Writing Gender-Neutral Text in Spanish With AI

“Write this in gender-neutral Spanish” is four different instructions, and a model asked that way will use all four in one document. Pick one, state its rewrite rule, and the output becomes consistent enough to review.

The four conventions, and what each changes

Spanish marks gender on determiners, nouns, adjectives and some pronouns. Any neutral convention has to say what happens to each of those slots, and the four in circulation answer differently.

source:  Todos los alumnos están contentos con sus profesores.

-e:      Todes les alumnes están contentes con sus profesores.
-x:      Todxs lxs alumnxs están contentxs con sus profesorxs.
-@:      Tod@s l@s alumn@s están content@s con sus profesor@s.
rewrite: El alumnado está contento con el profesorado.
         (or: Quienes estudian aquí están a gusto con quienes enseñan.)

• The -e ending is the only one that is pronounceable and the only one with a matching pronoun set — elle, elles, and the article le/les in the fuller versions of the system. It is the convention with actual spoken currency, mostly in Argentina, Chile and Uruguay.
• The -x ending is written-only and unreadable aloud in Spanish, since lxs has no syllable. It travelled from US English usage (Latinx) rather than from Spanish-speaking countries.
• The @ symbol is the oldest of the four and explicitly a both rather than a neither: the glyph is read as an a inside an o. It is invisible to screen readers and breaks search and sorting, so it is now rare in professional copy.
• Circumlocution — collective nouns (el alumnado, la plantilla, el personal, la clientela), relative clauses (quienes soliciten), and impersonal constructions (se recomienda) — changes no morphology at all. It is the only option that produces text nobody can object to on grammatical grounds, and it is the one most style guides for institutional Spanish actually recommend.

The words no convention covers cleanly

The -e rule is usually stated as “replace final -o/-a with -e”. That rule has three holes, and they are where output degenerates.

• Nouns already ending in -e are epicene already: estudiante, presidente, representante, cliente. Nothing to change — but a model applying the rule mechanically can produce back-formations, or can over-correct presidenta (a standard, RAE-accepted feminine) into something nobody writes.
• Nouns in -or form the feminine in -ora (profesor/profesora), so the neutral is profesore — an extra syllable, not a substitution. Likewise trabajador → trabajadore. Models get this one wrong more than any other because it does not fit the swap-the-last-letter pattern.
• Words where -e collides with an existing word. The neutral form is not always free; and irregular pairs (hombre/mujer, padre/madre, rey/reina) have no morphological neutral at all. These require a lexical choice — persona, madre o padre, familiar — which is a decision your prompt has to make in advance.

Why the model drifts back to the masculine

You will see the first two sentences honour the convention and the fifth revert. The mechanism is ordinary: todes les alumnes is a rare token sequence and todos los alumnos is an extremely common one, so the moment the immediate context stops reinforcing the instruction, the higher-probability continuation wins. Length makes it worse, and so does anything that pushes the instruction further back — a long retrieved document, a big few-shot block, a multi-turn conversation.

The practical consequences are that few-shot examples beat prose instructions, that examples placed near the end of the prompt beat examples at the top, and that a long output should be generated in sections rather than in one pass. This is the same drift you get from any low-frequency style constraint; it is not specific to inclusive language.

Build the prompt

1. Choose one convention and name it in the system prompt. Not “inclusive Spanish” — say “terminación en -e”, or “lenguaje no sexista por reformulación, sin -e, -x ni @”.
2. State the rewrite rule as a closed transformation, including the -or case, and state what is left alone: proper nouns, quotations, existing epicene nouns, and words with a lexicalised feminine such as presidenta.
3. Give a substitution table for the words with no neutral form — hombres y mujeres → personas, los padres → las familias, el usuario → la persona usuaria — rather than leaving the model to invent one per occurrence.
4. Add three to five worked pairs of source and target immediately before the input, at least one containing an -or noun and one containing an already-epicene noun.
5. Generate in sections for anything over a few hundred words, re-sending the rule with each section.

SYSTEM
Reescribe el texto en español con lenguaje inclusivo por terminación en -e.

Reglas:
1. Sustantivos y adjetivos con -o/-a de género → -e:  alumno/alumna → alumne.
2. Sustantivos en -or → -ore:  profesor/profesora → profesore.
3. Artículos y determinantes: el/la → le, los/las → les, un/una → une.
4. NO cambies: nombres propios, citas textuales, sustantivos ya epicenos
   (estudiante, cliente, representante), ni femeninos lexicalizados
   (presidenta, jueza).
5. Sin -x, sin @, sin desdoblamiento ("los y las").
6. Si una palabra no admite forma en -e (hombre, madre, rey), usa la
   sustitución de la tabla; si no está en la tabla, reformula la frase.

Tabla: hombres y mujeres → personas | los padres → las familias |
       el usuario → la persona usuaria

Devuelve solo el texto reescrito.

Check the output

Do not trust a second model call to grade this. The check is deterministic and cheap: scan the output for the endings you have banned and for the masculine plural articles you have replaced.

# any surviving masculine plural determiner, or a banned convention
grep -nE '\b(los|unos|aquellos|todos|nuestros)\b' out.txt
grep -nE '[a-zá-ú]+(x|@)s?\b' out.txt

# the -or trap: profesore is right, profesore→profesor(a) is a miss
grep -nE '\b\w+or(es)?\b' out.txt

A hit is not automatically an error — los datos is a grammatically masculine inanimate plural that most style guides leave alone, and that distinction is exactly the thing a regex cannot make. Use the scan to produce a short review list, not to fail a build. If your text is going to a Latin American audience across several countries, the register question sits alongside this one; see what “neutral Spanish” means in generated text and how to pin formal and informal register in a prompt, since tú, vos and usted drift for the same reason the -e ending does.