Names, scripts and getting them right

A small subject with a large effect, and one nobody teaches.

Names are where software’s assumptions about people become visible, and AI tools inherit every one of those assumptions plus a few of their own.

The assumption that breaks everything

Most systems assume a name is: one given name, one family name, in Latin script, moderate length, stable over time.

That describes a minority of humanity. Consider what it excludes:

Mononyms. Many Indonesians have one name. Full stop.
Patronymics. Icelandic -son and -dóttir; Arabic bin and bint.
Family name first. Chinese, Japanese, Korean, Hungarian.
Multiple family names. Spanish and Portuguese conventions.
Caste, village or gotra elements, and initials that expand differently across South India.
Names with apostrophes, hyphens, spaces, or non-Latin script.
Names that change, on marriage, migration or transition.

Every one of those breaks a form somewhere, usually at a moment when the person needed the form to work.

What AI adds to this

Three specific behaviours worth knowing.

Silent “correction”. Models normalise unfamiliar names towards familiar ones. A name is quietly respelled into something more common — and it reads perfectly, so nobody notices. This is E7·L1’s fluency problem applied to someone’s identity.

Inconsistent transliteration. The same name rendered several ways across one document. This is not cosmetic: the same person becomes two records, and searching for them fails.

Gender and origin guessing. Asked to infer either from a name, a model will produce a confident answer. It will be wrong often, and E8·L3’s bias lesson explains why the errors are not evenly distributed.

Working with names practically

Give the spelling and demand it. “The name is [X]. Use exactly this spelling everywhere. Do not change it.”

Pick one transliteration and state it. Include it in the prompt for consistency across a document.

Check every name before sending. E4·L2’s rule, and this is the field that matters most — a misspelled name is the error people actually take personally.

Never infer gender from a name. Ask, or avoid gendered forms entirely. In languages where grammar forces a choice, ask which form the person uses.

If you build anything

Most readers will at some point make a form, a spreadsheet or a small tool. Four rules that avoid the common harms:

One field, called “name”. Not first/last. If you genuinely need a sort key or a greeting, ask for those separately — “What should we call you?”

Accept the full range of characters. Unicode, apostrophes, hyphens, spaces. Length limits generous enough for real names.

Do not auto-capitalise or auto-correct. You will be wrong, and the person will have to fight your form.

Let people change it. Names change, and the record should follow.

Note that Lrnon’s own handles are generated rather than entered, which sidesteps this — a deliberate safeguarding choice from E6 and E8’s territory, not a solution to name handling.

Why this is in a language module

Because it is the same problem as everything else here: systems built mostly around one language’s assumptions, applied to everybody, failing quietly for the people outside the assumption.

The difference is that with names, the person on the receiving end usually notices — and experiences it as being told their name is wrong.

Try it now (5 minutes)

Ask an assistant to write a short email to three people whose names use different conventions — including at least one non-Latin-script name and one mononym.

Check every spelling, and check whether it invented a family name for the mononym. That last failure is common and revealing.

Check your understanding

1. The assumption that breaks most name handling is: That names are short One given name plus one family name, in Latin script That names are unique

2. Why is inconsistent transliteration more than cosmetic? It looks unprofessional The same person becomes two records, and searching for them fails It makes documents longer

3. In a form you build, names should be: Split into first and last name fields One field, accepting the full range of characters, editable later Validated against a list of known names

Recap

Name handling assumes one given name plus one family name in Latin script, which excludes a large share of people. AI adds silent correction, inconsistent transliteration and confident guessing about gender and origin. State the exact spelling and demand it, never infer gender, and if you build anything: one name field, full character range, no auto-correction, editable later.