Automated Translation at Scale: Bias, Erasure, and Accountability in Encoder-Decoder Systems

Every day, billions of sentences pass through Encoder Decoder Architecture models and arrive on the other side as something that resembles translation. The words are mostly correct. The grammar holds. But something has already been quietly settled — what counts as meaning, whose meaning gets preserved, and whose never gets asked about.

The Language We Pretend to Hear

The uncomfortable question is not whether automated translation works. It does, by most measurable standards. The question is what “works” has come to mean — and what that definition allows us to ignore.

Transformer Architecture models, particularly encoder-decoder systems rooted in the architecture Vaswani and colleagues introduced in 2017 (Vaswani et al.), treat translation as a compression-and-reconstruction task. The encoder reads a source sentence and distills it into a Context Vector — a fixed-dimensional numerical representation. The decoder generates target-language output from that compressed form, guided by an Attention Mechanism that selectively revisits parts of the source. It is a powerful capability. But it is also a bottleneck, because everything that does not fit the model’s learned distribution — tone, cultural register, gendered nuance in languages where gender inflects meaning differently than in English — passes through the same compression, and what cannot be quantified gets silently dropped.

Two Hundred Languages and a Noble Promise

The counterargument deserves its strongest framing. Meta’s NLLB-200 supports two hundred languages, including roughly 150 that most commercial systems never attempted. It improved BLEU scores by 44% over previous benchmarks, with gains exceeding 70% for some African and Indian languages (Nature). It processes around 25 billion translations daily on Meta platforms alone (Meta AI Research). Models like T5 and Bart continue to serve as widely used encoder-decoder baselines in research and applied settings.

For millions of speakers of languages previously excluded from digital communication, this represents real access — to information, to participation in conversations that were happening without them. Dismissing that accomplishment would be intellectually dishonest.

But accomplishment and accountability are not the same thing. The distance between them is where the damage accumulates.

When Accuracy Becomes a Mask for Erasure

The hidden assumption in large-scale translation is deceptively simple: if the output is statistically accurate, the translation is adequate. BLEU scores measure n-gram overlap between machine output and human reference translations. They measure resemblance. They do not measure whether meaning survived.

A decade-long review of gender bias in machine translation examined 133 studies and found that 105 treated gender as binary only (Savoldi et al.). Of those, 118 operated at the sentence level — disconnected from the discourse context that often carries gendered information. The research itself cannot see the full shape of the problem it is trying to measure. And it is overwhelmingly English-centric: the most common language pairs were English-German, English-French, and English-Spanish, leaving the languages with the most complex gender systems under the least scrutiny.

The Finnish pronoun “hän” — grammatically gender-neutral — still produced “He is a doctor. She is a nurse” as recently as February 2025 (Savoldi et al.). Google Translate has demonstrated strong masculine defaults when translating STEM occupations from gender-neutral source languages (Prates et al.). These are not edge cases. They are systematic patterns, reinforced through Teacher Forcing during training and propagated through Beam Search at inference time. The model does not decide to be biased. It learns what the training data treats as normal — then scales that normal to billions of daily interactions.

What disappears follows the same fault lines that colonial translation projects traced for centuries — the assumption that the target language’s categories are universal, that what cannot be expressed in the dominant grammar does not need expression at all.

The Translator Who Cannot Be Questioned

Every previous era of large-scale translation had identifiable translators. Colonial administrators who chose which indigenous terms to flatten into European equivalents could, at least in principle, be named and challenged. The decisions were visible — written in dictionaries, codified in grammars, contested by affected communities.

Encoder-decoder systems produce translations that emerge from statistical patterns distributed across billions of parameters. There is no single decision point to audit, no translator to question. The output arrives with the authority of scale and the anonymity of mathematics.

The EU AI Act, reaching full applicability on August 2, 2026 (EU AI Act), introduces requirements for data lineage tracking and human-in-the-loop checkpoints. But as of early 2026, no MT-specific regulatory standard exists, and the Act does not explicitly classify translation systems into a specific risk tier — classification depends on deployment context. A translation error in a social media comment and a translation error in a medical intake form carry vastly different consequences, yet the same model may produce both.

Who is accountable when an encoder-decoder system erases a patient’s gender identity in a clinical translation? The model developer? The platform that integrated the API? The hospital that trusted the output? There is no clear answer — because accountability was never part of the architecture.

Translation as Governance, Not Service

Here is what the compression metaphor obscures: translation at this scale is not a service. It is governance. When a single model determines how two hundred languages are rendered into each other, it sets norms — defining which meanings are portable and which are treated as residue.

Thesis (one sentence, required): Automated translation at scale, absent transparent accountability structures, functions as invisible cultural governance — determining whose meanings survive and whose are quietly overwritten.

This is uncomfortable because the alternative — not building multilingual systems — is worse. The discomfort is the point. The response to “this system has bias” cannot be “then stop translating.” It must be “then make the translation accountable.” Accountability requires something encoder-decoder architectures, by their mathematical nature, resist: legibility — the ability for affected communities to see what was decided, understand why, and challenge it.

The Debt We Have Not Named

If a model serves two hundred languages but the bias research covers only a handful of language pairs, how do we know what we are not seeing? If the standard evaluation metric cannot capture cultural meaning, are we measuring progress or measuring our own comfort? If the communities most affected by translation erasure are also the least resourced to audit these systems — what does inclusion actually mean?

The temptation is to reach for technical fixes: better datasets, debiasing algorithms, more representative benchmarks. Those matter. But they do not address the structural question — which is not “how do we make the model fairer?” but “who gets to define fairness for a system that speaks on behalf of cultures it has never listened to?”

Where This Argument Breaks Down

If encoder-decoder models significantly improve access for low-resource language communities — and the evidence suggests they do — then the cost of withholding these systems could exceed the cost of their biases. Imperfect translation may be better than no translation at all. If future work demonstrates that community-driven fine-tuning can meaningfully reduce erasure without centralized control, the governance framing becomes less urgent. That possibility should not be dismissed.

The Question That Remains

We built machines that can speak in two hundred languages. We have not yet built the institutions that can listen in any of them. Until the communities whose meanings are compressed, flattened, and reconstructed have a genuine voice in how these systems are designed and evaluated, translation at scale will remain a monologue dressed as conversation.