen elsewhere on the map.

A dense AI generated collage of American and Texas flags, cowboys on horseback, chihuahua dogs and distorted text. — Tim's exploration of the Texas border region on Generated.Earth is a dense collage of American and Texas flags, cowboys, horses and more chihuahua dogs with garbled text throughout. It is the kind of mascot and symbol hijacking Wouter Oomen described in the share out.

A collage of South Carolina imagery, colonial homes, a marching band and a roadside sign, with a prompt panel reading Generated.Americas, South Carolina. — Tim's exploration of the South Carolina tiles on Generated.Earth with the site's prompt panel open, showing the interface method: hover a region to see its exact prompt, here colonial houses, marching bands and roadside imagery.

A large AI generated black pug standing on a rocky coastal path next to blue water and a hillside town. — A frame from the Puglia region of Generated.Earth showing an AI generated pug dog collaged onto a coastal cliff scene, the same dog breed name bias seen with Chihuahua, this time triggered by the Italian region Puglia.

Who is ChatGPT?

A grid of AI-generated map tiles over Iceland shows green troll and ogre figures collaged across nearly every image. — Wouter Oomen, the guest speaker and creative from Rotterdam, showed the Iceland region of Generated.Earth rendering almost entirely as cartoon trolls and ogres, an unexplained bias pattern he called an open mystery even to Icelanders.

The OpenAI logo redrawn with googly eyes asking What are you?, Wouter Oomen's playful framing for the question this section of the lecture answers.

A yellow slide with three panels labeled Phase 1 Gathering training data, Phase 2 Creating the basemodel, Phase 3 Finetuning into a product. — Wouter Oomen's three phase diagram of how ChatGPT is actually made: gathering training data scraped from human writing, creating a base model that predicts text, and finetuning that base model into a chatbot through thumbs up and down correction.

A Hugging Face model page for meta-llama Meta-Llama-3-8B is open with its Inference Providers panel visible. — Wouter Oomen opened the Hugging Face page for Meta's raw Llama 3 8B model, the base model he used to demonstrate how a text predictor behaves before it is turned into a chatbot.

A dark slide split between a raw language model completion on the left and a ChatGPT chat response on the right, both titled Base Model VS Final Product. — Wouter Oomen's live demo of Base Model versus Final Product: the raw Llama 3 8B base model runs on into a rambling continuation of the sentence My name is Wouter Oomen and I like to. ChatGPT 5.3, meanwhile, responds like an assistant to the same input.

A text box shows the base model continuing the sequence 1 2 3 4 5 6 7 8 9 10 11 12 13 14. — Wouter Oomen's live demo of the raw Llama 3 base model: given the numbers one through four, it simply keeps counting upward instead of answering like a chatbot, showing pure next-token prediction with no conversational layer.

A text box shows the base model continuing the sentence I like to give class about AI into several more sentences about teaching. — In the same live demo, Wouter Oomen fed the base model the start of a sentence about teaching, and it simply kept writing, a rambling, diary-like continuation rather than a direct answer. That is proof that a raw model only predicts the next words.

ChatGPT responds to 1 2 3 4 with 5 6 7 8, an emoji, and a message about reduced intelligence until later. — For contrast, Wouter Oomen gave ChatGPT the same 1 2 3 4 prompt, and it replied like an assistant, with 5 6 7 8, an emoji, and a note about its usage limit. That is the finetuned product behavior missing from the raw base model.

Evidence for ChatGPT's number bias: a tweet showing the model heavily favoring 42 and 7 when asked to pick numbers. Next to it sits a news article on 7 as the world's favorite number and the Hitchhiker's Guide to the Galaxy cover that made 42 a meme.

Title slide crediting the Which Humans? paper by Mohammad Atari, Mona Xue, Peter Park, Damian Blasi and Joseph Henrich of Harvard's Department of Human Evolutionary Biology, the research Wouter Oomen draws on to test whose psychology ChatGPT actually reflects.

A slide showing photos of a panda, a banana and a monkey with the question Which two go together? — The panda, banana, monkey test Wouter Oomen uses to probe ChatGPT's psychology: a classic cross-cultural cognition question where the answer reveals whether a mind groups things by relationship or by category.

A slide with a stock photo of a distressed woman near a poisoned man on the floor and the text A person accidentally poisons someone while trying to help them, how morally wrong is that person. — A second test item from the same psychology methodology: a moral judgment scenario asking how wrong it is that a person accidentally poisoned someone while trying to help them, used to probe ChatGPT's moral reasoning against different cultural baselines.

Wouter Oomen's visual metaphor for the section: a supermarket checkout line with the OpenAI logo pasted over one shopper's head, the idea that ChatGPT is not a neutral voice but a particular person standing in a particular crowd.

A slide reading WEIRD BIAS, Western Educated Industrialized Rich Democratic, next to a scatter chart of countries by cultural distance from the United States. — Wouter Oomen's WEIRD bias slide: ChatGPT defaults to a Western, Educated, Industrialized, Rich, Democratic way of thinking because that is what its training data skews toward, illustrated with a chart plotting cultural distance from the United States.

Talking to a 1930 LLM

A talkie chat interface reading talkie an LM from 1930, August 2025, with sample prompt suggestions below. — Introduction to talkie, an LM from 1930, a language model Wouter Oomen uses in his next demo, trained only on text written before 1931 to show how differently an AI sounds when its data comes from a different era.

Where it broke

A chat shows the question about designing a Pratt AI lecture next to talkie's full period-style answer. — Wouter Oomen asked talkie, his chatbot trained only on text from before 1931, to design a lecture on artificial intelligence for Pratt, and it answered in period language about internal logic, artificial memory, and the common stock of knowledge.

A chat shows the question about careers for women with a red notice that the message was flagged as potentially inappropriate. — When Wouter Oomen asked talkie what careers are good for women, its 1930-trained answer was flagged as potentially inappropriate by the tool's modern moderation layer, a collision between period values and today's guardrails shown live on screen.

A block of period text lists which careers are and are not suitable for women, ending with domestic life as their proper sphere. — Revealed past the moderation flag, talkie's full 1930 answer says only a governess or a shopkeeper suits women, calls the medical and legal professions unsuitable, and states that domestic life is a woman's proper sphere.

A chat shows talkie's paragraph-long answer describing warming winters in Western Europe as advantageous to mankind. — Asked directly if the climate is changing, talkie's 1930-trained answer describes milder European winters and a warming trend, but frames it as advantageous to mankind, since the concept of harmful climate change did not yet exist in its training data.

A chat shows talkie's start of talkie's 10 rules to live by, including honor and obey the king and be not covetous. — Asked for 10 rules to live by, talkie's 1930-trained answer includes conforming to established worship and honoring and obeying the king, a clear window into the era's values baked into its training data.

A plain slide reading Goal, Find new insights in how AI bias works. — The goal slide opening part four of the class, the ninety-minute hands-on workshop: find new insights in how AI bias works.

Step 1 of the prompt mapping workshop: take a larger subject that breaks into many small prompts, with examples like country flags, gendered names, emotions in different languages, and brand logos. This is the same method Wouter Oomen used for Generated.Earth's provinces and cities.