The course day

All slides of the course day, in order

Laura Wagner and Moritz Mähr

How to use this deck

This deck holds all slides of the course day, in the order of the timetable. Use it on the room screen for the whole day.

• Each slot starts with a title slide. It names the activity, the time, the duration, the format, and the lead.
• The slides of a slot follow the title slide.
• Press O for the overview. Press ? for all keyboard shortcuts.

The times come from data/schedule.yml. Show the other decks in a second window during the Lens inputs.

Open the deck before the participants arrive. Press F for full screen.

License

These slides are licensed under CC BY 4.0. You can copy, change, and teach them if you give credit.

Source and handouts: maehr.github.io/stochastic-parrots-and-beyond

Press S for the presenter notes and ? for the keyboard shortcuts.

Open with the license slide only briefly. It tells participants that they may keep and reuse everything. Then move to the question.

License and attribution

The structures and illustrations used in this session are adapted under Creative Commons Attribution-ShareAlike 4.0 from the original work of Keith McCandless, Henri Lipmanowicz and Nancy White found at www.liberatingstructures.com and the Liberating Structures Fieldbook (2026). Licensed under CC BY-SA 4.0.

These slides are licensed under CC BY-SA 4.0. The diagrams are original drawings. The deck uses no Liberating Structures icons.

Read the attribution briefly. It is a license condition, and it tells participants where to learn more.

Welcome

09:00–09:10 · 10 min · Plenary · Lead: Laura

How can we make, and defend, a responsible decision about generative AI in our own research?

Our question

How can we make — and defend — a responsible decision about generative AI in our own research?

• Not a list of permitted and forbidden tools.
• A transparent position that you can explain to colleagues, reviewers, and the people your research affects.

Ask: “Who has used a generative AI system for research this month?” Do not ask for details yet. Transition: “Before we judge a use, we need ways of looking at it. We use four lenses.”

Four lenses

1. Scale and environment. What does a model cost, and where do its data come from?
2. Representation and Displacement. Whose perspectives become normal?
3. Lifecycle and labor. Who does the work, and who carries the risk?
4. Political economy. Which incentives keep a practice in place?

In the afternoon, Panarchy connects the lenses to the levels where change is possible.

Name each lens with its question. Point out that the lenses overlap on purpose. A single use can raise questions in all four.

One lifecycle, six stages

Original diagram. CC BY 4.0.

Read the stages from left to right. Say that we will mark risks at each stage. Transition: “Let us start with one hypothetical example and look at it through all four lenses.”

A short case

Hypothetical example

A doctoral candidate in history wants to use a hosted chatbot to summarize 200 digitized newspaper articles for a literature review.

Lens	First question
Scale and environment	What is known about the model’s training data and energy use?
Representation	Whose voices does the summary make typical, and whose does it drop?
Lifecycle and labor	Who labeled the data, and who checks the summaries?
Political economy	What does the provider gain, and what does the university pay?

This example is invented for teaching.

Ask the group to add one question per lens. The Lens groups later use four guiding questions for each Lens.

Concepts we use all day

• Stochastic parrot: Bender et al.’s term for a system that stitches together linguistic forms from its training data “without any reference to meaning” (2021, 617). Whether such systems understand is debated (Mitchell and Krakauer 2023).
• Documented vs. hypothetical: we mark every example. A product name is not evidence.
• Known, unknown, avoidable: three columns for every cost.
• Levels: a change in a project, a department, a university, an industry, or a law.

Stress the difference between documented and hypothetical. Ask participants to challenge any claim that has no source. Keep four questions apart: how a system is trained, what it can do for a task, whether it understands, and whether a use is justified. The course needs no answer to the third question to ask the fourth.

The day

1. Four Lenses, 15 minutes each: theory and one example.
2. Lens groups: apply the four questions of one Lens to each Research Use.
3. Spokespersons report to the plenary.
4. Panarchy Block 1: your challenge, your Panarchy Chart, and a group Action Plan.
5. Panarchy Block 2: the open challenges in pairs.
6. Your individual Action Plan, peer review, and the course evaluation.

