Ethical AI in Language Services. What Does It Actually Look Like?
It all started with a question. In 1950, Alan Turing1 asked: “Can machines think?” Five years later, John McCarthy2 gave the field its name, Artificial Intelligence. Fast forward to the Generative AI Revolution, and AI is all we hear about. Algorithms, models, and innovations dominate our lives, and there is no opting out. The old “I’ll Google it” has been hijacked by a neat AI summary sitting right at the top of the page. Industries are being remade so quickly that regulation and safeguards are struggling to keep up. What is true today will no longer be true tomorrow.
The language services industry has been hit by this storm head-on, and it is still grappling with the effects of an advancement that at times felt forced. Language professionals have worried about their roles. Language companies have worried about their survival. Yet with real resilience, everyone we have talked with is trying, again and again, to embrace the technology, make the most of it, stay relevant, protect jobs, and pivot toward new possibilities for improving language access and quality.
Among language rights advocates, though, harder questions are surfacing. Because while we talk about possibilities that are difficult to materialize, we also talk about safeguards, guardrails, human oversight, and ethical and responsible use, and we struggle to find concrete examples and guidance for any of them. AI governance has been inconsistent at best. For ordinary people with ordinary lives, simply trying to get by in a time of multiple crises, these are buzzwords, with no concrete plan or backing to protect them. And the language services industry lives this contradiction daily: we are asked to adopt the tools, reassure the clients, and absorb the risk, all while the safeguards everyone keeps invoking remain promises on a slide.
We have been looking for words to help us make sense of this all. To name what we are trying to advocate for. And some very smart people did find them. Enter the Cagliari Statement on Linguistic and Conceptual Justice in AI4.
The Cagliari Statement
The Cagliari Statement, drafted in Cagliari at a satellite workshop of the Linguistic Justice Society and signed in September 2026 by researchers from universities on five continents, raises a deeper point. It is not enough to ask whether AI speaks your language (access) and whether it translates accurately (quality). We also have to ask whether AI allows your language to mean what its speakers mean by it.
Let's backtrack a second. Languages carry meaning, memory, and cultural identity. Different languages organize, categorize, and describe the world differently. Consider the research by Caitlin Fausey and Lera Boroditsky3, which found that English speakers were more likely to describe an accident by naming the person responsible ("She broke the vase"), while Spanish speakers tended to use non-agentive language that emphasizes what happened rather than who did it. Remarkably, the researchers also found differences in what people remembered: English speakers were better at recalling who caused accidental events. It is a small but tangible example of the larger point the Statement makes: languages are not interchangeable sets of words. They can lead us to frame and attend to the same experience differently.
Conceptual Justice
This difference matters when we turn to AI. Language models learn meaning from enormous amounts of existing text, and the Statement warns that those resources are heavily dominated by English. As a result, the meanings and categories an AI learns can reflect English-language usage disproportionately. The problem is not limited to whether there is enough data for a language or whether a system can produce fluent sentences in it. A system might be perfectly capable of communicating in another language while still calibrating its judgments to English semantic norms.
Here the Statement draws a distinction that is easy to miss: inconsistency within a language versus inconsistency across languages. If an AI responds differently to essentially identical situations in the same language, that is a genuine technological defect. But what if it gives different judgments when the language changes? Our instinct might be to correct the AI so everyone receives the same answer. The Statement cautions that this is not always right. Sometimes the discrepancy reflects legitimate differences in how concepts are understood and used within those languages.
So should AI make the languages agree, or should it preserve the difference between them?
Linguistic Justice
This is where linguistic inequality quietly enters the picture. Because English dominates the data used to train current AI systems, striving for consistency across languages can, in practice, bring other languages into line with English meanings and categories. The Statement calls this cross-linguistic deference: resolving disagreement between languages by treating English, or another dominant language, as the authoritative standard.
That is why the Statement frames this as a question of linguistic and conceptual justice, not merely better translation or better technology. If different languages encode concepts somewhat differently, no language inherently deserves to be the authority against which the others are measured. The Statement names this equal semantic standing: “the semantic norms of no language carry inherent authority over those of any other”4.
It also reveals the gap between language inclusion and language equality. Expanding AI into more languages matters, but availability alone does not guarantee justice. A language can be fully supported, fluently generated, and well represented in an AI system while its concepts are still being interpreted through English semantic standards.
What the Statement Calls For
The Statement does not stop at diagnosis. It issues five distinct calls, each aimed at the people with the power to change how these systems are built and judged.
To AI developers and providers: treat cross-linguistic alignment policies as a published, reviewable commitment. Say openly whether a model is built to return the same judgment in every language, or to stay faithful to the ordinary usage of each language. And do not suppress divergence between languages as mere contradiction.
