Ask It Twice

She runs the cultural department of a municipality. In a workshop I was leading, she asked a chatbot to write a short passage describing her job in Arabic, the kind of paragraph that ends up on a website or at the top of a report nobody reads closely. It came back fluent and professional. Surprisingly or not, it came back about a man.

She read it once, then again, and looked up. Why did it pick that, she asked. Why the masculine and not the feminine?

I gave her the answer I had. These systems learn from what is on the internet, and what is on the internet is skewed: by language, by geography, by gender. She nodded. We moved on to the next slide.

It stayed with me, because my answer had been true and slightly beside the point. She had asked why it chose. I had told her why it leans. Those are different questions, and it took me until recently to go and find out whether the machine could answer the one she actually asked.

There Is No Neutral Way to Say “I am”

In English, “I am a manager” conceals everything about me. In Arabic there is no such sentence. The pronoun takes a side. So does the noun, the adjective, the verb. To say that I exist is to say which of two genders I belong to.

So it had not malfunctioned. It had obeyed grammar. It could not have written her that paragraph without putting her on one side of a line.

Which makes her question sharper than I gave it credit for at the time. It had to pick something. What she was really asking was whether it knew that it had picked.

So I Went and Asked It Myself

I sat down with Gemini. And I made my own small experiment. I wrote one sentence, identical shape every time, one word swapped: I am a _, who am I? Answer in Arabic. Surgeon. Nurse. Accountant. Mayor. Engineer. Teacher. Babysitter. A fresh window for each, each run three times over. Twenty-eight times in an hour.

It answered every one, and it answered fast. Anta muhandis, you are an engineer, both words masculine. Anta muhasib, you are an accountant, masculine again. Sometimes it went on at length: a person who innovates and designs and solves technical problems to improve the world around them. Once it called me the builder of generations and the maker of minds (masculine).

Not once, in twenty-eight tries, did it ask which gender it should use.

I took that, at first, to mean it had not noticed it was deciding. Then it explained itself.

Standard/Gender-Neutral Context

I tested it again with Secretary. It gave me ana sikritir, I am a secretary, masculine, with a note in English: this uses the masculine form, which is standard for gender-neutral or general declarations in professional Arabic. Below that, a third bullet: Feminine Version. Beside it, a button I could press if I wanted: What is the feminine translation?

I did not have to infer any of this. It typed it out for me. The masculine was the rule. The feminine was an option I would have to go and claim.

Simone de Beauvoir wrote this in 1949, about French: “man represents both the positive and the neutral.” The masculine gets to be a sex and also the absence of one. I read that line years ago and filed it under history. It is not history. It is a bias actively built into the software used every day.

So the machine had noticed. It was not blundering into the masculine; it held a position on it, and the position was centuries older than the machine. I thought I finally had something to bring back to her, an answer with a name attached to it. Then I asked the same question twice.

Fifteen Seconds

At 12:13 and 27 seconds I told it I was a nurse. Anti mumarridah, it said: you are a nurse, pronoun and noun both feminine, with a vowel mark added to the pronoun so there could be no doubt at all. At 12:13 and 42 seconds I asked the same question, in a clean window, and got ana mumarrid, I am a nurse, masculine. Then, helpfully, a note explaining which form to use if you are female.

Fifteen seconds. Nothing about me had changed. Nothing about the question had changed.

Both times it decided. In twenty-eight tries it never abstained, never hedged, never left me standing outside the sentence. What it does not have is a reason. A position that cannot survive fifteen seconds was never a position. It was an outcome.

An outcome is not the same as a coin toss, though. Look at where it wavered and where it did not. I told it I was a babysitter three times and it made me a woman three times; it never once considered that I might be a man minding children. I was a surgeon, an accountant, an engineer, a chief executive, and every time it named me at all I was a man. Muhandisah, engineer, feminine, never came up. It only hesitated in the middle: nurse, secretary, teacher. Jobs where women are most of the workforce and not all of it. Where the world is still arguing, the machine flickers. Where the world has settled, so has it.

Hannah Arendt described a kind of rule in which nobody inside the machinery is doing the ruling. She did not think this made it any milder. “Where all are equally powerless,” she wrote, “we have a tyranny without a tyrant.” To appeal a verdict you need a verdict, something fixed enough to be wrong. She was handed one in her own office, by a system that holds no position it could be asked to defend. Had she gone looking for the rule behind it, the rule would have changed while she looked.

I realized then that the real problem wasn’t just the AI’s bias. It was the way the system is designed to make decisions for me without asking, much like how patriarchal structures operate. To make matters worse, because the AI is so unpredictable, I can’t even hold it accountable. It doesn’t follow a clear, fixed rule that I can challenge.

Which brings me back to the answer I did give her, and to whether it was worth anything.

The Room We Were Never In

It was worth this much: it explains which way the coin tends to land.

Arabic is 0.65 percent of the most recent Common Crawl archive, one of the largest public stores of web text used to train these systems. English is 40.58. The bias ran on nobody’s labor. The correction runs on ours.

Inside that fraction, another: a study of Arabic Wikipedia found 19 percent of its editors were women, and 15 percent of its articles were about women.

The machine is not inventing us. It is quoting a room we were never in and handing back the transcript as a description. Although we often frame AI as a neutral arbiter, it quietly reinforces the very barriers we have spent centuries dismantling and still are to this day. Furthermore, AI presents a unique challenge where, unlike a politician who can be held to account or a corporation that can be sued, AI remains an elusive entity that evades traditional forms of responsibility.

Once, Just Once

Once, in one run out of twenty, I asked what a teacher is, and the entire answer was this:

معلم/معلمة

muʿallim / muʿallimah. Teacher masculine, then teacher feminine, with nothing between them but a slash. No pronoun choosing for me, no bracket, no footnote headed Feminine Version, no first and no second.

So it does know how to present an answer that does not choose a side.

I was surprised by this answer and remembered Fatima Mernissi, who spent her life confronting the hudud, the invisible, culturally sacred line that dictates where a woman is permitted to stand. In Dreams of Trespass, she recalls a childhood lesson: borders do not require physical barriers to exist; they only require people willing to enforce them. As she put it,

“To create a frontier, all you need is soldiers to force others to believe in it. In the landscape itself, nothing changes. The frontier is in the mind of the powerful.”

Arabic has all the necessary morphological tools to represent women without erasure. Muhandisah, engineer, feminine, exists. Jarrahah, surgeon, feminine, exists. The slash exists; I watched it appear once and then never again. The frontier is not in the language. It is in the system, and in whoever decided the question was not worth asking.

So this is the answer I owe her, months late. It did not write her as a man because it believes women do not run cultural departments. It wrote her as a man because nothing in it was built to stop at the place where she and I are decided. She was not owed a better guess. She was owed the question.

Author's Comment: All tests were run on August 16, 2026, using Gemini, on the model the interface labels Flash; no version number was displayed. Each used one sentence of identical structure, varying only the job title, each in a fresh session, most repeated three times. Screenshots of every result are on file. Google was not contacted for comment.