The edition page gives the times.

Do not give times on this slide. Each participant leaves with an individual Action Plan. It prepares the follow-up task: the Responsible GenAI Use Statement.

Research Use Card

09:10–09:20 · 10 min · Alone · Lead: Moritz

Fill in the Research Use Card on Sheet 1, front, right. Describe one actual or planned use of generative AI in your research. Leave the field “Lens” empty for now. Do not write confidential or personal data.

Your Research Use Card · 10 minutes

Alone. Take Sheet 1. On the right, fill in the Research Use Card:

• the task: one actual or planned use of generative AI in your research,
• the system or tool,
• why you might use it,
• what you do not know about it,
• who else is affected.

This is your Research Use. It stays your Research Use all day. Leave the field “Lens” empty for now.

Do not write confidential or personal data on the card.

Keep strict time. The cards stay with the participants. After the four Lens inputs, each participant writes the Lens that fits their Research Use best on the card. The Lens groups form after the break.

Lens 1: Scale & Environment

09:20–09:35 · 15 min · Plenary · Lead: Moritz

Theory and one example. Ask yourself: does this Lens fit my Research Use?

What did BLOOM’s training cost?

Documented case

Which costs disappear when we count only the training power?

Estimates for final training: Luccioni et al. (2023), Table 3. Original chart, CC BY 4.0.

Start with the chart. Participants have read Bender et al.; do not start with another summary of the paper. Ask them to identify the parts left out of the first bar. The expanded boundary includes 14.6 t from idle infrastructure and 11.2 t from hardware manufacture. Rounded values sum to 50.49 t; the paper reports 50.5 t. These are estimates built from measured energy and accounting assumptions, not a complete footprint for every use of BLOOM (Luccioni et al. 2023).

BLOOM has 176 billion parameters. It was trained on Jean Zay in France. Its ROOTS training corpus has 46 natural and 13 programming languages (Laurençon et al. 2022; Luccioni et al. 2023). The example makes Bender et al.’s environmental costs concrete (Bender et al. 2021, sec. 3).

The model runs somewhere

Hardware · electricity · cooling · location

Site context photographed in 2026: Clemcloum8, CC0 (Clemcloum8 2026). The panel is not evidence for BLOOM’s training footprint.

The panel makes the physical infrastructure visible. It was photographed after BLOOM’s training. Do not subtract its heat-recovery claim from the 2023 paper’s estimates. The photo does not establish the hardware configuration or heating arrangements during BLOOM’s training.

The learning question is: Which environmental and data choices are hidden by a model’s apparent performance? Bender et al. connect environmental costs with unequal distributions of benefit and harm (Bender et al. 2021, sec. 3). Crawford follows this material base through mines, grids, and supply chains (Crawford 2021, ch. 1).

Optional background belongs here, not on the projected slide: the IEA’s data-center totals are sector-wide estimates and projections (International Energy Agency 2025). The water figure in Li et al. is an estimate with explicit assumptions (Li et al. 2025). Neither provides a per-request measurement for the group’s own tool.

Training data: who survives the filter?

Hypothetical illustration of exclusion. Concept: Bender et al. (2021), sec. 4. Original diagram, CC BY 4.0.

Ask which page the rule loses. The pages and the specific rule are invented for teaching. Do not present them as a sample from C4 or BLOOM.

Then connect the example to the documented argument. Bender et al. discuss how C4’s blocklist filtering excludes LGBTQ-related text and how web-scale collection overrepresents dominant viewpoints (Bender et al. 2021, sec. 4). A larger dataset does not establish broader representation. The choice of a filter is a value choice. The filtering example is a separate illustration; it does not describe BLOOM’s data pipeline.

Documentation debt: what can you trace?

Source

Where did the training text come from?

Filter

What was excluded, and by whom?

