TL;DR for operators

An AI translation service can produce accurate text, support many languages, and satisfy a vendor’s stated compliance requirements without resolving the ethical questions surrounding its development and deployment.

In Translation and data ethics, Joss Moorkens and Monja Jannet bring together research on the ownership and reuse of translation data, algorithmic bias, minority-language inequality, digital exclusion, privacy, and accountability. Their central argument is that linguistic data originate in human labour and unequal social relationships. Treating those data as neutral inputs can conceal who contributed to a system, who controls its resources, and who bears the consequences of its outputs.

For organizations developing or procuring translation AI, the analysis suggests expanding evaluation beyond accuracy, price, language coverage, and vendor compliance. Dataset provenance, contributors’ rights, cultural representation, language-community control, privacy practices, and responsibility allocation all deserve examination.

The evidence boundary matters: this is a narrative scholarly synthesis, not an experiment or a validated governance framework. It identifies concerns that procurement and deployment processes may need to address, but does not establish which controls are effective or how much risk they reduce.

An accurate translation can conceal an unresolved data problem

A company plans to extend its customer-support service into several additional languages. Its procurement team compares translation accuracy, supported languages, latency, cost, and the provider’s legal assurances. A candidate system passes the evaluations and appears ready for deployment.

Yet the evaluation may have overlooked the translations that helped build the system. Who produced them? Were their creators entitled to control subsequent reuse? Did the dataset represent the intended language communities? And who is responsible if the service produces a culturally inappropriate or harmful result?

Moorkens and Jannet’s Translation and data ethics1 examines these questions through a narrative synthesis spanning translation studies, AI ethics, legal scholarship, and philosophy of technology.

The chapter challenges the treatment of language data as a resource available for collection and extraction independently of its human origins. Translated texts result from intellectual work, institutional decisions, and particular cultural circumstances. Those relationships do not disappear when texts become entries in a machine-learning dataset.

For the purchasing company, this changes what successful evaluation means. An acceptable translation demonstrates something about system output. It does not establish that the material used to build the system was obtained, represented, or governed appropriately.

Translation datasets contain intellectual labour, not just text

In Sections 1 and 2.1, the authors examine how large-scale collection and reuse can separate translated material from the people who created it.

A translation may be protected by copyright, subject to contractual arrangements, or covered by particular exceptions. The rights involved depend on the applicable jurisdiction and the terms governing the original work and translation. Consequently, neither unrestricted reuse nor a universal prohibition on training-data use follows automatically from the presence of copyrighted material.

The difficulty increases when many translations are aggregated into a training corpus. Individual contributions become harder to identify, while the commercial value of the collection may accrue primarily to the organization controlling the dataset or model.

The authors’ argument extends beyond legal permission. Even when reuse is legally permissible, questions remain about attribution, recognition, compensation, and the distribution of benefits from human translation work.

Cognaptus inference: A developer assembling translation corpora should distinguish three decisions: whether material can lawfully be used, whether its provenance and reuse conditions are understood, and whether contributor-related commitments are adequately addressed.

For a buyer, these considerations support asking vendors about dataset origins, licensing arrangements, and mechanisms for addressing rights disputes. Such inquiries can improve procurement transparency, although the chapter provides no evidence that any particular documentation standard prevents violations.

Language coverage does not guarantee fair representation

Sections 2.2 and 2.3 extend the problem from ownership to the content and social consequences of translation data.

The chapter discusses gender stereotypes in machine translation, particularly when translating from a relatively gender-neutral language into one that requires grammatical gender choices. A model may select gendered occupational terms according to patterns prevalent in training material rather than information justified by the source sentence.

This is an illustrative manifestation of representational bias, not an error rate estimated by the chapter. It shows why conventional assessments of linguistic adequacy may miss socially consequential choices.

The authors also discuss risks associated with synthetic training data. Drawing on previous experimental research, they note that repeated use of generated material without sufficient new, diverse observations can reduce output diversity and quality. The chapter does not test that mechanism itself or establish a universal threshold for safe synthetic-data use.

A further complication arises with minoritised languages. Adding a language to a model’s supported-language list does not necessarily ensure that its speakers’ cultural practices, institutions, and priorities are represented appropriately.

The chapter highlights Māori linguistic data sovereignty as an example of communities asserting authority over how language resources are collected, stored, accessed, and used.

This introduces a form of control that cannot always be reduced to individual copyright permissions. A language community may have collective interests in decisions about linguistic resources even when a technology provider can technically collect or process them.

For a company developing a product around community language resources, the decision is therefore not merely whether sufficient training examples can be acquired. It is also whether collection and commercialization arrangements respect the community’s interests and authority.

Community participation and governance may be appropriate responses, but the chapter does not test a particular engagement model or establish how competing claims should be resolved.

Making a translation service available is not the same as making it usable

The authors connect minority-language representation to a wider problem: digital language equality depends on more than access to a model.

A translation service may technically support a language while intended users lack reliable connectivity, suitable devices, digital literacy, privacy awareness, or institutional support to use it effectively.

