Data Labor & User-Generated Data

• Many recent technological advances, such as ChatGPT and search engines, rely on massive amounts of user-generated data obtained through user interactions or web scraping.
• Data producers often have limited control over what data is collected, how it is utilized, and who benefits from it.
• Organizations like OpenAI and Google have significant influence in shaping the technology landscape due to their ability to access and process user data.
• Synthesizing existing data labor literature provides opportunities to empower data producers in their relationships with tech companies.
• Researchers, policymakers, and activists can advocate for transparency in data reuse, create feedback channels, and develop mechanisms to share data revenue more broadly.
• Data labor can be characterized in six dimensions: legibility, end-use awareness, collaboration requirement, openness, replaceability, and livelihood overlap.
• These dimensions are based on parallels between data labor and other types of labor in computing literature.

Introduction

• Technology users generate vast amounts of data through their interactions with computing systems, including behavior logs, content, and personal information.
• Currently, only a handful of tech companies with the resources to collect, process, and model data at scale benefit from this data.
• The creation of generative AI models like ChatGPT and Dall-E was made possible by publicly available texts and artwork that model developers scraped and processed from billions of web pages.
• Data producers such as artists, writers, and users have scant power over decisions about how their data is used or who benefits.
• This imbalance of power has led to public criticism of tech industry practices, including unapproved reuse of work and implications for employment opportunities.
• It is argued understanding data generation as a form of labor, or "data labor," can pave the way for more broadly distributing the power and benefits of data.
• Proposals include supporting "data unions," mediators of individual data, legislation granting users greater control, and tools supporting user-driven collective action.

Defining Data Labor

• The research community has yet to provide concrete guidelines for actionable data labor.
• A clear characterization is essential to guide researchers, data producers, and policymakers in addressing the existing power imbalance between the public and large tech companies.
• A clear characterization will highlight how different data labor types require different strategies and interventions to empower data producers through research, development, and policy practices.
• A working definition is offered: Activities that produce digital records useful for capital generation.
• An activity must meet two criteria to qualify as data labor: creating or enhancing data ("digital records") and helping to generate capital.
• Data labor subsidizes prominent tech companies due to the emerging, substantial power inequity between these entities and the data-generating public.

Computer-Mediated Labor

• Focus of HCI and CSCW research, encompasses both compensated and uncompensated activities.
• Data labor includes both compensated activities and unwitting ones like content creation on social networks.
• Crowdsourcing, peer production, and content moderation are examples of data labor that advance computing systems and benefit technology companies financially.

Digital Labor

• This refers to monetized online activities, regardless of traditional workplaces or compensation.
• The internet is animated by cultural and technical labor, continuously producing value.
• Not all instances considered digital labor are instances of data labor and vice versa.

Crowdwork

• Tasks completed by distributed laborers for payment fall under computer-mediated labor.
• Image labeling, text production for spam filters are examples.
• Completing behavior experiments may not be data labor unless it leads to capital generation.

Data Work

• This is a relatively new term includes user data generation, data labeling, and cleaning.
• It primarily concerns the upstream activity of generation, not downstream processing activities such as data cleaning and filtering.
• Data scientists cleaning datasets may also perform data labor by labeling images, for example.

Frameworks of Power and Data Leverage

• This work expands on existing literature about power imbalance and social inequalities in computing systems.
• Actionable steps support equitable social relationships by following this roadmap.
• Power-to: ability to choose not to participate in data-generating activities or delete data.
• Power-over: ability to influence technology operators around data-driven technology decisions.
• Data leverage involves data strikes, data poisoning, and conscious data contribution.
• Leverage requires collective action and critical mass participation.

Dimensions of Data Labor

• Six key dimensions of data labor are described: legibility, end-use awareness, collaboration requirement, openness, replaceability, and livelihood overlap.
• Each dimension is examined for relation to power & avenues to empower data laborers.

Legibility

• Addresses whether data laborers know their labor is being captured.
• Illegible data labor involves activities where individuals are unaware their actions create digital records.
• User interaction logs and completing reCAPTCHA are examples of illegible data labor.
• Legible data labor includes activities such as contributing ratings and completing clearly disclosed crowdwork tasks.

Relationship with Power

• Legibility is essential for data laborers to exert control over technology companies.
• A lack of awareness inhibits the power to withhold or change labor, and it limits collective action.
• An imperative is to move illegible data labor to legibility.
• Tools can measure and communicate the economic value of data labor, activists can use to equip data laborers with knowledge about how to effectively leverage it against tech companies.
• Mitigating illegibility may not always empower data producers due to external social restraints.

End-Use Awareness

• Centers on whether data laborers know how their labor is used to generate capital.
• Data laborers may unknowingly be captured with no end-use awareness, benefiting technology companies.
• Content creators, Wikipedia editors, and journalists may be unaware of downstream uses like search engine performance or large language model training,
• Data laborers who are end-use aware are informed of or understand how their output is used.
• This includes supporting particular functions or features with targeted advertising, personalized newsfeed algorithms, and recommender systems.
• End-use awareness can empower data labor, giving them an understanding of the downstream capital generation implications of the data they produce.

Mitigating a lack of awareness

• Policymakers may mandate end-use awareness for sensitive data, requiring tech companies to disclose how data will be used.
• Opt-in mechanism will provide laborers control over their end use.
• Data laborers' concern can disincentivize production of data labor.
• Activists may develop tools illustrating effects of collectively withholding or poisoning data.
• Strengthening legal frameworks enables data control prevent downstream usage.

Collaboration Requirement

• A spectrum from non-collaborative to collaborative activities determines the extent of teamwork among its laborers.
• A computer-mediated framework shows a distinction between team vs individual work.
• Non-collaborative data labor: isolated data production.
• Examples include completing a reCAPTCHA or a MTurk task.
• Collaborative data labor: Deliberation, communication, and interaction among data laborers, especially in social computing systems.

Relationship with Power

• Requires social connections to facilitate withholds, but those connections also create cost.
• Data laborers lack shared identity without collaboration.
• Shared identity among non-collaborative laborers is a crucial step in empowerment for future action.
• Collaboration can theoretically leverage network against the "employers."

Openness

• Data labor is characterized by the accessibility of downstream data to the public.
• Private systems offer closed labor excludes others.
• Maximal openness from copyleft licenses provide labor publicly.

Examples of Open Labor

• Academia
• Examples: Pushshift Reddit data, dataset
• Regulations include GDPR in European Union and CCPA in California.

Replaceability

• Data labor will have more ability directly impact technology performance when responsibility is directly is on labor.
• For performance to be at it's maximum in any situation, the requirements set on it must be met to full capacity.
• Gaining and exerting labor power will be harder for tasks that can be easily replaced or by many similar people.

Livelihood Overlap

• This notes if data labor overlaps occupational activities.
• Foundation models will increase this dimension.
• Companies may not establish formal jobs in place of data labor, leaving the workers without control over their work and outcome.
• No overlap = rare activity of web searching activity; low paying usually
• Overlap = writing code for systems by GitHub; higher paying

Discussion of the Article

• The six dimensions identified and articulated above are only a starting point for understanding the rich variety of data labor activities.
• Additional considerations emerge from the perspective of data-dependent operators.
• A discussion on revenue generation must happen as a direct of the complexity.
• More analysis and study on this needed.

Conclusion

• The study's synthesis has been done as review from labor and data, a roadmap is now made to empower data producers has been constructed with six defined dimensions to use in order to take the power back.