The Moral Exploitation

of Data Workers

​

July 9, 2026

Conny Knieling, University of Pittsburgh�Anthony Nguyen, Florida State University

Thesis

Many data workers are exploited distinctively for their moral agency

​

​

Outline

1. Terms and Conditions
2. An Example of Moral Exploitation
3. Moral Exploitation: What’s Distinctive About it?
4. Neo-Colonialism
5. Moral and Social Implications

AI Background (Terms and Conditions)

​

Ghost Work:

• Invisible labor and human workforce that powers AI
• Term coined by Gray & Suri (2019)�
• Most successful AI rely on humans-in-the-loop
• The work of these “ghost workers” is indispensable but �is rendered invisible

Some Examples of Ghost Work in AI

Data Annotation & Labeling

• Image Labeling: Workers draw boxes around objects (and often, people)
• Text Classification: Workers categorize sentiment or intent in millions of sentences

Content Moderation

• Social Media: Workers review flagged content

Reinforcement Learning via Human Feedback (RLHF)

• Human raters compare AI responses and rank which is "better"

​

Philosophy Background (Terms and Conditions)

Moral Agency: A person’s ability to make deliberate moral decisions

​

This involves:

• The capacity to make autonomous, deliberate moral decisions
• The capacity to make choices based on some notion of right or wrong
• The ability to act based on one’s ethical judgment and understanding of consequences�

Philosophy Background (Terms and Conditions)

Moral Labor: presupposes moral agency

Examples for moral labor: Parenting, Counseling, Soldiers …

​

Exploitation of Moral Labor: take unfair advantage of somebody’s moral labor, exploit their vulnerability to provide moral labor

Conditions for moral exploitation: No adequate compensation, support, nor recognition

​

Example 1: Single parent, raising children alone, no support or compensation

Example 2: ER psychiatrist, calling law enforcement

Philosophy Background (Terms and Conditions)

Moral Labor: presupposes moral agency

Examples for moral labor: Parenting, Counseling, Soldiers …

​

Exploitation of Moral Labor: take unfair advantage of somebody’s moral labor, exploit their vulnerability to provide moral labor

Conditions for moral exploitation: No adequate compensation, support, nor recognition

​

Example 1: Single parent, raising children alone, no support or compensation

Example 2: ER psychiatrist, calling law enforcement

Philosophy Background (Terms and Conditions)

Moral Labor: presupposes moral agency

Examples for moral labor: Parenting, Counseling, Soldiers …

​

Exploitation of Moral Labor: take unfair advantage of somebody’s moral labor, exploit their vulnerability to provide moral labor

Conditions for moral exploitation: No adequate compensation, support, nor recognition

​

Example 1: Single parent, raising children alone, no support or compensation

Example 2: ER psychiatrist, calling law enforcement

Examples of the Moral Exploitation of Ghost Workers

Moral Exploitation: exploit vulnerable workers to perform intense moral labor, severe harm through labor conditions and work itself (dehumanizing and traumatizing), absence of fair compensation, support, or recognition of harm

​

Ghost

Workers

​

​

​

​

Ghost Workers

performing

Moral Labor

Our Examples:

​

• NSFW Data Annotation
• Content Moderation

Example of Moral Exploitation

“AI Sweatshops” in Kenya to detox ChatGPT

Severe exploitation of workers:

• Emotionally taxing moral labor
• Inhumane working conditions (sweatshops conditions)
• Lack of transparency about work details
• Erasure of contributions to AI success
• Often: lasting psychological trauma with no adequate support

Moral Exploitation: exploit vulnerable workers to perform intense moral labor, severe harm through labor conditions and work itself (dehumanizing and traumatizing), no adequate compensation, support, or recognition of harm

​

AI Sweatshops

The labor conditions for these workers are so bad, that they have been said to work in “AI sweatshops” (Wasike 2025)

