English claims in German that nobody chose

Patrick Roye Published 30 June 2026 · Expanded canonical 1 July 2026 Read on LinkedIn

An English claim in a German campaign does two opposite things at once. In English it signals reach and confidence, whereas for a German reader the same line often produces the opposite: it reads as imported, unspecific, a little evasive, and so it gets quietly discounted. Phrases such as “industry-leading platform” are ordinary B2B claims in English, yet in German – “branchenführende Plattform” – they can read as self-awarded and unsupported, a boast the reader has no reason to grant and every reason to file under marketing. The English claim is an old problem; the new question is who chooses it when AI writes the German.

What has for twenty years been a marketing problem is now becoming a governance problem – and that shift is the part nobody is talking about. The market already produced English-sounding German deliberately, for two decades. What AI changes is not the output but where the decision is made, and how visible that decision stays.

The marketing problem is well documented. When Endmark and YouGov tested twenty English claims on German audiences, 64 percent could not correctly understand them (Claimstudie 2016).¹ The share of English slogans had climbed nonetheless, from roughly 3 percent in the 1980s to around 30 percent after 2000 (Androutsopoulos). The German was never the issue: a foreign-language claim simply carries less emotional weight for native speakers than the same claim in their own language (Puntoni et al., 2009).

Every English claim in German used to be a decision with an author, since someone chose the English line, signed it off, and could have been asked why. A visible choice can be questioned, and occasionally it is reversed.

In EN→DE claim review nowadays, that original decision has become harder to locate. A claim is built backward from an intended effect, for a particular reader, whereas the model only ever sees it the other way around: it reads the words that were written, not the marketing decision that produced them. It then reproduces the English register of its training data, where that register is heavily represented. The result emerges from a process rather than from a single identifiable decision, and because the behavior varies from model to model, no single technical fix resolves it.

The human sign-off was the step that did the catching.

None of this makes English the wrong call everywhere. When Deutsche Telekom replaced “Erleben, was verbindet” with “Connecting your world” in 2023, the brand team understood the trade-off and could account for it.² That is precisely the point: it was a decision, made by people who could be asked why.

So the failure that used to be authored is now automated. The English claim in German still misfires the way it always did – only now the accountable person who once caught it is gone.

Discussion

“The catcher is gone – so just detect the AI”

The reflex answer to a missing human reviewer is a technical one: if a machine wrote the claim, let a machine flag it. That exit is closed, and the reason it is closed is the same observation the body makes about the claim itself – a detector reads the words that were written, not the decision that produced them.

When Google evaluated its own translation-specialized model against its own general one, professional translators scored ten language pairs under MQM. The specialized model clearly outperformed the baseline for most pairs, with two exceptions – one of which was German, where the report states that both models are on par.⁷

MQM human evaluation · WMT25 · lower is better

Language pair	TranslateGemma 27B	TranslateGemma 12B	Gemma 3 27B
English→Italian	1.8	2.0	2.5
English→German	2.3	3.2	2.2

Google technical report, not peer-reviewed. The human evaluation drew on literary, news, and social text from the WMT25 test set – not B2B or marketing German.

Italian improved when the model was specialized for translation, and German did not. The best of the three systems on German, the specialized 27B, still landed a fraction behind the general baseline it was built to improve on, and it scored worse on German than the same model scored on Italian.

The evaluation domain is worth naming twice, because it cuts in an uncomfortable direction. Literary, news, and social text is rigidly structured, consistent material – the kind where machine translation performs at its best. German being the outlier there proves nothing about B2B German, though it does not encourage confidence about it either, and it might even cut somewhat against the assumption that a better score would have caught the problem.

The empirical record

Detectors are unreliable, and they harm exactly this writing

An evaluation of fourteen detection tools found them “neither accurate nor reliable,” with none reaching 80 percent accuracy and a consistent bias toward labelling machine text as human.³ Worse for the case at hand: GPT detectors misclassified 61 percent of non-native English writing as AI-generated, while flagging native writing almost never.⁴ That is the inverse of this article’s own subject. Foreign-influenced text – the register that carries over when a claim moves between languages – is precisely what a detector reads as machine-like. A tool built to catch the AI would disproportionately flag the human writing most shaped by another language.

The structural bound

The unreliability is not an engineering gap that improves away

A recent analysis argues the limit is structural, not a matter of better tools: no text-only, one-shot detector can escape a size–power trade-off wherever the human population and the model’s output overlap in distribution – a barrier the author frames as logically independent of how good the AI becomes.⁵ The claim to take from this is narrow and defensible: surface-and-style detection is unreliable and fading, not that detection is impossible in principle.

