Grammatical Gender in Generated Job Titles Across Languages

The text said “Marie is a researcher” and you got Marie est un chercheur. The CV said “she” four sentences earlier and the German came back Ingenieur rather than Ingenieurin. These are two different bugs and only one of them is the model’s.

The output you are looking at

Job titles concentrate this problem for a reason that has nothing to do with the model’s opinions. In most gendered languages, an occupational noun has a masculine and a feminine form; the masculine is also the historical generic, so it is far more frequent in any corpus; and the gender cue that would select the feminine is often distant from the title, in a different field, or absent entirely. High-frequency default plus weak, distant evidence is a reliable recipe for the default winning.

The published evidence for the direction of the default is the WinoMT challenge set from Stanovsky, Smith and Zettlemoyer (ACL 2019), which constructed sentences where the correct gender of an occupation is unambiguous from syntax and measured whether translation systems used it. They found systematic reversion to the stereotypical or masculine form across the systems tested. Treat that as evidence about the mechanism rather than as a current figure for any specific model.

French: a form exists for almost everything now

The gap between “there is no feminine form” and “the model did not use it” closed some time ago in French. The Académie française adopted a report on the feminisation of names of trades and functions on 28 February 2019, concluding that there is no obstacle in principle to feminising them. Quebec’s Office québécois de la langue française had recommended it since 1979, which is why Quebec French usage is decades ahead and why a model trained on mixed francophone data will produce both conventions in one document.

masculine        feminine           formation
professeur       professeure        + e
ingénieur        ingénieure         + e
auteur           autrice            -teur → -trice  (auteure also current)
directeur        directrice         -teur → -trice
chercheur        chercheuse         -eur  → -euse
vendeur          vendeuse           -eur  → -euse
pompier          pompière           -ier  → -ière
écrivain         écrivaine          + e
médecin          médecin / la médecin   (no accepted feminine form)
journaliste      journaliste        epicene — only the article changes

Three patterns are worth separating because they fail differently. The -teur nouns split into -trice and -teuse along lines a model has to have memorised, and where two feminines compete — auteure and autrice — output is inconsistent rather than wrong. Epicene nouns such as journaliste carry the gender only on the article, so an error is invisible in the noun and appears in le/la and later in the pronoun. And a handful, including médecin, still have no settled feminine, so Marie est médecin is correct and a model “fixing” it to médecine has produced the word for the discipline.

German: the suffix is regular, the umlaut is not

German derives the feminine with -in, which is regular enough that a model rarely fails to produce a plausible form. What it fails on is the stem change and the plural.

Lehrer      → Lehrerin      → Lehrerinnen        regular
Ingenieur   → Ingenieurin   → Ingenieurinnen     regular
Arzt        → Ärztin        → Ärztinnen          umlaut
Koch        → Köchin        → Köchinnen          umlaut
Bauer       → Bäuerin       → Bäuerinnen         umlaut
Kaufmann    → Kauffrau      → Kaufleute          -mann/-frau, suppletive plural
Beamter     → Beamtin       → Beamtinnen         adjectival noun

The umlaut set is small but common, and the plural doubles the n, so a wrong feminine is often wrong twice. The -mann compounds are the interesting case: the feminine is not Kaufmännin but Kauffrau, and the mixed plural is neither — it is Kaufleute, using the suppletive -leute. Related neutral formations exist and are widely used in German job advertisements: Fachkraft, Lehrkraft, Pflegekraft, and the (m/w/d) tag appended to a title.

Beyond that sit the typographic conventions — Lehrer*innen, Lehrer:innen, Lehrer_innen, LehrerInnen — which are style choices, not grammar, and which a model will mix within one document unless you pin one. They also interact badly with downstream processing: the colon and asterisk variants break naive tokenisation, sorting and screen readers in different ways, so pick one at the source rather than normalising later. The agreement chain that a job title then controls is in German grammatical gender agreement in AI-generated text.

Why structured data is worse than prose

In running prose the model usually has something to work with: a name, a pronoun, an agreement earlier in the paragraph. The systematic failures happen where that context has been engineered away.

• Field-by-field translation. A CV, a directory or a product catalogue translated a row at a time gives the model the string Researcher and nothing else. There is no cue to ignore, so “the model got it wrong” is a misdiagnosis — it answered the question it was asked.
• The cue is in a sibling field. The gender may be in a salutation or pronouns column that your prompt does not include, three fields away in the same record. This is the most common and most fixable version.
• Inference from a name is a second guess. Names carry gender unreliably across cultures, and a model inferring from a name and then agreeing with its inference produces confident, consistent, wrong output — which is harder to spot than inconsistent output.
• Generic listings genuinely have no answer. A job advertisement is addressed to nobody in particular, and the correct output is a convention — Ingenieur (m/w/d), ingénieur ou ingénieure, Fachkraft — not a gender. Decide the convention once, in the prompt.

Fixing it

1. Find out whether the cue was in the input at all. Re-run one failing example with the gender stated explicitly. If it comes back correct, this is a context-assembly bug in your pipeline, not a model choice, and no prompt engineering is the right fix.
2. Pass gender as a labelled parameter alongside the text — subject_gender: f — rather than relying on a name or a pronoun buried in prose.
3. Supply a glossary for the titles you use repeatedly. A short table of masculine and feminine forms in the prompt removes both the -trice/-teuse guess and the umlaut guess, and makes output consistent across documents in a way that instructions alone do not.
4. State the convention for the unknown case explicitly — feminine when known, doublet when addressing an audience, neutral noun for advertisements — so that the model is not choosing a policy per row.
5. Check the pronouns and adjectives, not just the noun. A corrected title with the original masculine article or participle still around it is the usual result of a targeted fix, and is worse than the original because it looks reviewed.