2025-05-21 · Baduno Editorial Team · 8 blog.readMin · Blog & Knowledge
Detecting Machine Translation: Seven Typical Error Patterns
AI translations have become good – their errors have too: rarer, but subtler. How reviewers spot them before clients do.
The Obvious Four
False friends ('eventuell' becomes 'eventually'), inconsistent technical terms in the same text, literal idioms, and address breaks between 'du' and 'Sie': A trained eye finds these classics in minutes.
The Subtle Three
Confidently wrong numbers and proper names, culturally off-key examples, and an overly smooth uniform tone that sands away any brand character. Dangerous because the text looks flawless – it's just not correct.
Systematic Review
Machines check structure (numbers, placeholders, completeness, terminology), humans check meaning and impact – in this order, with documented results. Spot checks do not replace systematic review on money pages.
Implications for Buyers
Don't ask service providers if they use AI – ask how they find its errors. If they can't describe a review process, they don't have one. Our answer is public on this website.
Misleading Idioms and Metaphors
Machine translations often fail with idiomatic expressions that are not directly transferable. For example, 'to kick the bucket' quickly becomes 'den Eimer treten' in German, whereas the idiom is 'den Löffel abgeben'. Likewise, metaphors like 'the ball is in your court' should be culturally adapted to the target market, such as 'Sie sind am Zug'. Reviewers must recognize not only the literal meaning but also the intended effect. A typical example: 'break a leg' is often translated by AI as 'brich dir ein Bein', which is understood as a bad wish in German – the correct version is 'Hals- und Beinbruch'. In marketing texts, such errors can be embarrassing and distort the message. Therefore, reviewers should look for common idioms in the source text and manually check their target language equivalents. A good workflow: AI translates, the reviewer marks all idioms and validates them using a reference database or native speaker intuition.
Errors in Numbers, Dates, and Formats
AI translations often handle numbers correctly, but not always – especially with thousand separators, decimal separators, and date formats. A US date '03/04/2024' can be interpreted as April 3 or March 4. Currencies: '$1.50' becomes '1,50 $' in German, but some AIs omit the comma or keep the period. Also units: '10 miles' should become '16 kilometers', not '10 Meilen'. Even trickier are percentages with cultural differences: 'a 50% increase' can mean 'eine Steigerung um 50 %' or 'auf 150 %'. Reviewers should therefore systematically compare all numbers, dates, units, and currencies with the source text. A checklist helps: check decimal separators, date format (DD.MM.YYYY vs. MM/DD/YYYY), currency symbols (€ vs. $) and units (km vs. miles). Particularly important for price information in online shops – a wrong comma can multiply the price tenfold. Document every deviation and correct the formatting according to the target language rules.
Incorrect proper names and brand terms
Proper names, brand names, and product names should typically remain untranslated. However, AI tends to translate them as well: 'Apple' becomes 'Apfel', 'Microsoft' becomes 'Mikrosoftware'. Even more critical: specialized terms like 'SAP' or 'CRM' are sometimes replaced by supposed equivalents ('Sozialistische Arbeiterpartei' instead of systems). Acronyms like 'NASA' should also not be spelled out unless the target culture uses the translation. Reviewers must therefore maintain a terminology database that specifies which names should be retained. During review, all proper names, company names, product names, and acronyms should be marked and cross-checked against this database. A common mistake: 'Köln Bonn Airport' becomes 'Cologne Bonn Airport' – depending on the target language, both can be correct, but consistency is key. Document every decision so that the same name is treated uniformly in future translations.
AI translations have become good – their errors have too: rarer, but subtler. How reviewers spot them before clients do.
Missing long-term context and updates
Machine translations usually only consider the current sentence or paragraph, not the entire text or previous translations from the same client. This leads to inconsistencies with recurring terms, phrases, or messages. Example: A company changes its slogan from 'Quality first' to 'Quality & Speed'. The AI translates the new text correctly, but all old translations remain – visitors see different claims on different pages. Also with product launches: one version is called 'Pro 2024', the next 'Pro 2025'. Without long-term context, the AI might translate 'Pro 2024' as 'Pro 2024' and 'Pro 2025' as 'Pro 2025', but if the client wants the brand name to always remain untranslated, that must apply globally. Reviewers should therefore consult the client style guide before starting and check existing translations for consistency. Even better: use AI systems with Translation Memory (TM) that recognize previous translations. Nonetheless, human review remains indispensable to ensure that brand identity remains consistent across all channels.
