German Grammatical Gender Agreement in AI-Generated Text

Generated German that says Ich habe einen neuen Wagen gekauft. Es ist rot. got the noun, the article and the adjective right and the pronoun wrong. That distribution is not random, and it points straight at the cause.

The output you are looking at

German marks gender in four places, and they are separated by different distances from the noun:

der  neue     Wagen,   den  ich  gekauft  habe,  ist  rot ;  er   fährt gut
ART  ADJ      NOUN(m)  REL  —    —        —      —    ADJ    PRON(m)

die  neue     Karte,   die  ich  gekauft  habe,  ist  rot ;  sie  liegt hier
ART  ADJ      NOUN(f)  REL  —    —        —      —    ADJ    PRON(f)

das  neue     Auto,    das  ich  gekauft  habe,  ist  rot ;  es   fährt gut
ART  ADJ      NOUN(n)  REL  —    —        —      —    ADJ    PRON(n)

The article and the adjective sit immediately before the noun, so a model producing them is making a local, heavily-attested choice — and gets them right nearly all the time, because der Wagen and das Auto are frequent bigrams. The relative pronoun is one clause away. The personal pronoun can be a sentence or a paragraph away, and often has to be recovered from an English source that gave no gender at all. So the error rate rises with distance, and the visible error lands on the pronoun.

The es trap: English “it” has no German equivalent

This is the largest single source of wrong gender in translated German and it is a property of the source language, not of the model. English uses it for every inanimate referent. German uses er, sie or es according to the grammatical gender of the noun being replaced. There is no default and no neutral option: every it in the English forces a three-way decision that the English text does not contain the answer to.

It is worse than a coin flip, because the model must first identify which noun the English it referred to — often the wrong one in a sentence with two candidates — and only then look up that noun’s gender. Two chances to fail, and the second is invisible in the output. The same pronoun problem also runs backwards: German sie is she, it (feminine noun), they, and capitalised Sie is formal you, so German-to-English inherits an ambiguity German-to-German never had.

Note also that natural German sometimes uses semantic rather than grammatical agreement for people: das Mädchen is grammatically neuter, and while es is the prescriptively correct pronoun, sie is common in real usage and increasingly accepted. A model producing sie here is following attested German rather than making an error, which is worth knowing before you “fix” it in review.

Adjective endings encode gender and case together

German adjective endings do not encode gender alone. They encode gender, number, case and the kind of determiner in front of them, all in one or two letters, across three declension patterns. That is why an error here is usually a case error rather than a gender error, and why telling a model “Wagen is masculine” does not fix it.

masculine "Wagen", nominative vs accusative:
  der  neue  Wagen      →  den  neuen  Wagen        (weak, after der)
  ein  neuer Wagen      →  einen neuen Wagen        (mixed, after ein)
  neuer Wagen           →  neuen Wagen              (strong, no determiner)

neuter "Auto" — nominative and accusative are identical:
  das  neue  Auto       →  das  neue  Auto
  ein  neues Auto       →  ein  neues Auto

Two things follow. First, masculine is the only gender where nominative and accusative differ, so der/den and ein/einen are where a case slip becomes visible — and a model that writes Ich habe der neue Wagen gekauft has made a case error that looks like nothing in particular. Second, feminine and plural share endings in the nominative and accusative (die, die neue / die neuen), so a feminine-plural confusion is silent in two of the four cases and appears only in the dative (der feminine singular versus den plural with an -n on the noun).

The consequence for review is that scanning generated German for “wrong gender” will miss most of the errors, because they are in the case slot of a shared ending.

When the gender is predictable from the word

German gender is arbitrary for basic vocabulary, but a large part of the lexicon is derived, and derived nouns take their gender from the derivation. These rules are near-exceptionless and are worth knowing because they tell you which words a model is unlikely to get wrong and which it will.

• Feminine: nouns in -ung, -heit, -keit, -schaft, -ion, -tät, -ik, -ei. die Verbindung, die Sicherheit, die Konfiguration, die Qualität. Almost all technical abstractions land here.
• Neuter: diminutives in -chen and -lein without exception (das Mädchen, das Fenster­chen), and any infinitive used as a noun (das Laden, das Speichern).
• Masculine: agent nouns in -er derived from verbs — der Server, der Nutzer, der Zeiger — and most days, months and seasons.
• Compounds take the gender of their last element, always. das Auto but der Autoschlüssel (because der Schlüssel); die Datei but der Dateiname (because der Name). This is the single most useful rule for technical German, where most nouns are compounds, and it means the gender question for a long compound reduces to the gender question for its short final noun.
• Loanwords are the unpredictable set. die E-Mail in Germany, das E-Mail in Austria and Switzerland; der Blog and das Blog both current. A model will produce the mixture that exists in its training data, which means inconsistency within a single document rather than a stable wrong answer.

What to do about it

Give the model the antecedent. Sentence-level translation removes the noun that er depends on; translate at paragraph level or larger and the pronoun errors drop for structural reasons rather than stylistic ones. If your pipeline must segment, carry the previous sentence as context even if you discard its output.

Constrain the noun before you constrain the pronoun. If you are generating text about a specific product or entity, put the German noun and its article in the prompt as a glossary line — der Wagen (m) — rather than letting the model choose a translation and then agree with its own choice. Half of the pronoun errors are really translation-choice instability: the model rendered car as Auto in one paragraph and Wagen in the next, and then agreed correctly with each.

Check deterministically where you can. A pronoun that disagrees with every noun in the preceding two sentences is findable without a model, and so is der/die/das against a lexicon. What is not machine-checkable is whether the model picked the right antecedent, and that stays a human review item. The parallel problem in French, where the agreement chain is longer but the case system is absent, is in gendered nouns in French AI-generated text.