---
title: Why linguists keep writing about John and Mary
date: 2026-10-15
summary: In linguistics papers, 30% of male names are John and 31% of female names are Mary. A good list of names doesn't fix that, because people still pick from it. A random generator does.
---

Open a linguistics paper and look at the example sentences. "John thinks
Mary left." A survey of 927 papers in three major journals found that
30% of all male names were John and 31% of all female names were Mary.
Male arguments showed up twice as often as female ones, and they were the
ones doing things. Textbooks look the same, and they haven't changed much
since the 1990s.

Nobody chooses this. When an example needs a name and you're at a
whiteboard, you reach for the first one in your head, and the first one in
your head comes from the culture you grew up in.

## a list isn't enough

The obvious fix is a curated list of diverse names, and good ones already
exist. But choosing from a list is still choosing. Over a semester the
same unconscious pull comes back, even with a balanced list and good
intentions.

So with Maura O'Leary and Rainey Williams we built the
[Diverse Names Generator](https://diversenamesgenerator.com/), which takes
the choice away. You press generate, or swipe a card in the Android app,
and get a random name from a database linguists curate. The bias never
gets a chance to act.

## making random usable

Random only works if people actually use it in a real classroom, so every
name comes with what an instructor needs at that moment:

- an IPA transcription, so you can say it confidently
- its usual gender association and the languages it comes from
- sometimes a recording, its native script, and famous people who share it

Filters make sure the randomness still fits the example. Semanticists need
one name per first letter so the initials don't collide in logic notation,
syntacticians need to match gender features, and one filter keeps to
English phonotactics for instructors worried about unfamiliar clusters.
Inside those constraints the pick is still random. The app downloads the
filtered set once with a single database call, then generates from that
local list, so it works offline in a classroom with bad wifi and never
shows the same name twice in a row.

The names come from users and students. Phonetics instructors can give
their students credit for submitting a name with a recording, which turns
an IPA exercise into something that ends up in front of other classes.
Linguists check every submission by hand before it goes into the pool.
When we published, the database had over 200 names from more than 30
languages.

## the point

The bias here isn't in anyone's values, it's in which name comes to mind
first. You can't train that away, but you can design it out. When the
default changes, the examples change.

The paper is in the
[*Proceedings of the LSA*](https://journals.linguisticsociety.org/proceedings/index.php/PLSA/article/view/5541)
(2023), and we presented it at the LSA 2023 Annual Meeting.