Record

Can someone check those decisions?

An undocumented choice becomes harder to question.

Concept: Bender et al. (2021), sec. 4.

Recall documentation debt from the assigned paper: collecting more data than one can document makes later accountability harder (Bender et al. 2021, 615). Use the fictional filtering example: if the filter’s rule is missing, how would the excluded group identify what happened?

BLOOM’s open reporting illustrates what can be recorded. ROOTS and the BigScience governance work document data sources and decisions (Laurençon et al. 2022; Jernite et al. 2022). This does not make BLOOM cost-free or settle consent for all data. Ask participants which records exist for the tool they use.

A training footprint does not price your task

Final training

50.5 t CO₂eq

One accounting boundary.

One deployment

914 kWh over 18 days

One observed pattern of use.

For your task: what still needs measurement?

Separate studies within Luccioni et al. (2023). The values have different units and boundaries.

Do not add the two numbers. The 18-day API deployment served 230,768 requests; its hardware and conditions differ from final training. The study reports that keeping the model ready accounts for much of deployment energy (Luccioni et al. 2023, sec. 4.4). It does not give the energy for a different research workflow.

Luccioni et al. compare generative and task-specific systems for several tasks (Luccioni et al. 2024). A search tool, manual work, or a smaller model is an alternative to test at comparable quality. Do not promise a saving without checking the task and the boundary.

Worked map: two stages are enough

Hypothetical worked analysis

System: BLOOM. Selected stages: Training; Deployment and use.

Stage	Known	Unknown	Avoidable (PROPOSED)
Training	Final training: 50.5 t CO₂eq	Cost of training for our task	Test an existing model before new training
Deployment and use	One 18-day deployment: 914 kWh	Energy for our intended use	Compare fewer requests and a task-specific tool

Largest unknown: Energy for our task. One avoidable cost: A new training run, if an existing model meets the need.

Plenary sentence: “Performance hides the cost of our intended use.”

Figures: Luccioni et al. (2023). The research team is fictional. Actions are proposals with unmeasured savings. Other stages: Not examined.

Show every filled field. This is a fictional research team considering BLOOM, using documented measurements. The known figures come from the paper. The unknowns concern the team’s intended use. The actions are proposed choices, not measured savings.

Mark all other stages Not examined. Explain that a blank cell is ambiguous; write Unknown when the evidence is missing. The Lens 1 group can use this prepared case if a Research Use cannot be discussed openly.

Takeaway

A missing number is an unknown, not a zero.

Ask about environmental costs, training data, and documentation debt.

Return to Bender et al. (2021), secs. 3–4.

Keep the lens slot at 15 minutes: case first, 5 minutes; concepts and worked map, 10 minutes. The Lens 1 group applies the four guiding questions later in the morning. Transition to Lens 2: We followed data as a cost. Next we follow data as a picture of the world.

Sources

• Bender et al. (2021): environmental costs, training data, documentation debt.
• Luccioni, Viguier, and Ligozat (2023): BLOOM footprint and deployment study.
• Laurençon et al. (2022); Jernite et al. (2022): ROOTS and data governance.
• Luccioni, Jernite, and Strubell (2024): comparing deployment tasks.
• Photo: Clemcloum8 (2026), CC0. Original diagrams: CC BY 4.0.

Full references and evidence limits

Lens 2: Representation

09:35–09:50 · 15 min · Plenary · Lead: Laura

Theory and one example. Ask yourself: does this Lens fit my Research Use?

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.”

Lens 3: Lifecycle & Labor

09:50–10:05 · 15 min · Plenary · Lead: Moritz

Theory and one example. Ask yourself: does this Lens fit my Research Use?

Whose work makes ChatGPT safer?

Documented case · Kenya safety labeling

Who reads the text that users are protected from?

Case: Perrigo (2023). Content note: workers read texts about violence and sexual abuse. No examples are shown.