Exploitation: A Characterization

Exploitation

• Taking unfair advantage of someone’s vulnerability:�Excessive extraction & lack of reasonable alternatives�
• Intuitively, “to wrongly exploit someone is to…use the fact that his [sic] back is to the wall, so to speak, to get him to accept lopsided and outrageous terms of exchange” (Valdman 2009, 13) �
• Thought Experiment: $20,000 for an antidote that retails for $10 (Valdman 2009)�
• Real Example: LA Garment Industry (LA Times 2023)
• Paid per piece sewn, some paid as low as $1.58/hour
• Minimum wage in LA: $17.87/hour

​

Exploitation: A Characterization

Exploitation

• Taking unfair advantage of someone’s vulnerability:�Excessive extraction & lack of reasonable alternatives�
• Intuitively, “to wrongly exploit someone is to…use the fact that his [sic] back is to the wall, so to speak, to get him to accept lopsided and outrageous terms of exchange” (Valdman 2009, 13) �
• Thought Experiment: $20,000 for an antidote that retails for $10 (Valdman 2009)�
• Real Example: LA Garment Industry (LA Times 2023)
• Paid per piece sewn, some paid as low as $1.58/hour
• Minimum wage in LA: $17.87/hour

​

Exploitation: A Characterization

Exploitation

• Taking unfair advantage of someone’s vulnerability:�Excessive extraction & lack of reasonable alternatives�
• Intuitively, “to wrongly exploit someone is to…use the fact that his [sic] back is to the wall, so to speak, to get him to accept lopsided and outrageous terms of exchange” (Valdman 2009, 13) �
• Thought Experiment: $20,000 for an antidote that retails for $10 (Valdman 2009)�
• Real Example: Los Angeles Garment Industry (LA Times 2023)
• Paid per piece sewn, some paid as low as $1.58/hour
• Minimum wage in LA: $17.87/hour

​

Exploitation and Objectification

• Exploitation involves objectification—this is inappropriate treatment given the exploited individuals’ status as persons�
• The moral exploitation of ghost workers �involves three features of objectification:�
• Instrumentality,�
• Fungibility, and�
• Denial of subjectivity (Nussbaum 1995)

Exploitation and Objectification

• Exploitation involves objectification—this is inappropriate treatment given the exploited individuals’ status as persons�
• The moral exploitation of ghost workers �involves three features of objectification:�
• Instrumentality,�
• Fungibility, and�
• Denial of subjectivity (Nussbaum 1995)

Typical Cases of Exploitation

Moral Exploitation

• Labor cannot be replaced by AI (yet)�
• Exploits distinctive human capacity for moral decision-making

Material Exploitation

• Labor can be automated

​

• Ignores any human agency; treatment as mere means

​

• Lacks any recognition

The Distinctive Moral Exploitation of Ghost Workers

​

Material Exploitation

• Labor can be automated

​

• Ignores any human agency; treatment as mere means

​

• Lacks any recognition

Moral Exploitation

• Labor can’t be automated

​

• Exploits distinctively human moral capacity�
• Lacks full recognition

Data-Driven Moral Exploitation as (Neo-)Colonialism?

Colonialism

• Land grabs, other forms of theft and material exploitation (e.g., slavery)�
• Objectification without even superficial recognition of moral agency�
• Effectively, “what the colonist was saying to the colonized subject was: ‘Work yourself to death, but let me get rich!’” (Fanon 1963, 135)

Data-Driven Moral Exploitation as (Neo-)Colonialism?