Section 2.4 draws on global connectivity reporting and published research involving Myanmar migrant workers in Thailand to illustrate barriers extending beyond model availability.

These cases serve an explanatory purpose. They do not constitute a representative measurement of translation-service adoption or demonstrate that a specific accessibility intervention works.

For an enterprise or public-service operator deploying translation tools to underserved users, the relevant evaluation should therefore include the conditions of use. Can the intended population access the service reliably? Can users recognize when output may be unreliable? Do they understand how submitted information may be processed?

A product’s language list is not evidence that these conditions have been satisfied.

Accountability becomes harder when responsibility is distributed

Sections 3 and 4 address another consequence of automated translation: the difficulty of assigning responsibility when outcomes involve many people and technological components.

A harmful output may reflect training data, model design, deployment configuration, user instructions, or interactions among these factors. Opaque systems can make the causal contribution of each component difficult to reconstruct.

The authors connect this situation to the problem of many hands, distinguishing responsibility for an assigned role from causal responsibility, blameworthiness, and legal liability. These concepts are related but cannot simply be treated as interchangeable.

A related concern is credit-blame asymmetry. AI systems may receive recognition when work succeeds, while human translators or operators remain responsible for correcting failures. The chapter draws this argument from earlier scholarship rather than measuring its prevalence in translation workplaces.

Interface design further affects responsibility. A human designated as accountable may have limited ability to inspect, understand, or modify the system’s output. This is part of what the chapter discusses through technological mediation: technology shapes the choices and control available to its users rather than merely executing instructions.

Privacy creates another responsibility problem. Translation prompts and training corpora can contain personal or confidential information, while model opacity complicates tracing, deletion, and enforcement of data-subject rights. The authors discuss GDPR and the limitations of anonymisation in this context.

Regulation, including the EU AI Act and other legal mechanisms discussed in the chapter, can impose obligations or provide avenues for governance. It does not automatically resolve causal uncertainty, unequal power, or every question about legitimate data use. Applicable obligations also depend on jurisdiction and legal developments.

What translation buyers and developers should evaluate

The chapter does not prescribe a tested governance procedure. Nevertheless, its conceptual analysis supports several questions that organizations can incorporate into their existing decisions.

Decision area Question for the organization Boundary
Dataset approval Can contributors, data sources, licensing conditions, and reuse rights be identified? Provenance records alone do not establish lawful or ethical use.
Multilingual quality assurance Are gendered choices, cultural representation, and language-specific biases assessed alongside accuracy? The chapter supplies no validated testing protocol.
Minority-language development Have relevant communities’ interests in controlling linguistic data been considered? Appropriate governance depends on community and institutional context.
Privacy and deployment What information is submitted, retained, accessed, or reused by the service? Legal requirements and technical protections require separate verification.
Accountability What can providers, developers, employers, and users actually inspect, change, or control? Assigned responsibility does not establish causal or legal liability.

These questions have different implications depending on the actor.

A model developer deciding whether to ingest a corpus has direct influence over collection and documentation. An enterprise buyer may instead need contractual assurances and evidence from a provider. A deployment operator controls some aspects of access, workflow, and escalation, but may have little visibility into model training.

The chapter’s analysis cautions against treating those actors as if they possess identical information or authority.

What the evidence does not establish

Moorkens and Jannet provide a literature-grounded conceptual synthesis. Their chapter contains no original experiments, quantitative dataset, systematic literature-selection protocol, causal estimation, or comparative test of governance interventions.

Its examples identify plausible and documented forms of ethical concern through previously published work. They do not show how frequently those concerns arise across all translation systems, whether one model is more ethical than another, or how much a proposed governance measure reduces harm.

The absence of original intervention testing is particularly relevant for procurement decisions. Organizations can use the chapter to identify questions worth investigating, but cannot treat its arguments as proof that provenance documentation, community consultation, bias testing, or a particular accountability arrangement will deliver a measurable reduction in risk.

Similarly, its legal discussion provides context for governance, not a current legal determination for a specific contract, dataset, or jurisdiction.

The chapter’s contribution is therefore stronger as an integrated explanation of what organizations may overlook than as a specification of how those organizations must solve each problem.

A successful translation is only one part of the evidence

The central correction is straightforward: translation accuracy and regulatory compliance do not exhaust the ethical evaluation of an AI translation system.

The underlying data were produced by people, often within unequal economic and institutional relationships. Their reuse can affect contributor rights and recognition. Their composition can influence representation. Their commercialization can shift control away from language communities. And the systems built from them can distribute responsibility in ways that do not match users’ actual ability to intervene.

For organizations considering AI translation services, the resulting decision is broader than whether a model works. It includes whether the organization understands the origins of its language resources, the communities affected by their use, and the limits of the control and accountability arrangements surrounding deployment.

Moorkens and Jannet do not establish a definitive governance solution. They establish why a technically successful translation cannot, on its own, answer those questions.


Cognaptus: Automate the Present, Incubate the Future.


  1. Joss Moorkens and Monja Jannet (2026). Translation and data ethics. arXiv:2610.28710. https://arxiv.org/abs/2610.28710 ↩︎