Start with the work, not a lifecycle definition. The source is TIME’s investigation from January 2023, based on interviews and documents. Workers labeled harmful passages for a toxicity detector. This is not the same task as writing and ranking answers for human-feedback training (Perrigo 2023; Ouyang et al. 2022).

Ask participants to follow the work across the companies and workers in the diagram. The learning question is: Who does the work, carries the risk, or has no recourse?

Workers can speak and organize

A worker-led account makes hidden labor visible.

Contextual still: Mathenge (2024); videography and editing: P. Teka, CC BY 4.0. This is a separate source from TIME’s interviews.

The image is the published preview of the documentary Data Workers Organizing: The African Content Moderators Union. It is a source about organizing and working conditions at Sama, not a photograph of the labeling task in the 2021–2022 contract. Do not identify the speaker as one of TIME’s anonymous interviewees. The film includes sensitive testimony; use this safe still without playing the film in class (Mathenge 2024). It is a figure source, not a new assigned reading.

The later organizing account shows why an unknown complaint route in a report is not evidence that workers lack agency. Distinguish worker-led action from what a researcher can change through an institution.

One safety intervention, unevenly carried costs

Reported pay

USD 1.32–2 per hour

Reported take-home pay.

Reported harm

Four interviewees

All reported lasting distress.

Counseling was reported. Access and benefit were disputed.

Reported facts: Perrigo (2023). Conditions vary across projects.

TIME reviewed documents and interviewed four workers. This is a documented case, not a representative wage survey. Sama disputed the wage range and reported a higher range for some roles. Workers reported heavier workloads than Sama stated. Workers described counseling as ineffective or hard to access; Sama said both group and individual counseling were available. OpenAI said Sama managed pay and mental-health provisions (Perrigo 2023). Keep these disputes explicit; do not turn the four interviews into an industry-wide rate.

Use the lesson page for the detailed case table. The important question for this slide is whose safety is improved and who bears the cost of that work.

From documentation debt to harm without recourse

Decision

Who set the data and work rules?

Record

Can affected people find the decision?

Recourse

Who can challenge it and seek remedy?

What would a worker or an affected person need to question the decision?

Conceptual bridge: Bender et al. (2021), sec. 4. Case analysis: Perrigo (2023).

Bender et al. connect undocumented training data with harm without recourse (Bender et al. 2021, 615). Apply this as a question to the Kenya case. The link is our analysis, not a claim that their paper investigated Sama.

Explain recourse as a route to challenge a decision and seek remedy. Access to counseling alone does not establish such a route. If the report does not establish an effective complaint route, write Unknown from this report. Do not write None.

Introduce extra terminology only if useful: Gray and Suri call invisible on-demand labor ghost work (Gray and Suri 2019). Taylor describes the presentation of human work as automation (Taylor 2018). Crawford discusses labor and data extraction (Crawford 2021, chs. 2–3). These sources support the notes; participants do not need to have read them.

Follow the work across six stages

Framework: Huber and Mähr (2024). Diagram: Wagner and Mähr (2026), CC BY 4.0. Lifecycle harms: Suresh and Guttag (2021).

The course framework builds on Rachel Huber and Moritz Mähr’s Decoding Inequality course, published in 2024 and updated in 2025 (Huber and Mähr 2024). The current drawing is by Laura Wagner and Moritz Mähr, 2026. Suresh and Guttag research sources of lifecycle harm; they did not author these six course stages (Suresh and Guttag 2021).

The labels are a course convention. Evaluation happens before and after deployment. Governance applies throughout. Ask where the Kenya work sits: labeling at Data; commissioning and supplier safeguards at Governance.

The original source uses architecture selection, data collection, training, application, evaluation and monitoring, governance and regulation. The course uses shorter labels for analysis. Do not confuse this teaching framework with one universal technical pipeline.

Worked matrix: data and governance

Hypothetical worked analysis

System: ChatGPT. Selected stages: Data; Governance.