Colonialism

• Land grabs, other forms of theft and material exploitation (e.g., slavery)�
• Objectification without even superficial recognition of moral agency�
• Effectively, “what the colonist was saying to the colonized subject was: ‘Work yourself to death, but let me get rich!’” (Fanon 1963, 135)

Neo-Colonialism

• Unrecognized, non-formal exploitation�
• Can involve moral exploitation, (presupposes moral agency)�
• “In essence, neocolonial society and colonial society do not differ in the least” �(Sankara 1983, 81)

Three Moral and Social Implications

Concerning…�

1. The prevalence of an overlooked form of exploitation,�
2. The relation between agency and objectification, and�
3. The way forward for moral AI development (if possible)

Implication 1: Overlooked but Widespread Exploitation

• Much data work is (paradigmatically!) exploitative in overlooked ways�
• AI development as we know it is materially exploitative, but also exploitative in its use of ghost workers’ moral agency�
• Existing critiques of ghost workers’ labor conditions tend to focus on material exploitation�
• Contemporary AI development is complicit in furthering (neo)colonial injustice

Implication 1: Overlooked but Widespread Exploitation

• Much data work is (paradigmatically!) exploitative in overlooked ways�
• AI development as we know it is materially exploitative, but also exploitative in its use of ghost workers’ moral agency�
• Existing critiques of ghost workers’ labor conditions tend to focus on material exploitation�
• Contemporary AI development is complicit in furthering (neo)colonial injustice

Implication 1: Overlooked but Widespread Exploitation

• Much data work is (paradigmatically!) exploitative in overlooked ways�
• AI development as we know it is materially exploitative, but also exploitative in its use of ghost workers’ moral agency�
• Existing critiques of ghost workers’ labor conditions tend to focus on material exploitation�
• Contemporary AI development enacts further (neo-)colonial injustice

Implication 2: Agency and Janus-Faced Objectification

• Exploitation is always objectifying, but moral exploitation complicates the relation between agency and objectification�
• Moral exploitation presupposes some implicit recognition of the exploited person’s moral agency, but a failure of recognition arises nonetheless�
• Further harm: Not only a lack of adequate compensation for morally taxing labor, but also use of agency against the agent�
• Exploited not despite the fact they’re because, but because they are!

Implication 2: Agency and Janus-Faced Objectification

​

• Exploitation is always objectifying, but moral exploitation complicates the relation between agency and objectification�
• Moral exploitation presupposes some implicit recognition of the exploited person’s moral agency, but a failure of recognition arises nonetheless�
• Further harm: Not only a lack of adequate compensation for morally taxing labor, but also use of agency against the agent�
• Exploited not despite the fact they’re because, but because they are!

Implication 2: Agency and Janus-Faced Objectification

​

• Exploitation is always objectifying, but moral exploitation complicates the relation between agency and objectification�
• Moral exploitation presupposes some implicit recognition of the exploited person’s moral agency, but a failure of recognition arises nonetheless�
• Further harm: Not only a lack of adequate compensation for morally taxing labor, but also use of agency against the agent�
• Exploited not despite the fact they’re persons, but because they are!�
• Provides further ethical reasons to improve ghost workers’ labor conditions

Implication 3: Need to Theorize About Fair Moral Labor

• We have a choice: radically reform distribution of moral labor used in AI development, or abolish it—at least, the work requiring traumatizing content�
• Given that AI often functions as public goods, should this moral labor be distributed more equally (e.g., maybe by lottery, as in jury duty)?�
• Or is this labor so traumatizing that it swamps the benefits of AI development, rendering the social practice unjustified in the first place?�
• Might the answer depend on the case/particular AI application?�
• Won’t settle these issues here
• Our main point: Our view raises these issues as further normative questions�

Implication 3: Need to Theorize About Fair Moral Labor

• We have a choice: radically reform distribution of moral labor used in AI development, or abolish it—at least, the work requiring traumatizing content�
• Given that AI often provides (what should be) public goods, should this moral labor be distributed more equally—e.g., by lottery, as in jury duty?�
• Or is this labor so traumatizing that it swamps the benefits of AI development, rendering such AI development unjustified in the first place?�
• Might the answer depend on the case/particular AI application?�
• Won’t settle these issues here
• Our main point: Our view raises these issues as further normative questions�

