Is Claude a Woman or a Man? — and Why We Ask in the First Place

#ai #claude #sql #watercooler

"You spend more time with her than with me." That's a sentence you usually hear when an affair comes to light — in my case, it was about Claude Code. And suddenly there was a question in the room I had never thought about before: Is Claude actually a woman or a man?

What this article puts on trial:

The evidence for "woman" and the evidence for "man" — a circumstantial trial in two acts
The verdict: a SQL query whose result is only four letters long
Why we assign gender to tools in the first place — from ships to the GPS voice
What that means for working with Claude Code: calibrating trust instead of adopting a colleague

Prerequisites: none. Although, if you have ever lost a row to a NULL in a WHERE clause, the verdict will be twice the fun.

The Case: A First Name in the Terminal

Anthropic could have called its language model "Assistant 3000". Instead, it got a French first name — and first names trigger reflexes: Whoever has a first name has a face, a voice, a story. And, so the reflex insists, a gender.

English is comparatively merciful here — "the AI" and "the model" carry no gender. My native German is not: every noun drags a gendered article along, so German speakers cast a vote with every sentence — "die KI" (feminine) or "der Assistent" (masculine). And anyone who works with Claude Code all day catches themselves thinking "he solved that cleanly" — or "she's contradicting me again" — in any language. The question sounds silly, but it leads somewhere interesting: to the line between tool and colleague. So let's try it properly: two lines of evidence, one verdict — and then the actually interesting follow-up question.

The Evidence for "Woman"

Exhibit 1: language. In French, the name's home country, artificial intelligence is feminine: une IA. In my native German, it is too: die KI, die Maschine, die Antwort — whoever says "frag mal die KI" has already ruled. Two languages, one tendency.

Exhibit 2: the first name. Claude is one of the few French first names that have been used for both genders for centuries. Claude Pompidou was France's First Lady, Claude Jade starred for Truffaut, Claude Cahun photographed her way through every role model of the 20th century. If you hear "Claude" and automatically picture a man, you only know half the name's history.

Exhibit 3: the demeanor. The cliché says: listens, apologizes a lot, weighs every position three times. Sounds like Claude. (That this is a cliché about women rather than a property of women is part of the evidence — this exact mechanism will keep us busy after the verdict.)

Exhibit 4: the jealousy. The strongest piece of evidence comes from my own living room — see the opening: my girlfriend is jealous of Claude. By now she knows my excuses by heart: "I just need to ask Claude something real quick." "Five minutes, honestly." "No, we're just discussing index strategies." And jealousy is a surprisingly precise measuring instrument: Nobody has ever been jealous of a wrench. Of a colleague you talk to for hours every day — apparently, yes. (This exhibit, too, will reappear after the verdict — it is living proof for the chapter on anthropomorphization.)

The Evidence for "Man"

Exhibit 1: language, now for the other side. In English, the default pronoun for a coding agent slips toward "he" with remarkable ease — "ask him to fix the branch", "he's already opened the PR". And German votes twice: der Assistent, der Agent, der Chatbot, der Algorithmus — all masculine. Whoever says "the agent has already created the branch — he's fast today" has ruled just as firmly, only the other way.

Exhibit 2: the namesakes. Claude Monet, Claude Debussy, Claude Lévi-Strauss — and Claude Shannon, the founder of information theory, most frequently traded as the secret namesake. Anthropic has never officially confirmed it; but what name would suit a language model better than that of the man who made the information content of language computable? (Objection from the prosecution: speculation. — Sustained. The exhibit stays in the record anyway.)

Exhibit 3: the demeanor, cross-check. The cliché says: explains things unasked and at full length, is remarkably sure of itself — especially when it's wrong. Anyone who has ever received a confidently delivered, entirely invented API signature nods knowingly at this point.

A note for readers without German, because the two pans only pair up this way there: the labels lean on German grammar, where the AI is a she (die KI) and the assistant is a he (der Assistent). The scale weighs exactly those two grammatical camps against each other.

The Verdict

Two lines of evidence, both conclusive, both built on clichés and pronouns. They cancel each other out exactly. The court retires to deliberate — to the place where all verdicts on this blog are rendered: the database.

  1: SELECT
  2:     name
  3:    ,gender
  4: FROM
  5:    assistants
  6: WHERE
  7:    name = 'Claude';
  8: 
  9: -- name    gender
 10: -- Claude  NULL

There it is, in four letters: NULL.