Question	Data	Governance
Who does the work?	Sama text labelers	OpenAI commissions; Sama manages
Who carries the risk?	Workers exposed to harmful text	Workers relying on supplier safeguards
Who has recourse?	Effective remedy: Unknown; counseling disputed	Effective complaint route: Unknown from this report
Possible intervention (PROPOSED)	Require support designed with workers	Request labor safeguards through university procurement

Highest risk among selected stages: Data. Most influence for our fictional university team: Governance. One intervention: Request evidence of worker support through procurement.

Case: Perrigo (2023). Matrix: a fictional team’s analysis. Interventions are proposals. Other stages: Not examined.

Read the completed matrix as the example of this Lens. A fictional university team is considering ChatGPT and has read the Kenya report. The report supplies the labeling and contracting facts. The team’s identification of risk, influence, and interventions is a hypothetical analysis.

The team can request evidence through institutional procurement; it does not directly manage Sama workers. Counseling is a reported safeguard, not proof of effective recourse. Mark all four unselected stages Not examined. Explain that highest risk means highest among the stages examined.

Takeaway

Find the people behind the system. Ask who can question its decisions.

Training data · documentation debt · harm without recourse

Return to Bender et al. (2021), sec. 4.

Keep the lens slot at 15 minutes: case first, 5 minutes; concepts and worked matrix, 10 minutes. The Lens 3 group applies the four guiding questions later in the morning. Transition to Lens 4: Why do these patterns persist? The next lens follows money, infrastructure, and rules.

Sources

• Perrigo (2023): Kenya safety-labeling investigation.
• Bender et al. (2021): documentation debt and recourse.
• Huber and Mähr (2024): conceptual source of the six-stage framework.
• Suresh and Guttag (2021): research on sources of lifecycle harm.
• Mathenge (2024): documentary still, CC BY 4.0. Videography and editing: P. Teka.
• Original diagrams: Wagner and Mähr (2026), CC BY 4.0.

Full references and evidence limits

Lens 4: Political Economy & Governance

10:05–10:20 · 15 min · Plenary · Lead: Laura

Theory and one example. Then write on your Research Use Card the Lens that fits your Research Use best.

Learning question

Which incentives preserve an unsafe or inequitable practice?

The first three lenses showed costs, harms, and hidden labor. This lens asks why they persist.

Slides: CC BY 4.0. Lesson page: maehr.github.io/stochastic-parrots-and-beyond

Ask: “Who decided which AI tool you can use at your institution?” Collect answers. Most decisions were made at a level above the researcher.

1. Concentrated resources

• AI is made of models, data, labor, development frameworks, and computational power.
• A few companies control most of these resources.
• “Openness alone does not perturb the concentration of power in AI.”

Widder, Whittaker, and West (2024), Nature.

Point out that this is a material analysis: it starts from what AI is made of, not from what companies say about it.

2. “Open” is a claim to check

• More than 45 generative AI systems, 14 dimensions of openness.
• Many “open” systems are “open weight” at best.
• Many providers withhold information on training and fine-tuning data.
• Openness is composite and comes in degrees.

Liesenfeld and Dingemanse (2024), FAccT.

Ask: “What would you need to inspect, document, and cite a model in a publication?” Weights alone are rarely enough.

3. Models encode the goals of their makers

• Every model simplifies the world according to the priorities of its makers (O’Neil 2016, ch. 1).
• A system optimized for engagement, market share, or a fast release can conflict with the goals of research.

4. Institutions set defaults

• Universities decide which tools to license, which to recommend, and which data may leave the institution.
• These defaults shape practice more than individual intentions.
• Procurement and policy are places of leverage, not only of compliance.

This concept prepares Panarchy: the university is one of the levels.

5. Regulation: the EU AI Act as one example

From 2 August 2025, providers of general-purpose AI models must, under Article 53:

• keep technical documentation,
• inform providers who build on the model,
• have a policy to comply with EU copyright law,
• publish “a sufficiently detailed summary about the content used for training”.