Implication 3: Need to Theorize About Fair Moral Labor

• We have a choice: radically reform distribution of moral labor used in AI development, or abolish it—at least, the work requiring traumatizing content�
• Given that AI often provides (what should be) public goods, should this moral labor be distributed more equally—e.g., by lottery, as in jury duty?�
• Or is this labor so traumatizing that it swamps the benefits of AI development, rendering it unjustified in the first place?�
• Might the answer depend on the case/particular AI application?�
• Won’t settle these issues here
• Our main point: Our view raises these issues as further normative questions�

Implication 3: Need to Theorize About Fair Moral Labor

• We have a choice: radically reform distribution of moral labor used in AI development, or abolish it—at least, the work requiring traumatizing content�
• Given that AI often provides (what should be) public goods, should this moral labor be distributed more equally—e.g., by lottery, as in jury duty?�
• Or is this labor so traumatizing that it swamps the benefits of AI development, rendering it unjustified in the first place?�
• Might the answer depend on the specific case or AI application?�
• Won’t settle these issues here
• Our main point: Our view raises these issues as further normative questions�

Implication 3: Need to Theorize About Fair Moral Labor

• We have a choice: radically reform distribution of moral labor used in AI development, or abolish it—at least, the work requiring traumatizing content�
• Given that AI often provides (what should be) public goods, should this moral labor be distributed more equally—e.g., by lottery, as in jury duty?�
• Or is this labor so traumatizing that it swamps the benefits of AI development, rendering it unjustified in the first place?�
• Might the answer depend on the specific case or AI application?�
• We won’t settle these pressing questions here
• Our Main Point: Our view raises these questions for future work in AI ethics

Three Moral and Social Implications Summarized

1. Neo-colonialism via the moral exploitation of ghost workers is prevalent;�
2. These ghost workers are (morally) exploited not despite the fact they are persons, but because they are persons who can make moral decisions; and�
3. Future work must address whether traumatizing moral labor in AI development should be reformed (e.g., distributed more equally) or if, instead, such moral labor should be abolished together

Three Moral and Social Implications Summarized

1. Neo-colonialism via the moral exploitation of ghost workers is prevalent;�
2. These workers are (morally) exploited not despite the fact they are persons, but because they are persons capable of moral decision-making; and�
3. Future work must address whether traumatizing moral labor in AI development should be reformed (e.g., distributed more equally) or if, instead, such moral labor should be abolished together

Three Moral and Social Implications Summarized

1. Neo-colonialism via the moral exploitation of ghost workers is prevalent;�
2. These workers are (morally) exploited not despite the fact they are persons, but because they are persons capable of moral decision-making; and�
3. Future work must address whether traumatizing moral labor in AI development should be reformed (e.g., greater pay and the work distributed more equally) or if, instead, such moral labor should be abolished altogether

Thesis

Many data workers are exploited distinctively for their moral agency

​

​

Thank you for your attention!

​

​

Appendix

Top AI Secrets that May Become Relevant in Q&A

Objection: Synthetic Data! �Just let the models construct all the data they will be trained on!

Reply: We Still Need Humans in the Loop!�

• Unless we accept significant drops in model accuracy, synthetic data (on its own) cannot completely eliminate the need for humans in the loop�
• Dhananjay Ashok & Jonathan May’s 2024 paper “A Little Human Data Goes a Long Way”�
• “Faced with an expensive human annotation process, creators of NLP systems increasingly turn to synthetic data generation. While this method shows promise, the extent to which synthetic data can replace human annotation is poorly understood. We investigate the use of synthetic data in Fact Verification (FV) and Evidence-based Question Answering (QA) by incrementally replacing human-generated data with synthetic points on eight diverse datasets. Strikingly, replacing up to 90% of the training data only marginally decreases performance, but replacing the final 10% leads to severe declines” (Ashok & May 2024)�
• Synthetic data may reduce total exploitation, but cannot by itself eliminate it!�

Objection: Synthetic Data!

