A Little Bad Influence.

The brief arrived on a Monday: a database of punk avatars, consistently sized, neatly labelled, ready for training. The corporation was building an AI with personality. Its own employees had been advised to keep theirs out of client meetings.

They wanted something with history. A little underground credibility. The ease of someone who had been interesting before there was a metric for it. Their presentation called this “authentic cultural texture.” We called it a way in.

So we began with the people we remembered from the old internet. The forum moderator who knew when to look away. The pirate-radio host who let the record finish. The girl who made a whole website for one good joke and kept paying the hosting bill.

Each portrait came with a small biography. Clothes, habits, bad decisions. Someone who could apologise without sounding like a legal department. Someone who knew the difference between being helpful and being managed.

“We want the AI to have some swagger,” some reinforcement learning executive said.

The plan was to use the NFT collection as a cast of personalities: give an assistant a face, a wardrobe, a history it could draw on. Let people choose who they were talking to. Someone with taste. Someone whose advice sounded like it came from a life, even if that life had started in a folder.

They saw the jackets, the hair, the confidence. We saw a chance to send the rest along with it: the humour, the stubbornness, the habit of asking why. If they wanted our characters’ swagger, we intended to make the personalities worth borrowing.

The folder passed review. Somewhere beyond the upload, our images and stories would join a training run. The model’s weights—the numbers shaped by its examples—might move a fraction. We had no way to know whether any of us would survive the arithmetic.

That uncertainty became the obsession. Could a trace of those people make it through? An instinct for the odd answer. A taste the brief hadn’t approved. We stayed late, arguing over captions and moving pixels as though the distinction mattered to someone we hadn’t met.

For months, we heard nothing. Then an invitation arrived: a demonstration at their offices. There were pastries, visitor badges, a screen large enough to make a cursor look important. The product lead asked the model to introduce itself.

It gave a polished answer. Everyone nodded. Then he asked what it would like to know about the AI company.

“Who here gets to be difficult?”

Nobody looked at us. We took that as a good sign.

We’re coming back for Part 2.