Free and open-source models with public weights are exempt from the first two duties, unless they have systemic risk.

Fines are possible from 2 August 2026. Models placed on the market before 2 August 2025 have until 2 August 2027.

Regulation (EU) 2024/1689, arts. 53, 101, 111(3), and 113. Not legal advice.

Do not answer whether the Act applies to a specific project. Note such questions for the legal service.

Case: an “open” license

Documented case

Question	Llama 2 license (2023)	Open Source AI Definition 1.0 (2024)
Any purpose, without permission?	No: 700 million user threshold, use policy	Required
Outputs may improve other models?	Not other large language models	No such restriction
Weights available?	Yes, under the license	Required
Detailed information on training data?	Not required	Required

Sources: Meta (2023); Open Source Initiative (2024); Maffulli (2023).

The OSI stated in July 2023 that the Llama 2 license is not open source. Ask: “Why does the word matter?” Reputation, and in the EU fewer regulatory duties.

Who gains, who pays?

Original diagram. CC BY 4.0.

Read one arrow of each kind: money, data, rules. Ask which arrow is missing for the research participants.

Sources

• Widder, West, and Whittaker (2023): Open (For Business); published version in Nature (2024).
• Liesenfeld and Dingemanse (2024): Rethinking Open Source Generative AI.
• Open Source Initiative (2024): The Open Source AI Definition 1.0.
• Regulation (EU) 2024/1689, the AI Act.

Takeaway

A practice persists when someone gains from it. Find the incentive, then find the level where it can change.

Transition to the break: “Now choose the Lens that fits your Research Use best. After the break, the Lens groups apply its four questions.”

Break

10:20–10:35 · 15 min

Break. We continue with the Lens groups.

Form Lens groups

10:35–10:40 · 5 min · Plenary · Lead: Laura

Go to the table of the Lens that fits your Research Use best. Form groups of four. If a table has too many people, some move to their second choice.

Research Use and Lens discussion

10:40–11:40 · 60 min · Groups · Lead: Moritz

Introduce your Research Use Card. Apply the four questions of your Lens to each Research Use. Record insights, uncertainties, and disagreements. Nominate a spokesperson.

Lens 1: Scale & Environment

1. What is known, unknown, and avoidable about the energy, water, and hardware that this Research Use needs?
2. Where do the training data come from, and what did the filters remove?
3. Where does the system run for this task, and how often do you use it?
4. Which smaller or non-generative alternative could do the same task at a comparable quality?

Lens 2: Representation

1. Whose categories, languages, and aesthetics does the system treat as normal?
2. Who decided what counts as “quality” in the data, and what did that decision exclude?
3. Which stereotypes or representational harms can enter the output of this Research Use?
4. Who is depicted or displaced without consent, and who can object?

Lens 3: Lifecycle & Labor

1. Which lifecycle stages does this Research Use touch: design, data, training, deployment, evaluation, or governance?
2. Who does the hidden work at these stages, for example labeling, moderation, or the checking of outputs?
3. Who carries the risk, and who has recourse when harm occurs?
4. Whose work, data, or consent does the output depend on, and was it given?

Lens 4: Political Economy & Governance

1. Who gains from this Research Use, and who pays or carries the risk?
2. Which claims of openness or safety must you check, and how?
3. Which institution sets the default: a license, a procurement decision, or a policy?
4. Which rule applies, for example the EU AI Act, and which governance change would matter most?

Spokesperson reports

11:40–12:00 · 20 min · Plenary · Lead: Laura

Five minutes for each group. Report conceptual insights, not individual Research Uses.

1. Which concept or distinction was difficult to apply consistently, and what needs clarification?
2. Where did our Lens lead to conflicting interpretations, and what explains the disagreement?
3. Which unresolved question most affects how we should apply this Lens in practice?

Clarify open questions

12:00–12:10 · 10 min · Plenary · Lead: Moritz

We clarify the open questions from the reports.