The counter-case, and why it confirms the narrow claim

Where detection does hold, it reads structure, not surface

The strongest recent counter-example makes the point rather than refuting it. A University of Maryland and Google DeepMind pipeline distinguishes AI-written fiction from human-written fiction at roughly 93 percent – but only by analyzing discourse-level narrative structure, not surface style, precisely because surface artifacts are editable and fade as models improve.⁶ That is the whole argument in miniature. Detection that works reads the structure behind the words; detection that reads the words is the kind that fails. And the structure a claim-review needs to judge – was this the right claim, for this reader, provable in this market – is not in the words at all. It is the marketing decision the model never saw.

So the catch the body mourns cannot be automated back into place. A detector sees the text, never the decision that produced it – the same gap, one layer up, that the body identifies in the model itself. The reviewer’s judgment was never a detection function. It was a governance function: deciding whether this claim, for this reader, in this market, was the right one – and being accountable for the answer. That function does not come back with a better detector. It comes back with a person whose job is to hold it.

Industry convergence · corroboration, not evidence

The direction is now named by the industry itself, though the sources sell the positioning and should be read with that in mind. RWS CEO Benjamin Faes has argued that AI “can produce grammatically perfect text in 50 languages without understanding any of them,” that grammatically correct is not the same as culturally intelligent, and – naming Germany specifically – that the opportunity is to become “the critical governance layer that makes enterprise AI trustworthy across cultures” (MultiLingual #253, July 2026; RWS is productizing exactly this).

Christine Clay (Alexa Translations / Apertera, same issue) frames the same movement as governance shifting “from policy to execution,” with human input as system design rather than post-editing. Vendor framing, affiliation noted – useful only as a sign that the upstream-governance shift is now the industry’s own stated direction, not as independent authority.

Sources

Endmark / YouGov, Claimstudie 2016 (n = 1,204): 64 percent of respondents could not correctly understand English-language claims, and only 28 percent could translate them in the sender’s intended sense. This is a later wave of the same ongoing Endmark study whose earlier waves (2003 / 2006 / 2009) appear in Chapter 02. ↩
Deutsche Telekom replaced “Erleben, was verbindet” with “Connecting your world” in September 2023, announced at Digital X in Cologne. Chief Brand Officer Ulrich Klenke framed the change as a deliberate move toward a globally uniform brand. ↩
Debora Weber-Wulff et al., “Testing of detection tools for AI-generated text,” International Journal for Educational Integrity 19:26 (2023). Fourteen tools evaluated; found “neither accurate nor reliable,” none reaching 80 percent accuracy, with a bias toward classifying output as human-written. ↩
Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou, “GPT detectors are biased against non-native English writers,” Patterns 4(7), 100779 (2023). Seven detectors misclassified an average of 61.3 percent of non-native TOEFL essays as AI-generated, while classifying native-speaker essays almost perfectly. ↩
N. A. Garland, “AI Detectors Fail Diverse Student Populations: A Mathematical [Analysis],” arXiv:2603.20254 (2026), unreviewed preprint. Argues that no text-only, one-shot detector can escape the size–power trade-off wherever distributional overlap exists between the human population and AI output – a structural barrier logically independent of AI model quality, cited as the theoretical explanation for the empirical unreliability documented by Weber-Wulff and Liang. ↩
Jenna Russell, Rishanth Rajendhran, Chau Minh Pham, Mohit Iyyer, and John Wieting, “StoryScope: Investigating idiosyncrasies in AI fiction,” arXiv:2604.03136 (2026), University of Maryland / Google DeepMind. Discourse-level narrative features distinguish human- from AI-written fiction at 93.2 percent macro-F1, deliberately avoiding surface-style signals because those are editable and fade as models improve. ↩
Google Translate Research Team, “TranslateGemma Technical Report,” arXiv:2601.09012 (January 2026). A Google technical report, not peer-reviewed. MQM human evaluation on the WMT25 test set across ten language pairs, drawing on literary, news, and social domains; professional translators annotated error spans with severity and category (Table 3, lower scores better). The report states that translation-specialized TranslateGemma clearly outperforms the general Gemma 3 baseline for most pairs, with two exceptions: German, where the models are on par, and Japanese→English, where the specialized model regresses. ↩