Lens 3: Lifecycle and Labor — Materiality and Extraction
Who does the work, carries the risk, or has no recourse?
- Slot: 15 minutes in the morning: theory and one example.
- Instructor: Moritz Mähr.
- Guiding questions: four questions for the Lens 3 group. See Activity.
- Handout: the Research Use Card on Sheet 1.
- Slides: Lens 3 slide deck.
- Anchor sources: Crawford (2021), chapters 2 and 3; Bender et al. (2021).
The case describes workers who read texts about violence and sexual abuse. The page describes their working conditions. It quotes none of the texts.
Question
Who does the work, carries the risk, or has no recourse?
A generative AI system looks automatic. Behind it, people collect and label data, rate answers, moderate content, maintain hardware, and write the texts and images that the system learned from. This lens follows the lifecycle of a model and asks who works at each stage, who carries the risk, and who can object.
Core concepts
1. The lifecycle has six stages
The course uses six stages: design → data → training → deployment → evaluation → governance. This framework builds on Rachel Huber and Moritz Mähr’s Decoding Inequality course, published in 2024 and updated in 2025 (Huber and Mähr 2024). Its original labels are architecture selection, data collection, training, application, evaluation and monitoring, and governance and regulation. The course uses shorter labels. The order is a convention of the course. Evaluation happens before and after deployment, and governance applies at every stage. Each stage involves different people and different risks. Suresh and Guttag show that harm can enter at many points of the machine learning lifecycle, from the collection of data to the deployment of a model (2021). Weidinger et al. sort the risks of language models into six areas, from discrimination and misinformation to environmental and socioeconomic harms (2022).
3. Extraction concerns people who never agreed
Training data come from the work of writers, artists, photographers, and translators. Their consent was rarely asked. Crawford describes the collection of data at this scale as a form of extraction, next to the extraction of minerals and labor (2021, chs. 2–3). Copyright, consent, and provenance are therefore questions of labor too.
4. Recourse is unevenly distributed
A person who is harmed by an output can often not find out which data, which worker, or which decision caused it. Bender et al. warn that undocumented training data “perpetuates harm without recourse” (Bender et al. 2021, 615).
Evidence and limits
- Investigative journalism and interview studies document the conditions of outsourced data work (Perrigo 2023; Miceli and Posada 2022; Gray and Suri 2019).
- Developers’ own papers document that human feedback is part of the training (Ouyang et al. 2022).
- Most contracts between AI companies and suppliers are confidential. Single cases come from leaks, investigations, or interviews.
- Working conditions differ across suppliers, countries, and years. One case does not describe the whole industry.
- The companies dispute some details of the reported cases. The case below states where they do.
Case: safety labeling for ChatGPT in Kenya
This case comes from an investigation by TIME, based on interviews and documents (Perrigo 2023).
To make ChatGPT less toxic, OpenAI needed a detector for violent, hateful, and sexual content. The detector needed labeled examples. From late 2021, OpenAI contracted Sama, a San Francisco-based outsourcing firm, to label tens of thousands of text passages in Nairobi, Kenya (Perrigo 2023).
| Item | Reported fact |
|---|---|
| Task | Label text passages that describe violence, hate speech, and sexual abuse |
| Take-home wage | About USD 1.32 to 2 per hour, by seniority and performance |
| Workload | Workers reported 150 to 250 passages per nine-hour shift. Sama said 70. |
| Contracts | Three contracts worth about USD 200,000. OpenAI said it paid about USD 150,000, because the work ended early. |
| End of the work | Sama ended the work in February 2022, eight months early |
TIME interviewed four Sama employees who worked on the project. All four “described being mentally scarred by the work”. They had access to “wellness” counseling, but all four said that the sessions did not help (Perrigo 2023). OpenAI said that it works “hard to build safe and useful AI systems that limit bias and harmful content” (Perrigo 2023).
This case concerns safety labeling. The separate work of writing and ranking model answers for human-feedback training is documented by Ouyang et al. (Ouyang et al. 2022).
Why this case. The safety of a product for its users depended on work that exposed other people to harm. The users never saw this work. The case asks where the responsibility of a researcher who uses the product begins.
Worked matrix
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.
The fictional university team is considering ChatGPT. It uses the Kenya investigation as evidence about labeling and contracting (Perrigo 2023). Its judgments about risk and influence are a teaching analysis. Its interventions are PROPOSED. The report does not establish every route to complaint or remedy. Write Unknown from this report when evidence is missing. Do not infer that no route exists.
Activity
This Lens has no individual activity. After the four Lens inputs, each participant chooses the Lens that fits their Research Use best. The participants form Lens groups of four, ideally one group for each Lens. The Lens 3 group works for 60 minutes:
- Each person introduces their Research Use Card.
- The group applies the four guiding questions below to each Research Use.
- The group records insights, uncertainties, and disagreements on a flipchart or in a shared document.
- The group nominates a spokesperson for the plenary.
Guiding questions for Lens 3
- Which lifecycle stages does this Research Use touch: design, data, training, deployment, evaluation, or governance?
- Who does the hidden work at these stages, for example labeling, moderation, or the checking of outputs?
- Who carries the risk, and who has recourse when harm occurs?
- Whose work, data, or consent does the output depend on, and was it given?
Write Unknown where evidence is missing: an unknown is a finding. Describe the use, not the data. Do not show confidential, personal, or unpublished research data. If a Research Use cannot be discussed openly, use the Kenya case instead.
The questions of all four Lenses are on one shared slide of the course day deck.
Output
The flipchart or the shared document of the Lens 3 group: insights, uncertainties, and disagreements for each Research Use. In the plenary, the spokesperson reports the conceptual insights, not the individual Research Uses. The syllabus gives the three plenary questions.
Reading
Slides
Open the Lens 3 slide deck. Press S for the presenter notes.
Timing (15 minutes). The Kenya case, 5 minutes. The concepts and the worked matrix, 10 minutes. The Lens group applies the four guiding questions later in the morning.
Prompts.
- “Which work in your own research is invisible in the final publication?”
- “If an output of this system harms someone, whom can that person contact?”
- “Which stage of the lifecycle does your institution control through procurement or rules?”
Link to the afternoon. Keep unknown routes to recourse visible. Panarchy asks at which level someone could investigate or change them. An unknown route is not evidence that no route exists.
Difficult moments. Some participants work with outsourced labor themselves, for example through crowdsourcing platforms. Invite them to share practice, not to defend it. Point to fair-pay guidelines of their discipline where they exist.
Preparation. Show the completed matrix as the example of this Lens. The slide deck starts with the case. The lesson page keeps its standard section order. Its documentary image comes from Mathenge’s worker-led account (Mathenge 2024), a separate source from TIME’s interviews. It does not show the labeling task. Do not play the film without reviewing its content note.
Lens group. Circulate during the group work. If a group gets stuck, ask one guiding question at a time. If a Research Use cannot be discussed openly, offer the case of this Lens.