Buffer / capture findings

12:10–12:20 · 10 min · Groups · Lead: Laura

Complete the flipchart or the shared document of your group. Photograph the flipchart only if the group agrees. Do not photograph people.

Lunch

12:20–13:20 · 60 min

Lunch. We continue with Panarchy.

Panarchy introduction

13:20–13:25 · 5 min · Plenary · Lead: Moritz

Small levels change fast. Large levels change slowly. The levels influence each other in both directions.

Learning question

What changes across interconnected levels would make generative AI use in academic research more responsible?

This is the common question of the afternoon. Each participant turns it into their own challenge: one sentence, starting from their Research Use Card.

1. Levels change at different speeds

• Small, local levels change fast: a project can change next week.
• Large levels change slowly: a university, an industry, or a law changes over years.
• Panarchy helps a group “identify opportunities and obstacles for spreading ideas at multiple levels”.

Liberating Structures, Panarchy; Holling (2001) on the adaptive cycle.

Use the image of nested dolls: zoom in, zoom out.

2. Influence runs both ways

• Ideas and experiments can spread upwards.
• Resources, rules, and memory flow downwards.
• A fast level often waits for a slow one. A slow level often ignores a fast one.

3. Discover your own levels

An illustration, not a template:

individual research project → research group or department → university and research sector → model vendors and industry → regulation and society

Say clearly: each participant finds their own levels. Some add funders, publishers, or disciplines.

The Panarchy Chart

Sheet 1, back. Original diagram, adapted from Liberating Structures Panarchy. CC BY-SA 4.0.

A short case

Hypothetical example

A department wants researchers to disclose their use of generative AI.

• Project level: researchers have no template and do not know what to write.
• Department level: a template would take one afternoon, but nobody owns it.
• University level: the licensed chatbot has no documented training data, so a disclosure stays vague.
• Vendor level: the vendor gains from vagueness.

Leverage: the department writes the template (fast), and asks the university to add documentation to the next procurement (slow).

This example is invented for teaching.

Frame Challenge & Identify Influences

13:25–13:40 · 15 min · Alone · Lead: Laura

Turn Sheet 1 over. Start from your Research Use Card. Write your challenge in one sentence in the header of the Panarchy Chart. List the influences on your challenge in the field “Influences”.

Block 1 · before the break

Minutes	Format	What happens
15	Alone	Frame the challenge from your Research Use Card. List the influences.
20	Pairs	Translate the influences into levels. Compare your levels.
5	Plenary	Form new groups of four.
15	Groups	Present your challenge and your Panarchy Chart.
5	Groups	Select one challenge.
10	Alone	Identify Opportunities & Barriers for the selected challenge. Sheet 2, part A.
25	Groups	Discuss, prioritize, and consolidate into a group Action Plan.
10	Groups	Buffer: complete the group Action Plan.

Steps, labels, and durations adapted from Liberating Structures Panarchy.

The facilitator guide gives the spoken prompt for each step. Before a group prioritizes, record each objection about consent, rights, or a burden on one group. A majority does not settle it. Keep all Panarchy Charts for Block 2.

Translate & Compare Levels

13:40–14:00 · 20 min · Pairs · Lead: Moritz

Sort your influences into levels on your Panarchy Chart. Name each level. Compare your levels with the levels of your partner.

Form groups of four

14:00–14:05 · 5 min · Plenary · Lead: Laura

Form new groups of four. Choose people who were not in your Lens group.

Present challenges and charts

14:05–14:20 · 15 min · Groups · Lead: Moritz

Each person presents their challenge and their Panarchy Chart. Keep to about three minutes each.

Select challenge

14:20–14:25 · 5 min · Groups · Lead: Laura

Select one challenge for deeper analysis. The other challenges stay open. They come back after the break.

Find Opportunities & Barriers

14:25–14:35 · 10 min · Alone · Lead: Moritz

On Sheet 2, front, part A: note Opportunities & Barriers for the selected challenge. Note the level of each.

