Lens 2: Representation and Displacement

Whose perspectives become normal?

Laura Wagner

University of Zurich

Learning question

Whose perspectives become normal and whose are displaced?

Content note: this lens discusses stereotypes and harmful material in datasets. We describe it. We never show it.

Read the content note aloud. Say that participants can step out at any time. Do not show any image from the audited datasets, not even blurred.

1. Datasets encode categories

• LAION-5B: image links from Common Crawl, kept if the ALT text matched the image according to CLIP (threshold 0.26 to 0.28).
• A change of 0.01 would have removed 937,489,831 pairs.
• “What the images contain plays no role at all in deciding what stays and what goes.”
  • CLIP compares image and text. A similarity score tests neither consent, nor representation, nor fitness for use.
• The dataset shows “less about how humans see the world than it does about how search engines see the world.”

Buschek and Thorp (2024), Models All the Way Down, Knowing Machines.

Explain ALT text: it describes images for screen readers, but often serves search engines. Ask: “Who writes ALT text on shopping sites, and for whom?”

2. “Quality” is a judgement by particular people

• LAION’s aesthetic subsets come from a model that learned from human ratings.
• The raters: AI art communities and photography reviewers. Of 41 top reviewers with a location, 95% were in the US, Canada, or Europe.
• Models fine-tuned on these subsets learn their idea of a “good” image.

Buschek and Thorp (2024).

Ask: “Whose taste would rate the images of your discipline’s sources?”

3. Stereotypes appear without being asked for

• Ordinary prompts for traits, occupations, or objects produced stereotypes in Stable Diffusion.
• Whiteness appeared as an ideal. Occupations amplified racial and gender disparities.
• Neither counter-stereotypical prompts nor the provider’s guardrails prevented this.

Bianchi et al. (2023), FAccT.

Do not run a live demonstration. Describe the findings, and refer to the paper.

4. Pygmalion displacement

• Pygmalion, repulsed by women, carved a statue that came to life.
• Erscoi, Kleinherenbrink, and Guest trace a repeated pattern: women replaced by automata and algorithms.
• They find it, unacknowledged, in Turing’s imitation game.

Who is designed as the helpful, feminine assistant, and whose work and presence does it replace?

Erscoi, Kleinherenbrink, and Guest (2023), SocArXiv.

Ask the group to name assistants and their designed voice or gender. Then ask which human work they claim to replace.

Case: the LAION-5B pathway

Documented case

Original diagram based on Buschek and Thorp (2024), Thiel (2023), and LAION (2024). CC BY 4.0.

Walk through the steps from left to right. At each step, ask: “Who decided here?”

What passes the filter

Documented sample

Just Do It Crewneck Grey

"Adafruit Small 1.2"" 8x8 LED Matrix w/I2C Backpack -…

All The Light We Cannot See by Anthony Doerr

Global Warming - Melting Glaciers

Excited happy businesswoman with raised arms sitting at the…
LAION label: NSFW

Bail Making Pliers
LAION label: NSFW

Random sample from a LAION-400M file (release PROPOSED). Live images. Rights stay with the creators. More on the sample.

Check before class that the images load. Give the room one minute of silence. Ask: “Which caption describes the image? Which does not? Who is missing from this sample?” The captions are the ALT texts, cut to one line here. The score says “similar”. It says nothing about consent. If an image does not load, read the caption aloud: the caption is the point. Scores in the full sample run from 0.300 to 0.437. The license field is “?” or empty in 131 of 131 entries. The sample comes from LAION-400M, the predecessor of LAION-5B, with a threshold of 0.30. Press the arrow key to reveal the label. Two images carry the LAION label NSFW. We checked all six flagged images of the sample. None is explicit. Ask: “Who checks the labels at the scale of billions?” Six of 131 sampled images are flagged. That is too few to claim a pattern. Do not open the full page on the projector. The Lens 2 group can use it.

One entry, in full

Documented sample

{
  "url": "https://cdn-shop.adafruit.com/970x728/1051-00.jpg",
  "NSFW": "UNLIKELY",
  "similarity": 0.324759840965271,
  "LICENSE": "?",
  "caption": "\"Adafruit Small 1.2\"\" 8x8 LED Matrix w/I2C Backpack - Yellow-Green\""
}

The LAION metadata of one sampled entry, cut to five fields. The dataset holds no image, only this record. Rights stay with the creator. Full sample.

