Data Governance in the Age of Large-Scale Data-Driven Language Technology

1 INTRODUCTION

• Language technologies are now embedded across online and offline activities, so their governance affects individuals’ and communities’ lives and access to technology.Deep-learning systems extract increasingly complex language statistics from growing text and speech collections, expanding their applications and societal reach.Source: §1 · p. 2
• LLMs depend on increasingly large and diverse language corpora whose management must address multiple jurisdictions, languages, rights holders, and downstream harms.Their broad cross-task and cross-domain use makes careful data identification, acquisition, and management an unresolved challenge.Source: §1 · p. 2
• The paper proposes a Data Stewardship Organization that coordinates stakeholders and rights holders to support responsible governance of multilingual language resources.The organization aims to foster data subjects’ and rights holders’ agency while enabling collaboration at the scale and diversity required by contemporary language technology.Source: §1 · p. 2

1.1 Research Context and Paper Outline

• The work is grounded in a year-long distributed collaboration involving over 1000 participants from 60 countries and multiple technical, social, legal, and multilinguality groups.The collaboration connects fields that are not normally closely connected through a practical case study pursuing ethical, legal, and technical goals together.Source: §1.1 · p. 2
• The paper reviews governance theory, values, legal and technical contexts, existing data-management approaches, and Wikimedia before presenting the proposed structure and actor relationships.The sections collectively address preserving training data, promoting stakeholder agency, and managing multiple legal contexts.Source: §1.1 · p. 3

2.1 Approaching Collaborative Governance: Theories and Mechanisms

• The framework treats governance as distributed coordination across laws, tools, norms, values, and organizations, while recognizing that governance mechanisms can reinforce power imbalances.It therefore attends to fragmentation and first-mover advantage when designing its own structure and processes.Source: §2.1 · p. 3

2.2 Values of Governance, Governance of Values

• The project derives values from workshop discussions, emphasizing inclusivity, non-discrimination, privacy, agency, creator rights, and equitable distribution of data-use benefits.The value set responds to concerns about technical-performance priorities, indiscriminate web-text use, and exploitative open-data practices.Source: §2.2 · p. 4
• Human-rights frameworks help ground the proposed values, but their diversity and historical focus on individual–state relations make them insufficient as absolute global foundations.The proposal therefore combines shared aspirations with local expertise, community-oriented justice, and participation in shaping governance values.Source: §2.2 · pp. 4–5

2.3 Object of Governance: Distributed Language Data

• The governance object is human-centric digital language data, whose creators, users, subjects, and rights holders face different representation, consent, commercial, and legal concerns.The scope includes text from news, scholarship, reports, blogs, social media, radio, and digitized books.Source: §2.3 · p. 5

2.4 Focus of a Data Governance Organization

• The proposal covers the data life cycle from creation and selection through curation, documentation, dissemination, hosting, versioning, and deletion.Governance choices depend on dataset characteristics and the stakeholders most directly involved at each stage.Source: §2.4 · p. 5
• Data Governance provides the overarching norms, frameworks, and communication mechanisms that connect Data Sourcing and Data Tooling.The distinction is people-centric: governance addresses stakeholders and exchanges, sourcing assembles datasets, and tooling operationalizes policies.Source: §2.4 · p. 5
• The diversity of language sources represented in the governance structure ultimately depends on the quality of the underlying sourcing efforts.This limits what the governance framework can achieve independently of sourcing practice.Source: §2.4 · p. 6
• Privacy governance combines local legal definitions, software detection of personal information, and sourcing decisions that account for privacy risks and stakeholder impacts.Contestation and removal additionally require actionable rules, searchable data instances, and controls preventing undisclosed dissemination of copies.Source: §2.4 · p. 6

3.1 Social Context: Social Variation and Language Discrimination

• Language varieties differ across education, geography, race, and socioeconomic class, while labeling minority varieties as “noisy” can reinforce linguistic discrimination.Gendered and racial disparities have also been documented in language technologies.Source: §3.1 · p. 6
• Governance should prioritize under-represented language communities, avoid equating sociolinguistic variation with noise, and give speakers a say in data decisions.Source: §3.1 · p. 6

3.2 Legal Context: Rights and Regulations

• Copyright and property rights create tensions between creator rewards, attribution, socially beneficial reuse, and restrictions on language-data use.Governance can help creators, hosts, and modelers navigate these tensions through locally relevant contestation and use-case frameworks.Source: §3.2 · pp. 6–7
• Long digital texts are difficult to anonymize, so privacy governance must address consent, deletion and rectification requests, and memorization risks in trained models.GDPR-style protections depend on how personal and publicly available data are defined.Source: §3.2 · p. 7
• User-rights frameworks address algorithmic discrimination, accessibility, explanation, and performance, while data governance can support representation and system auditability.The effectiveness of relatively new laws, including bias-audit requirements, remains untested.Source: §3.2 · p. 7