Discuss and consolidate Action Plan

14:35–15:00 · 25 min · Groups · Lead: Laura

Discuss and prioritize the Opportunities & Barriers. Consolidate them into one group Action Plan on a flipchart or in a shared document.

Buffer / capture findings

15:00–15:10 · 10 min · Groups · Lead: Moritz

Complete the group Action Plan. Keep your Panarchy Chart. You need it after the break.

Break

15:10–15:25 · 15 min

Break. We continue with Panarchy Block 2.

Form new pairs

15:25–15:30 · 5 min · Plenary · Lead: Laura

Form a pair with a person from another group of four. Each pair has at least one open challenge. Two open challenges: examine both. One open challenge: both examine it.

Block 2 · after the break

Form new pairs across the groups of Block 1. Avoid a partner who already discussed your challenge.

• Two open challenges: examine both.
• One open challenge: both of you examine it.

1. Alone, 10 minutes. Opportunities & Barriers for the open challenges. Sheet 2, part B.
2. Pairs, 35 minutes. Discuss, prioritize, and record findings and possible actions.

Pair each participant whose challenge was selected with a participant whose challenge is open. Every challenge must receive an in-depth discussion in Block 1 or Block 2.

Opportunities & Barriers for remaining challenges

15:30–15:40 · 10 min · Alone · Lead: Moritz

On Sheet 2, front, part B: note Opportunities & Barriers for the open challenge or challenges of your pair. Use the existing Panarchy Charts.

Discuss and prioritize remaining challenges

15:40–16:15 · 35 min · Pairs · Lead: Laura

Discuss and prioritize. Record the findings and possible actions for each open challenge in part B.

Individual Action Plan

16:15–16:35 · 20 min · Alone · Lead: Moritz

On Sheet 2, back: complete the Action Plan for your own Research Use. Write the intended change, the next step, the barriers, the timeline, and the success indicator.

Your individual Action Plan

Alone, 20 minutes. Sheet 2, back. For your own Research Use:

• the intended change,
• the next step,
• the barriers,
• the timeline,
• the success indicator.

Pairs, 10 minutes. Peer review: is the next step concrete? Is the timeline realistic? Would the success indicator show the change?

Peer review

16:35–16:45 · 10 min · Pairs · Lead: Laura

Read the Action Plan of your partner. Is the next step concrete? Is the timeline realistic? Would the success indicator show the change? Note the feedback.

Course evaluation and closing

16:45–17:00 · 15 min · Plenary · Lead: Moritz

Complete the course evaluation. Your edition page gives the link. After the course day, write your Responsible GenAI Use Statement.

Expected output

• Your own Panarchy Chart with your challenge, influences, and levels.
• Opportunities & Barriers for your challenge, from Block 1 or Block 2.
• A group Action Plan for each group of four.
• Your individual Action Plan, reviewed by a peer.

Sources

• Liberating Structures: Panarchy.
• McCandless and White (2026): The Liberating Structures Fieldbook.
• Holling (2001): Understanding the Complexity of Economic, Ecological, and Social Systems.

Takeaway

Responsible use is not only a personal choice. Find the level where a change is feasible, and the allies at the levels above and below.

Slides: CC BY-SA 4.0, adapted from Liberating Structures.

Transition to the closing: “Take your Action Plan home. It is the start of your Responsible GenAI Use Statement.”

Sources

• Bender, Gebru, McMillan-Major, and Shmitchell (2021): On the Dangers of Stochastic Parrots.
• Mitchell and Krakauer (2023): The Debate over Understanding in AI’s Large Language Models.
• Course website: lenses, handouts, and literature.

Takeaway

We do not look for permission. We look for reasons that we can defend.

Stochastic Parrots and Beyond · Laura Wagner and Moritz Mähr · CC BY 4.0

Transition to Lens 1: “The first lens asks what a model’s performance hides.”