Errors in URLs, metadata, and hreflang tags
Machine translations usually focus on the visible text and ignore technical elements such as URLs, metadata, and hreflang tags. Yet these are essential for multilingual search engine optimization. Typical errors: The AI translates URL slugs, even though they should usually remain unchanged – from "/de/produkte" it incorrectly becomes "/en/products", leading to dead links. Meta titles and descriptions are also often automatically translated without respecting the character limit: German meta descriptions with 160 characters are often too long after translation into English and get truncated. Particularly critical are hreflang tags: These HTML tags indicate the language versions of a page to search engines. An AI might accidentally set hreflang="de" for an English page or point the canonical URL to the wrong language version. The result: Search engines index the wrong pages or penalize duplicate content. Reviewers should therefore systematically go through all links, metadata, and hreflang tags after translation. Use tools like Screaming Frog or manual sampling. Ensure that URL structures remain consistent (e.g., /de/, /en/ as language paths) and that metadata is optimized for each language. Document each adjustment to maintain the technical framework for future translations.
Localization of legal texts and safety instructions
Legal texts such as general terms and conditions, privacy policies, or legal notices often contain standardized wording that AI translates literally correctly but is legally insufficient. For example, the German 'Einwilligung gemäß Art. 6 DSGVO' often becomes a generic 'consent according to law' in English, without specifying the specific legal basis. This can lead to cease-and-desist letters. Safety instructions for products are also particularly vulnerable: Warning notices like 'Keep out of reach of children' are sometimes weakened by AI ('Keep out of reach' instead of 'Keep out of reach of children'). In Europe, precise warnings are legally required; inaccuracies endanger product liability. Reviewers should therefore check legal texts not only for linguistic correctness but for legal equivalence. A proven approach: Create a glossary with binding translations for clauses, paragraphs, and legal terms. Work with a native speaker who knows the legal system of the target country. Finally, have a legal expert proofread it. AI can speed up the workflow, but liability remains with the company – therefore, systematic human review is indispensable.
Hreflang Tags and SEO: The Invisible Pitfalls
While linguistic errors catch the eye immediately, invisible errors in the source code endanger the findability of your multilingual website. Machine translations often deliver correct texts but forget to adjust hreflang tags, meta data, and URL structures. A typical example: the AI translates the content, but the hreflang attribute still points to the source page – search engines then display the wrong language version in search results. Also, the translation of meta titles and descriptions is often neglected, so your page reads in German but appears under English snippets. Even more critical: if the AI translates page titles but leaves the URL unchanged, duplicate content issues arise. Reviewers should therefore systematically check each page for complete localization of SEO elements: hreflang, canonical, meta description, alt texts, and URLs. The devil is in the source code: a missing language code or an incorrect region (de-DE instead of de-AT) can ruin your international visibility. Our review therefore covers not only the visible text but also the technical infrastructure of localization. Only this way can you ensure that your content is found in all target markets – without ranking losses due to machine negligence.
Terminology Management as a Quality Anchor
One of the biggest challenges in machine translation is the inconsistent use of technical terms. While a single text may appear flawless, comparison with previous translations from the same client often reveals serious discrepancies. The AI has no memory for company-specific vocabulary – it translates each occurrence anew without considering the established lexicon. Therefore, centralized terminology management is essential. Before translation, create binding glossaries that define preferred terms, forbidden synonyms, and context-dependent variants. This glossary must be integrated both into the AI environment and the review environment. A reviewer should then systematically compare each technical term in the target text against the glossary. Especially critical are neologisms, product names, and legal formulations. Without a glossary, you risk that the same term is translated differently across various publications – a missed opportunity for brand consistency and an annoyance for your customers. Invest in maintaining your terminology base; it is the most effective lever to raise the quality of machine translations to a professional level.
blog.faqT
How do I handle incorrectly translated proper names?
Create a binding terminology database that defines the correct spelling of each proper name in the target language. Reviewers should mark all proper names and compare them against this database. Deviations are corrected and the database is updated when new cases arise. This ensures that brand and product names remain consistent.
Which tools support checking for number and formatting errors?
Use QA tools such as Xbench or Verifika, which automatically check for thousands separators, date formats, and currencies. These tools list discrepancies that then need to be validated manually. Alternatively, regular expressions can handle custom checks. The results are recorded in the inspection log.