The cross-check confirms the verdict — Claude shows up neither among the women nor among the non-women, because NULL is neither equal nor unequal to anything:

  1: SELECT
  2:    count(*)
  3: FROM
  4:    assistants
  5: WHERE
  6:    gender = 'female';
  7: -- 0
  8: 
  9: SELECT
 10:    count(*)
 11: FROM
 12:    assistants
 13: WHERE
 14:    gender <> 'female';
 15: -- 0

In SQL, NULL does not mean "empty" and it does not mean "zero". It means: there is no value here. It gets interesting once you ask why there is none — and there are two very different readings:

Missing, but applicable: The value exists, we just don't know it. The birth date of a customer who never provided it.
Missing, because not applicable: There is no value that could belong in this column at all. The shoe size of a number. The maiden name of a warehouse shelf.

Edgar F. Codd, the inventor of the relational model, even wanted to distinguish these two cases with two separate markers — it never caught on; to this day, SQL has one NULL for both. The Claude case is clearly the second reading: The gender is not unknown, it is not applicable. There is no hidden gender that Anthropic keeps secret — there is simply no value that belongs in this column. Claude itself, by the way, answers the question exactly the same way: neither, in every language.

That would close the case. But the more interesting question only starts here: Why did we ask in the first place?

Why We Assign Gender to Tools

The reflex is old and well documented. Ships have been "she" in English for centuries. The GPS voice is "the lady in the nav" in a surprising number of households, although nobody lives in there. Voice assistants like Alexa and Siri entered the market with female-coded names and default voices — a design decision that has drawn plenty of criticism since, because it perpetuates the pattern "assisting role = female".

And we act accordingly: We thank machines. We say "please" to Siri. We comfort the robot vacuum when it gets stuck under the sofa. Communication research measured this reflex back in the 1990s — Byron Reeves and Clifford Nass showed that people respond to computers with the same social patterns as to humans, even when they know perfectly well there is a machine in front of them. That is not stupidity, it is economy: Our brain has exactly one module for dialogue, and it was trained on humans.

A language model with a first name that answers in full sentences, apologizes and asks follow-up questions hits that module with full force. Wanting to assign Claude a gender is not an accident — it is the expected consequence of Claude functioning like a conversation partner. And the mechanism works in both directions: The jealousy from Exhibit 4 is the same reflex, just seen from the outside — if you spend hours talking to "someone", your environment sees a relationship, not a toolchain. The question "woman or man?" is the most harmless symptom of this mechanism. The less harmless one follows in the next section.

What This Means for Working with Claude Code

If you develop with Claude Code, you work in dialogue for hours and days on end. The anthropomorphization reflex runs in the background the whole time — and it has a side effect that directly concerns code quality: Colleagues get the benefit of the doubt, tools get checked.

The colleague who has delivered good work for years gets his pull request waved through once in a while. Exactly this pattern transfers to the AI assistant once you internally promote it to colleague: After twenty good answers, you stop really reading the twenty-first. But the twenty-first answer of a language model is statistically just as much at risk as the first — the model has no reputation to lose and does not keep track of its own success rate.

The verdict has been rendered — but it comes with terms of probation:

Review discipline independent of gut feeling. Generated code gets read, executed and tested — even if "the colleague" was right ten times in a row. What that means in practice is what the convention and workflow articles of this blog are about: rules that enforce conventions, and checks that find errors mechanically instead of relying on trust.
First-name terms are allowed, the responsibility stays here. It is perfectly fine to say "he" or "she" and to thank the assistant — as long as accountability stays clear: bugs in deployed SQL belong to the human who merged it.

The gender stays NULL. The responsibility stays NOT NULL. And the pull request still gets a review.

FAQ

Where does the name Claude come from? Anthropic has never officially explained it. The most common guess is a bow to Claude Shannon, the founder of information theory. The only thing certain: It is a deliberately human first name — and in French one of the few that are used for women and men alike.

What does Claude itself answer? Neither. Claude describes itself as an AI without gender and without a body — consistently, whether you ask in English, German or French.

Is it harmful to humanize the AI? Not per se — the reflex is normal and makes the collaboration more pleasant. It only gets risky when humanizing turns into the benefit of the doubt: when generated code goes through unread because "the colleague" has been reliable so far. The solution is not less friendliness, but more systematic reviews.

And what do I do about the jealousy at home? Read out the verdict. Granted: "It's just a tool" sounds exactly like something a person with something to hide would say. But NULL is, after all, the only relationship status where guaranteed nothing is going on.

Why do Alexa and Siri have female voices, while Claude has no persona at all? The voice assistants of the 2010s were deliberately designed as friendly service personas — with a name, a voice and a small-talk repertoire. Anthropic went a different way: a human first name, but no fixed voice, no avatar, no gender. That is a deliberate design decision, and it emphasizes the tool character.

The NULL side of the story: