Translation and data ethics

1 Introduction

• The “data as oil” metaphor makes data seem objective while separating it from the translators who created it.Kenny notes that resource metaphors can portray translation as an undifferentiated mass, obscuring individual translators’ contributions.Source: §1 · p. 1
• The shift from symbolic to subsymbolic AI replaced comprehensible steps with machine-learning processes that can be difficult to interpret.LLMs find patterns through self-supervision, further automating translation and partially or fully replacing human steps.Source: §1 · pp. 1–2
• Data ethics examines moral problems involving data, algorithms, and related practices to formulate and support morally good solutions.The chapter connects this field to translation ethics because translation brings human and data-driven technology into interaction.Source: §1 · p. 2
• The chapter charts ethical issues, discusses responsibility and regulation, and considers future research and discussion.Source: §1 · p. 2

2.1 Copyright, ownership and reuse

• Translation copyright can depend on jurisdiction and originality, yet translators often transfer rights to clients and their contributions are difficult to trace at scale.The chapter notes that copyright exceptions for LLM training have been introduced in many countries to balance expression protection with innovation.Source: §2.1 · pp. 2–3

2.2 Data bias

• Gender bias can arise when LLMs choose gendered forms for gender-neutral inputs by reflecting stereotypes, such as associating doctors with men and nurses with women.Source: §2.2 · p. 4
• Bias in LLM outputs is often learned from training corpora rather than fabricated, and synthetic training data may also cause bias.When synthetic training data lacks enough new, previously unseen data, model output quality and diversity can diminish.Source: §2.2 · pp. 4–5
• LLM translations into minority languages can carry high-resource grammatical structures, while output quality depends on the amount and quality of digitally available training data.Minority languages often lack large quantities of digital training data.Source: §2.2 · p. 5
• European-language coverage varies: English and German dominate technological and contextual resources, while small languages rank lower.Technological factors include tools, services, datasets, and lexical resources; contextual factors include economic, educational, legal, and funding conditions.Source: §2.2 · p. 6
• Minority-language communities may reject commodification of their data, as Māori advocates assert collective control over its creation, collection, use, and storage.The Māori Data Sovereignty network treats language data as a collectively owned taonga.Source: §2.2 · p. 6
• The digital divide now includes gaps in technology skills, use, and performance as well as access, and underrepresentation and information asymmetry can perpetuate disadvantage.Source: §2.2 · p. 6
• More than 25% of the world’s population is entirely offline, and only 14 per cent of rural residents in low-income countries are online.New users may lack digital literacy and privacy awareness, while Myanmar migrant workers in Thailand faced keyboard-input and translation-quality barriers.Source: §2.2 · p. 7

3 Responsibility and regulation

• AI governance varies by jurisdiction, from the EU AI Act and China’s guidelines to a US approach that has tended to leave governance to companies.The chapter notes that sidelining ethics experts in some companies raises questions about the importance of internal review.Source: §3 · p. 7
• Responsibility for harms in data-driven systems is difficult to attribute because black-box causation and the many-hands problem obscure individual accountability.The chapter also describes a credit-blame asymmetry when humans oversee and repair AI output, and notes proposals for transparent disclosure of roles.Source: §3 · p. 8
• Legislation has struggled to keep pace with technological development, and the proposed EU AI Liability Directive was withdrawn from the 2025 Work Programme.Recent legislation may place obligations on AI developers and providers without necessarily attributing liability.Source: §3 · p. 8
• Web scraping and user prompts can introduce personal data into AI development and use, while GDPR rights and consent requirements are difficult to enforce with black-box LLMs.Complete anonymisation has demanding requirements and is impractical in many cases.Source: §3 · pp. 8–9

4 Future directions

• Data-related harms are difficult to quantify, while AI hype and political pressure make regulation difficult.Source: §4 · p. 10
• Translation researchers use critical theory to examine how translation technologies can encode hegemonic values and enable data colonialism through linguistic-asset extraction and monetisation.Researchers also describe data as socially constructed through collection and as an incomplete representation of a population.Source: §4 · p. 10
• Postphenomenological analyses examine how machine translation can empower or misinform, amplify technologists’ input, conceal annotators, and constrain output to common past training-data patterns.These effects may heighten risks for low-resource language users.Source: §4 · p. 10

• Copyright, ownership, and bias are ethical issues that apply across data types.Source: §5 · p. 11
• Minoritised language communities and people on either side of the digital divide face distinct issues tied to data’s role in power imbalances.Source: §5 · p. 11
• The chapter discusses responsibility and regulation and reflects on growing attention to data ethics among translation studies researchers.Source: §5 · p. 11

• Floridi and Taddeo (2016) is recommended as a starting point for an introduction to data ethics.Source: §6 · p. 11
• Mellinger and Hanson’s 2025 discussion addresses data issues around data-driven optimisation in the translation industry and its risks.Source: §6 · p. 11