To researchers in AI, linguistics, and the study of concepts: build open, empirical, multilingual evidence on where conceptual divergence actually occurs, and when evaluating models, report within-language consistency and cross-linguistic agreement as separate findings rather than blending them.
To institutions and funders: support not only work that extends language coverage but also work that documents how languages other than English are actually used, on their own terms. And where AI outputs feed policy decisions, disclose the conceptual standard behind them as a condition of funding or procurement.
To standard-setting bodies and benchmark developers: stop treating agreement across languages as a measure of quality. “A benchmark that rewards a model for giving the same answer in every language encodes cross-linguistic deference as a metric”4. Report within-language consistency and cross-linguistic agreement separately, and reward only the former.
To all parties: recognize “cross-linguistic inconsistency in AI not as a bug to be quietly removed, but as the point at which linguistic and conceptual justice become visible”4.
LACC's Ask #5
Read this way, the Statement is more than an academic intervention. It is the concrete grounding that language rights advocates have been looking for, and it maps directly onto the Language Access Coalition of Canada's call to Set Ethical Guardrails for AI in Language Services: a clear ethical framework governing the use of AI in translation and interpretation, built on transparency, bias mitigation, multilingual equity, community consultation, data privacy protections, and human oversight.
The Statement supplies the intellectual foundation for several of those guardrails.
Its principle of transparency, where judgments are language-dependent, is transparency, given teeth. Imagine an AI system assesses a sponsorship application written in Arabic and finds it “inconsistent”. If the system simply outputs its judgment, the applicant never learns that a language difference shaped the outcome. The Statement’s transparency principle says: where a judgment depends on the language it was made in, and the stakes are not trivial, say so. That is transparency given teeth, because it shifts the burden of noticing from the person affected to the system itself.
Equal semantic standing is multilingual equity, made precise enough to test a system against. Ask an English-language AI whether a traditional healer “practices medicine,” and you may get one answer; ask the same question in the language of the community where that title carries recognized standing, and get another. Equal semantic standing means the system may not resolve that disagreement by quietly deferring to English. It is multilingual equity made precise enough to test: you can literally check whether a model’s judgements in one language overrule what speakers of another language mean by their own words.
Its insistence that the communities whose concepts are at stake have a legitimate claim to a voice is community consultation, stated as a right rather than a courtesy. When AI has to decide how a term like “family member” applies in a community where the concept extends to kinship relations English does not name, the people whose concept is at stake are the ones with the standing to say how it should be understood, not the engineers who trained the model. That is community consultation stated as a right: the community holds a legitimate claim to a voice in the resolution, not a seat at an optional feedback session.
And its call for language-specific alignment grounded in how real speakers actually use their words is bias mitigation at the deepest level, at the level of meaning itself. Most bias fixes try to strip harmful associations out of a model. The Statement’s call goes one level deeper: ground the model in how ordinary speakers of each language actually use their words, so that its judgements in Inuktitut reflect Inuktitut usage, not English translations of it. Bias at the level of meaning is invisible to the usual audits, and this is the first framework that names it and prescribes a fix.
What the Coalition's ask adds, and what language services uniquely demand, are the remaining pillars: data privacy protections for the words and stories communities entrust to these systems, and human oversight by the professionals who carry the accountability that no model can.
Access and quality got us asking whether AI works in every language. Linguistic and conceptual justice asks the harder question: whose meaning wins when languages disagree? The Cagliari Statement gives us the words for that question. The Coalition's guardrails are how we answer it.
List of Cited Sources
Turing and the 1950 question
Turing, A. M. (1950). "Computing Machinery and Intelligence." Mind, 59(236), 433–460. Published by Oxford University Press in the journal Mind, October 1950; this is the paper that opens with the question "Can machines think?"
Direct link: https://doi.org/10.1093/mind/LIX.236.433
McCarthy and the 1955 founding of the AI term
McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (2006). "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955." AI Magazine, 27(4), 12.
Direct link: https://ojs.aaai.org/aimagazine/index.php/aimagazine/article/view/1904
The Fausey and Boroditsky research
Fausey, C. M., & Boroditsky, L. (2010). "Who dunnit? Cross-linguistic differences in eyewitness memory." Psychonomic Bulletin & Review, 18, 150–157.
Direct link: https://link.springer.com/article/10.3758/s13423-010-0021-5
The Cagliari Statement
Linguistic Justice Society, “Cagliari Statement on Linguistic and Conceptual Justice in AI,” signed in September 2026: https://sites.unica.it/ljs/cs/.
Comments