Reply 1: We Still Need Humans-in-the-Loop to Get Model Accuracy�

• Unless we accept significant drops in model accuracy, synthetic data (on its own) cannot completely eliminate the need for human annotators�
• Dhananjay Ashok & Jonathan May’s 2024 “A Little Human Data Goes a Long Way”�
• “The extent to which synthetic data can replace human annotation is poorly understood. We investigate the use of synthetic data in Fact Verification (FV) and Evidence-based Question Answering (QA) by incrementally replacing human-generated data with synthetic points on eight diverse datasets. Strikingly, replacing up to 90% of the training data only marginally decreases performance, but replacing the final 10% leads to severe declines” (Ashok & May 2024)�
• Upshot: Synthetic data may reduce total exploitation, but cannot eliminate it!�

Objection: Synthetic Data!

Reply 1: We Still Need Humans-in-the-Loop to Get Model Accuracy�

• Unless we accept significant drops in model accuracy, synthetic data (on its own) cannot completely eliminate the need for human annotators�
• Dhananjay Ashok & Jonathan May’s 2024 “A Little Human Data Goes a Long Way”�
• “The extent to which synthetic data can replace human annotation is poorly understood…. We investigate the use of synthetic data in Fact Verification (FV) and Evidence-based Question Answering (QA) by incrementally replacing human-generated data with synthetic points on eight diverse datasets. Strikingly, replacing up to 90% of the training data only marginally decreases performance, but replacing the final 10% leads to severe declines” (Ashok & May 2024)�
• Upshot: Synthetic data may reduce total exploitation, but cannot eliminate it!�

Objection: Synthetic Data!

Reply 1: We Still Need Humans-in-the-Loop to Get Model Accuracy�

• Unless we accept significant drops in model accuracy, synthetic data (on its own) cannot completely eliminate the need for human annotators�
• Dhananjay Ashok & Jonathan May’s 2024 “A Little Human Data Goes a Long Way”�
• “The extent to which synthetic data can replace human annotation is poorly understood. We investigate the use of synthetic data in Fact Verification (FV) and Evidence-based Question Answering (QA) by incrementally replacing human-generated data with synthetic points on eight diverse datasets. Strikingly, replacing up to 90% of the training data only marginally decreases performance, but replacing the final 10% leads to severe declines” (Ashok & May 2024)�
• Upshot: Synthetic data may reduce total exploitation, but cannot eliminate it!�

Objection: Synthetic Data!

Reply 1: We Still Need Humans-in-the-Loop to Get Model Accuracy�

• Unless we accept significant drops in model accuracy, synthetic data (on its own) cannot completely eliminate the need for human annotators�
• Dhananjay Ashok & Jonathan May’s 2024 “A Little Human Data Goes a Long Way”�
• “The extent to which synthetic data can replace human annotation is poorly understood. We investigate the use of synthetic data in Fact Verification (FV) and Evidence-based Question Answering (QA) by incrementally replacing human-generated data with synthetic points on eight diverse datasets. Strikingly, replacing up to 90% of the training data only marginally decreases performance, but replacing the final 10% leads to severe declines” (Ashok & May 2024)�
• Upshot: Synthetic data may reduce total exploitation, but cannot eliminate it!�

Objection: Synthetic Data!

Reply 2: We Still Need Humans-in-the-Loop to Double-Check Data�

• Unless we accept significant risks model accuracy, synthetic data will generate a greater need for human verification that the synthetic data is realistic and usable�
• Shivani Kapania et al.’s 2025 “Examining the Expanding Role of Synthetic Data”�
• “Participants [AI developers] described validation of synthetic data as a consistent bottleneck. Ironically, the qualities that made synthetic data attractive—such as its scale, diversity, and the ability to simulate data for rare scenarios—also made it exceptionally challenging to validate since…validation was typically manual [i.e. by “eyeballing” the synthetic data].” (Kapania et al. 2025, 7) �
• Upshot (Again): Synthetic data may reduce total exploitation, but can’t eliminate it!

