Hindi Gender Agreement Between Verb and Subject in AI Text

Hindi marks gender on the verb, not just on the noun and the adjective. That gives generated Hindi two ways to be wrong that European languages do not have — and in the perfective, the verb may not be agreeing with the subject at all.

The output you are looking at

English   I went to the office yesterday.

Hindi     मैं कल दफ़्तर गया।     (speaker male)
          मैं कल दफ़्तर गई।      (speaker female)

English   The boy ate the bread.

Hindi     लड़के ने रोटी खाई।
          ↑ subject marked with ने     ↑ verb is FEMININE, agreeing with रोटी

The second example is the one that surprises people. The subject is masculine, the verb is feminine, and the sentence is correct. Any approach to reviewing Hindi output that checks “does the verb match the subject” will flag it as an error and will miss the real errors elsewhere.

The basic rule, and where English gives no answer

In the imperfective, the future and the progressive, the verb agrees with the subject in gender and number, and so does any adjective in the sentence:

मैं जाता हूँ।       I go.        (m)     मैं जाती हूँ।       (f)
मैं जाऊँगा।         I will go.   (m)     मैं जाऊँगी।         (f)
वह अच्छा छात्र है।   He is a good student.
वह अच्छी छात्रा है।  She is a good student.
हम काम कर रहे हैं।   We are working. (m/mixed)
हम काम कर रही हैं।   (all-female group)

Now count what the English source provides. “I will go” contains no information about the speaker’s gender, and Hindi requires it in the verb — there is no unmarked option. The same is true of “you”: आप गए versus आप गईं, on top of the separate honorific choice between तू, तुम and आप. So a large share of Hindi gender errors are not errors of agreement at all; they are the model filling a slot the English left empty, and it fills it with the masculine because that is the more frequent form.

This has a specific consequence for products: a chatbot speaking Hindi in the first person must choose a gender for itself in every past-tense and future-tense sentence it produces. मैंने देखा or मैंने देखी, मैं मदद कर सकता हूँ or कर सकती हूँ. There is no neutral form to fall back on, and an assistant that drifts between the two across a conversation reads as badly as one that drifts between formal and informal address. This is a decision to make once, in the system prompt, not a bug to fix later.

The ergative rule: the verb changes allegiance

Hindi is split-ergative. In the perfective aspect with a transitive verb, the subject takes the postposition ने, and once it does, it stops controlling agreement. Three cases, and the model has to decide which it is in before it can choose an ending:

• Intransitive perfective — verb agrees with the subject. लड़की गई (the girl went, feminine), लड़का गया (the boy went). No ने.
• Transitive perfective with an unmarked object — verb agrees with the object. लड़के ने रोटी खाई — subject लड़के ने, object रोटी (feminine), verb खाई (feminine). Swap the object and the verb changes with it: लड़के ने खाना खाया, with खाना masculine and खाया masculine, even though the subject has not moved.
• Transitive perfective with an object marked by को — verb agrees with nothing. लड़के ने लड़की को देखा. Both arguments are marked, so the verb takes the default masculine singular देखा regardless of the gender of either participant.

Three different controllers for the same verb slot, selected by aspect and by two postpositions. To get this right a model must identify the aspect, detect ने, detect whether the object carries को, and then find the gender of whichever noun ends up in control. Each of those is a parse step, and the English source signals none of them — English has no ergative marking, so there is nothing to translate the cue from.

The characteristic failure is therefore a plausible verb form agreeing with the wrong argument, most often the subject, because subject-agreement is the pattern everywhere else in the language and in the source. In a sentence where subject and object happen to share a gender, the error is invisible — which means the error rate you see when reviewing is lower than the rate at which the rule is being misapplied.

What Urdu has to do with it, and what it does not

It is often suggested that Hindi agreement errors come from a model’s Urdu data pulling it toward a different pattern. That explanation does not hold, and it is worth saying why, because it sends people looking in the wrong place.

Hindi and Urdu share this grammar almost exactly. Urdu has the same split ergativity, marks it with the same postposition written نے, and applies the same object-agreement rule in the perfective. The two standards differ in script (Devanagari versus the Perso-Arabic Nastaʿlīq), in high-register vocabulary (Sanskrit versus Persian and Arabic sources), and in some idiom — not in agreement.

What the Hindi–Urdu relationship does cause is a different class of problem: register and lexical drift. A model asked for Hindi may return Persian-derived vocabulary that a Hindi reader perceives as Urdu, or a formal register that does not match the request. And romanised Hindi-Urdu, which is a large share of what exists on the web, is script-ambiguous, so it feeds both standards from the same text. If the complaint about your output is “this does not sound like Hindi”, it is likely lexical. If the complaint is “the verb ending is wrong”, it is the ergative rule, and it comes from parsing, not from Urdu.

Fixing it

1. Fix the first person once, in the system prompt. State the assistant’s grammatical gender and the address level (आप unless you have a reason) so that every generated sentence is consistent rather than independently guessed.
2. Pass the user’s gender as data when you have it and as an explicit “unknown—use plural honorific forms” instruction when you do not. आप गए हैं with honorific plural agreement is the standard way to address someone whose gender you have no business assuming.
3. Translate whole sentences, never fragments. A pipeline that sends a verb phrase without its arguments has removed the noun the verb is supposed to agree with, and the model will supply the default.
4. Build the review checklist around the ergative rule. Search the output for ने, and for each occurrence check what the verb agreed with, rather than checking subject agreement globally. This is where the systematic errors are.
5. Prefer a glossary over correction. If your domain has recurring nouns, state their gender in the prompt — रिपोर्ट (f), खाता (m) — because agreement errors downstream are frequently caused by the model assigning a gender to a loanword differently in different sentences.