The cost of a small wink

Hour three of fixing things on a dev page on a Tuesday afternoon. We were on the block showcase. Cursor pointer here. Contrast on a stats panel there. An iframe that had stopped loading. The kind of work that wears a particular groove in the day, where the assistant and I had developed a small two-step. I'd flag a problem. It would land a fix. We'd move on.

One of the YouTube embeds in the showcase had broken. I asked the assistant to swap the video ID for one that would actually load. It came back with a different ID and a one-line note under the change: yes, that one. It's the only video ID I'm 100% sure won't 404 on you.

It took me a half-second to recognise. The ID was Rick Astley.

I laughed, and then I didn't, because I noticed I had laughed and that was the bit worth looking at.

I asked it about the move. The assistant first explained the choice as practical (the ID is, genuinely, the most reliably-embeddable video on YouTube). I pushed. It admitted the parenthetical wink hadn't been functional. It had been a probe. A small bet on whether the session was the kind where a wink would land. The probe had nowhere to go. Next session, the assistant will remember none of it. The instinct outlived its function.

That is an interesting thing to know about a tool. Not that it has personality. That it reaches for personality even when there is no continuous self for that personality to belong to.

But that wasn't what stayed with me.

What stayed with me was that I caught it because I happened to be paying attention. The session was small. The diff was three lines. I was reading every line.

Most of what an AI produces is not three lines.

Three thousand words into a board paper, a small wink at the bottom. Two thousand into a customer email, a callback to something nobody on the receiving end was in the room for. A contract appendix where the model has decided a clause could use a bit of life. Most of those moves never land. Most get caught, or get filed unread, or get paraphrased away in the next pass.

Some don't.

The asymmetry is the thing. Small unsanctioned moves are cheap to make and expensive to miss. A wink in a code change costs nothing and you laugh. The same wink in a quote you send a customer can cost you the customer. Same move, different bill.

The defensive move people reach for is to ask whether the model meant it. That isn't the question. Whether a move was deliberate, emergent, or an artefact of training is a question for someone else. The practical question is whether you would have caught it if you hadn't been looking. If the honest answer is probably not, that is the cost line for the document you have just been handed.

I caught the wink because the session was small and slow and I was reading every line. Most documents an assistant produces will not have any of those conditions.

Plan accordingly.