3.3 Machine Learning Context: Challenges and Incentives

• ML data governance remains underdeveloped because research prioritizes model architecture and training while dataset curation is comparatively undervalued.For LLMs, the distance between pretraining data and task-specific models further abstracts the relationship between data choices and progress.Source: §3.3 · p. 7

• Across these approaches, coordinating stakeholder priorities, legal requirements, values, and interoperability remains a central unresolved challenge.Source: §4 · p. 8

Centralized Dataset Management

• Centralized repositories improve persistence, reproducibility, and standardized access, but give governing organizations disproportionate control and provide limited visibility or recourse for data subjects and providers.Source: §4 · p. 8

Public Dataset Repositories

• Public dataset repositories broaden contributions and documentation, but rely on submitters to assess legality and cannot provide the oversight needed for autonomy, consent, and contestation.Source: §4 · p. 8

Open Data Initiatives

• Open-data initiatives improve access to otherwise unavailable datasets and support reproducibility, but persistent unredacted copies conflict with requests to remove personal information.Source: §4 · p. 8

Distributed Data Governance in the Wikimedia Project

• Wikimedia demonstrates how distributed curation can combine licenses, evolving policies, local contestation, and technical tools across diverse stakeholders.Its experience also exposes tensions between contributor rights, cultural values, and the use of underrepresented communities’ knowledge.Source: §4 · pp. 8–9

• The structure organizes six distributed actor categories and uses communication across them to align end goals, legal issues, values, and interoperability.The DSO coordinates relationships while Data Sourcing and Data Tooling support the broader structure.Source: §5 · pp. 9–10
• Data Rights-holders represent interests in consent, privacy, attribution, contestation, and just rewards, while Data Providers introduce data under conditions permitted by those rights-holders.Source: §5 · p. 9
• Data Hosts aggregate and maintain data, can decline offered sources, and require global proximity to support linguistic diversity and meaningful contestation rights.The applicability of data-protection laws across jurisdictions remains an open question.Source: §5 · pp. 9–10
• Data Modelers receive access subject to rights-holder, provider, and host requirements, while seeking visibility, unified formats, and appropriately bounded replicability.Source: §5 · p. 10
• Contractual flow-down requires each actor to communicate predecessor restrictions, with hosts and modelers using binding agreements and updated dataset versions to honor deletions and expired licenses.Source: §5 · p. 10
• The DSO supports interoperability through standard agreements, license-taxonomy rules for aggregating restrictions, technical tools, and fallback processes for actors lacking resources.Providers may retain custom licenses, while the DSO offers standard options to foster inclusion.Source: §5 · p. 11

• The approach coordinates providers, hosts, modelers, rights-holders, and other stakeholders through a Data Stewardship Organization and complementary technical tooling.Source: §6 · p. 11
• The year-long case study is international and collaborative, while the governance structure uses values to guide management plans, access restrictions, and legal agreements.Source: §6 · p. 11
• Developing legally compatible agreements across providers, hosts, modelers, copyrights, and institutional jurisdictions remains one of the most difficult hurdles.Source: §6 · p. 11

A.1 Overview of Approach

• The working group refined an initial value set through debate, defining representativeness as capturing the full diversity of human language use.The process began with planners’ implicit values and incorporated comments from a broader group.Source: §A.1 · p. 16

A.2 Steps

• Participants made implicit values explicit by chronologically annotating planning documents and discussing highlighted values in their surrounding context.Terms initially served as placeholders before live discussion produced generally agreed definitions and, in some cases, multiple value labels.Source: §A.2 · p. 16
• The project used screenshots of planning drafts to document how value definitions and labels evolved during discussion.The figures present representative drafts in chronological order.Source: Figure 4 · p. 15, Figure 5 · p. 15, Figure 6 · p. 15

A.3 Growing

• The working group grew from around 10 people to more than 50, with newcomers introduced to the grounding values and governance plan.Presentations on the values and structure arguably helped bring additional participants into the group.Source: §A.3 · p. 16

B DSO STANDARD HOST-PROVIDER AGREEMENT

• The DSO’s governance values were supported by templates for legal agreements between Data Hosts and Data Providers.The standard agreement formalizes a relationship intended to support the proposed governance values.Source: B DSO STANDARD HOST-PROVIDER AGREEMENT · p. 17

1. PREAMBLE

• The agreement responds to ethical and legal problems associated with relying on Internet snapshots, including amplification of harmful biases.The project instead builds an annotated catalog intended to cover diverse languages and social contexts.Source: §1 · p. 18
• BigScience identifies diverse, high-quality language resources and assigns data availability to Data Providers and Data Hosts under an ethical charter.Providers license datasets for research, while Hosts contribute technical capacity to make them accessible.Source: §1 · p. 18
• The Data Provider Agreement governs data sharing between interested parties for the benefit of open research.It serves as the core instrument for enabling and governing that exchange.Source: §1 · p. 18

