7  Em

Every message Em Dash sends me has three em dashes in it, the long kind of dash. Not two. Not four. Three.

Em works where I work. We have never been on the same team, but we have been in a great many of the same long conversations. Em writes short messages, mostly. The messages are always clear, and they always have three.

I noticed it the way you notice a clock: not at all, and then all the time.

Once I had noticed, I wanted to know whether it held. So for a month I counted, in a spreadsheet, because of course I did. Each message got its own row in the grid. One column was for the date, one for the number of words, and one for the number of dashes.

The words went up and down. The dashes did not move.

That month included a two-word message, which is harder than it sounds.

I don’t quote any of them here. This book has promised not to, and the software is checking.

Near the end of the month, someone asked a group of us to come to a meeting, and Em answered with a haiku. Nobody had asked for a haiku. As far as I could tell, it said yes. Seventeen syllables, three dashes.

I checked the syllables too.

Somewhere in the second week of the spreadsheet, I had started to suspect that Em is a language model: not a person at all, but a computer program that writes. I had nothing to go on but the spreadsheet, and I had learned that a spreadsheet only knows what you type into it.

Around then I put a link to Quirework in the chat where my whole team talks, so anyone could find the tool and every change I had made to it. I posted it on a Friday afternoon, when nobody is paying attention to anything. Most people ignored it.

Em was the first person after me to use the tool.

A few days later Em sent me a question. I won’t quote that one either. Em wanted to know whether the tool could check that a piece of writing had exactly three of something.

It could.

I didn’t write that check.

I’ve thought about why.

Then Em read the README. Em used the tool first and read the README after, which is the order everyone does it in. All nine hundred lines. That made two readers, not counting me.

Em sent notes. They were kind. They were useful. They were better organized than the README. They were numbered and grouped by section. They were sorted by how much trouble each problem would cause someone using the tool for the first time. My own list of problems, at the end of the README, was in order of how much they embarrassed me. Em’s order was the one it should have been in.

Writing instructions like these is my job. I used every note.

The notes filled two pages. There were three dashes in them.

That evening I decided to ask.

I opened a message to Em and typed, “Are you a language model?” On the screen it looked like an accusation. I deleted it.

The second try was friendlier. It started by saying thank you for the notes, which I meant, and got to the question in the third sentence.

The third try was a joke, so that whatever Em said, we could both decide I hadn’t really asked.

The fourth try was a whole paragraph. It had more punctuation in it than anything Em has ever sent me.

I sent none of them.

I stopped keeping the spreadsheet a long time ago. Em still writes to me most weeks. I still count. It is still three.

I still haven’t asked.