Objection: Synthetic Data!

Reply 2: We Still Need Humans-in-the-Loop to Double-Check Data�

• Unless we accept significant risks model accuracy, synthetic data will generate a greater need for human verification that the synthetic data is realistic and usable�
• Shivani Kapania et al.’s 2025 “Examining the Expanding Role of Synthetic Data”�
• “Participants [AI developers] described validation of synthetic data as a consistent bottleneck. Ironically, the qualities that made synthetic data attractive—such as its scale, diversity, and the ability to simulate data for rare scenarios—also made it exceptionally challenging to validate since…validation was typically manual [i.e. by “eyeballing” the synthetic data].” (Kapania et al. 2025, 7) �
• Upshot (Again): Synthetic data may reduce total exploitation, but can’t eliminate it!

Objection: Synthetic Data!

Reply 2: We Still Need Humans-in-the-Loop to Double-Check Data�

• Unless we accept significant risks model accuracy, synthetic data will generate a greater need for human verification that the synthetic data is realistic and usable�
• Shivani Kapania et al.’s 2025 “Examining the Expanding Role of Synthetic Data”�
• “Participants [AI developers] described validation of synthetic data as a consistent bottleneck. Ironically, the qualities that made synthetic data attractive—such as its scale, diversity, and the ability to simulate data for rare scenarios—also made it exceptionally challenging to validate since…validation was typically manual [i.e. by “eyeballing” the synthetic data].” (Kapania et al. 2025, 7) �
• Upshot (Again): Synthetic data may reduce total exploitation, but can’t eliminate it!

Objection: Synthetic Data!

Reply 2: We Still Need Humans-in-the-Loop to Double-Check Data�

• Unless we accept significant risks model accuracy, synthetic data will generate a greater need for human verification that the synthetic data is realistic and usable�
• Shivani Kapania et al.’s 2025 “Examining the Expanding Role of Synthetic Data”�
• “Participants [AI developers] described validation of synthetic data as a consistent bottleneck. Ironically, the qualities that made synthetic data attractive—such as its scale, diversity, and the ability to simulate data for rare scenarios—also made it exceptionally challenging to validate since…validation was typically manual [i.e. by “eyeballing” the synthetic data].” (Kapania et al. 2025, 7) �
• Upshot (Again): Synthetic data may reduce total exploitation, but can’t eliminate it!

Another Objection: Is Moral Exploitation Really a Distinctive Wrong?

�

Another Objection: Is Moral Exploitation Really a Distinctive Wrong?

Reply: Yes, it Infringes the Right to a Fair Division of Moral Labor!

• Consider the division of domestic labor in a patriarchal, nuclear family�
• In such families, the woman is highly pressured to perform an unfair amount of parenting as well as other “household labor” that keeps things running smoothly�
• She is wronged even if—as is often, though not always, the case—she finds this labor deeply meaningful and fails to be (very) psychologically harmed by it�
• Moral labor cannot be fully automated, but purely material labor can (c.f. “lights-out” factories)
• Moral labor is more intimately connected with who, and what, we are�� (continued on next slide)

Another Objection: Is Moral Exploitation Really a Distinctive Wrong?

Reply: Yes, it Infringes the Right to a Fair Division of Moral Labor!

• Moral exploitation brings with it an unfair risk of moral injury�
• So, the morally exploited person faces an unfair risk of performing “a betrayal of what’s [believed to be] right [and] shattering the ability to live with integrity in connection with others” (Gilligan 2014, 95)�
• Regarding the moral labor involved in some data-driven ghost work: At least arguably, this moral labor is so traumatizing that it shouldn’t exist in the first place �
• And even if it should exist, the psychological harms involved still raise difficult normative questions about how this traumatizing labor should be distributed

Examples for Agency

Young children

Unconscious people, animals, objects

Philosophy (Terms and Conditions)

Moral Agency: An individual's capacity to make intentional, deliberate moral decisions

​

​

​