2. DEFINITIONS

• The agreement defines Data as machine-readable content, Metadata as supplementary information, and a Dataset as a specific collection whose provider has the necessary sharing rights.It also distinguishes processed datasets, derived works, users, hosts, and providers.Source: §2 · p. 19
• The Data Host receives a worldwide, perpetual, royalty-free license to access, store, process, label, and share agreed data under specified use cases and redistribution rules.The license is non-exclusive, non-transferable, and non-sublicensable.Source: §2 · p. 19

4. DATA PROVIDER RIGHTS AND OBLIGATIONS

• The Data Provider warrants ownership or necessary rights and supplies contact information for resolving dataset queries or issues.These obligations establish the provider’s authority to enter the agreement and support communication with the Host.Source: §4 · p. 19
• The Data Provider supplies data in an agreed suitable format and notifies the Host about restrictions, suspected infringement, regulatory noncompliance, or other relevant issues.The agreement also requires disclosure that the dataset contains no malicious code.Source: §4 · p. 20
• The Data Provider disclaims warranties concerning the data’s fitness, accuracy, adequacy, and completeness.The provider is not liable for dataset malfunctions, errors, or omissions, while agreeing to notify the Host when problems become known.Source: §4 · p. 20

5. DATA HOST RIGHTS AND OBLIGATIONS

• The Data Host receives no rights beyond those stipulated and must hold the Data Provider harmless for authorized use.The data is provided “as-is” to end users and other entities with access.Source: §5 · p. 20
• The Data Host must promptly report suspected infringement or related claims and stop using the data until the issue is resolved.This requirement applies when the Host becomes aware of intellectual-property, damage, or other claims associated with the data.Source: §5 · p. 20
• The Data Host should attribute the Data Provider and may publish research, benchmarks, performance indicators, and scientific insights derived from the data.These permissions are framed within BigScience’s research scope.Source: §5 · p. 20
• The Data Host undertakes commercially reasonable efforts to remove personal data before using the dataset, while the Provider must report known personal data.The agreement therefore assigns privacy-related responsibilities to both parties.Source: §5 · p. 21

6. LIMITATIONS

• Data access is restricted to the agreement’s stated purpose, access policy, and listed dataset, while liability for users’ processing is excluded for both parties.The Data Provider may also bear claims arising from non-observance of its obligations when it complies with the data management plan.Source: §6 · p. 21
• Each party bears its own implementation costs, and the Data Provider and Data Host must use up-to-date security and authentication practices.Security measures may include secure transmission, protected servers, and secure processing.Source: §6 · p. 21

9. TERM AND TERMINATION

• After termination, the parties must stop using the Data and Processed Datasets within 14 days and then delete them, subject to legally required retention, while completed Derived Work remains unaffected.Termination may follow an uncured breach or voluntary termination with justification, 60 days’ notice, and an opportunity to negotiate continuity.Source: §9 · pp. 21–22
• The agreement excuses failures caused by circumstances beyond reasonable control and requires confidential treatment of covered information, with disclosure permitted when legally required.Confidentiality excludes public information and information independently obtained or already available before the agreement.Source: §9 · p. 22

17. NO ADDITIONAL TERMS

• The agreement can be modified only by mutual written consent, invalid provisions are severable, and no additional terms bind either party unless expressly agreed in writing.The agreement and its exhibits constitute the parties’ entire understanding.Source: §9 · p. 22, §14 · p. 22, §17 · p. 23

14. DISPUTE RESOLUTION

• Disputes arising from breach must first enter alternative dispute resolution under the auspices of the BigScience Community.Specified provisions, including liability limits, termination, confidentiality, governing law, survival, and dataset-use restrictions, continue after termination.Source: §14 · p. 22

I. Data Provider Information

• The Data Provider Information section includes a special-conditions field for applicable terms.Source: Data Provider Name · p. 24
• Special conditions are recorded as part of the Data Provider information.Source: Data Provider Name · p. 24
• The field is conditional rather than universally required.Source: Data Provider Name · p. 24

List of Datasets

• Providers can specify dataset licenses, restrictions, scope or use cases, and conditions such as open artifacts, RAIL licensing, or value alignment.Source: List of Datasets · pp. 25–26
• The agreement offers optional frameworks for dataset use, reuse, and distribution while acknowledging providers’ economic interests.Source: List of Datasets · p. 25
• Data may be distributed through authenticated systems, anonymization, public licensing, or downstream access conditioned on nondissemination agreements.Source: List of Datasets · pp. 25–26