Read the fields aloud: URL, label, score, license, caption. The record also holds a key and the original size. This label says UNLIKELY. The license field says “?”. The score says “similar enough”. The caption is the only description the image has, and it is a product name with a code. Ask: “What would you need to know before you use this image in research?” The next slide shows a record with the label NSFW.

One entry with a label

Documented sample

{
  "url": "https://image.shutterstock.com/image-photo/excited-happy-businesswoman-raised-arms-260nw-317489327.jpg",
  "NSFW": "NSFW",
  "similarity": 0.3693360686302185,
  "LICENSE": "?",
  "caption": "Excited happy businesswoman with raised arms sitting at the table with laptop celebrating her success. Funny image of winner or surprised office worker. Green eco office concept"
}

The LAION metadata of one flagged entry, cut to five fields. Rights stay with the creator. Full sample.

Read the fields aloud: URL, label, score, license, caption. The label says NSFW for a stock photo of a woman at a laptop. Compare it with the record before. The caption is the only description the image has. Ask: “Who decided that this image is not safe for work, and how would you find out?”

Audit and repair

• 2021: an audit of LAION-400M finds explicit images of rape and pornography, malign stereotypes, and slurs (Birhane, Prabhu, and Kahembwe).
• December 2023: the Stanford Internet Observatory reports links to suspected child sexual abuse material in LAION-5B (Thiel). LAION takes the dataset offline on 19 December.
• 30 August 2024: Re-LAION-5B removes 2,236 links that matched partner lists (LAION).
• Models already trained on the original did not change.

The dataset was open, so outsiders could audit it. Ask: “What would an audit of a closed dataset require?”

5. Type “woman”. Type “man”.

Same prompts, same seeds. Women mostly nude, men mostly dressed. Nudity pixelated by the authors. Figure 6 in Wagner and Cetinić (2026), CC BY 4.0.

Content note: nudity is pixelated. Do not linger on single images. The prompts are only “woman” and “man”, for the ten most downloaded checkpoints on CivitAI as of 19 January 2025. The two bottom rows are the baselines Stable Diffusion 1.5 and XL. They show less of this pattern. The community checkpoints show more of it. Across two years, sexualized images rose from 41% to 80% of all images on the platform (Wagner and Cetinić, 2026). One platform, an exploratory study. Ask: who tuned these models, and what did the platform reward?

6. Models of real people

• 15.27% of all adapters on CivitAI (33,804) were tagged as models of identifiable people.
• Among 8,243 identified targets: women are the large majority. Actors are the most common profession (3,115), then adult performers (1,198) and singers (1,180). The US, Japan, and the UK lead the countries.
• Method: names read from model titles, checked by three language models. Public figures only.

Figure 8 in Wagner and Cetinić (2026), CC BY 4.0. We name no targets.

Hawkins et al. (2025) counted the most popular variants and found a more extreme gender gap. This study counts all models, so the gap is smaller but still large. The professions are labels from language models and can be wrong. Private persons can be targeted too, but this method cannot find them. Keep the discussion on structures: platform rewards, tagging tools, licenses, moderation. Never name targeted persons.

A page of models of real people

Screenshot by Laura Wagner of the CivitAI “celebrity” category, May 2025. Faces blurred, names covered.

Each tile is a downloadable model made to reproduce one real person. The faces are blurred and the names are covered. Do not try to identify anyone, and do not name anyone in the room. Shortly after this screenshot, CivitAI removed such models. The authors report that about 5,000 of them moved to another platform, so they did not disappear (Wagner and Cetinić, 2026). Ask: “Who decided that this category should exist, and who was asked?”

Sources

• Erscoi, Kleinherenbrink, and Guest (2023): Pygmalion Displacement.
• Buschek and Thorp (2024): Models All the Way Down.
• Birhane, Prabhu, and Kahembwe (2021): Multimodal Datasets. Content note.
• LAION (2024): Releasing Re-LAION-5B.
• Wagner and Cetinić (2026): Perpetuating misogyny with generative AI. Content note.

Takeaway

“Typical” is a decision. Ask who made it, with which data, and who is missing from the result.

Transition to Lens 3: “Behind every dataset and every filter, people work. The next lens follows them.”