The concept of using English as a pivot to reach another language is nothing new in translation – when there was no direct pair, you went via English, the intermediary “bending” the results. Nor is it limited to writing. In 1999, Gottlieb and Grigaravičiūtė found that the Lithuanian voice-over for the Danish serial Charlot & Charlotte, made via an English pivot script rather than from the Danish, came out with flattened register, lost modal particles, ironed-out irony – the personality stripped from the dialogue.
When we moved to LLMs, the pivot didn’t disappear. It moved inside the model, no longer a procedural step you can avoid but the dominant conceptual layer in the architecture.
Research presented at the 2024 ACL Annual Meeting traced this inside the model: Wendler and colleagues showed that Llama-2, working with a non-English prompt, resolves meaning in its middle layers and assigns higher probability to the English version of the next word than to the input language, before the target language surfaces at all.
The effect is measurably heavier into German than out of it: in 2017, Evert & Neumann found the shining-through more pronounced from English into German than the reverse, pointing to a prestige effect.
It is the same failure mode the field documented decades ago, by a different mechanism. The human pivot was visible and optional; you could choose a direct pair. The model’s is structural, invisible, hard to detect in the output. AI industrialized a discredited practice and rebuilt it where you can’t route around it, and without a human reviser to catch it.
Why don’t prompts and engines help? Better prompts operate on the input only, while a different engine merely swaps one English-pivoting model for another. Neither reaches the layer where the pivot happens, because the model doesn’t reproduce the Anglo communication model from its training content. It reproduces it from where it resolves meaning, below anything a prompt or an engine swap can touch.
A system that routes conceptual content through English cannot reach what German B2B writing depends on – the register signals, proof structures, and expectation that a claim arrives with its substantiation.
The shape this takes in practice is consistent. English B2B copy leads with the claim and follows with support – and often the support isn’t in the text at all, but in a screenshot, a video, a linked page the model never sees. The model reproduces that structure faithfully: claim-first German, with the evidence trailing or missing entirely. Grammatically intact and built backwards for a reader who expects the evidence to come first. Nothing flags it as wrong, yet the reader stops trusting the text – often on instinct, without being able to say why.
The European Language Council’s 2025 roadmap names the result deceptive fluency – output that sounds convincing yet fails the reader it was written for. It calls the systems producing it multilingual but monocultural.1
The volume is not marginal. Bryan Murphy, CEO of Smartling, put numbers to it at a December 2025 conference: 84% of translated words are now AI-based, 43% with zero human review. He called the second figure a wild number.
So what can be done, if the model can’t be fixed from the inside? The models doing most of the commercial work are closed; their weights cannot be inspected, and no study has traced the pivot inside them. But the bias is not a hypothesis – the field documented it long before LLMs existed, and these models train on the same English-dominant data that produced the effect everywhere we can measure it. Even the one glimpse from behind a closed door points the same way: Anthropic’s interpretability work has noted English-default internal representations in Claude.
The profession solved pivot bias once, by choosing direct pairs wherever it could. Inside the model that choice is gone – you cannot tell it to stop resolving meaning in English. So pivoting away from EN no longer happens in the engine but downstream: in deciding what gets verified before it reaches the reader, and by whom.