We Are the Training Data
A generation raised inside the feedback loop is developing instincts no previous culture has needed: how to be legible to machines, and when to refuse to be.
Every prior generation performed for an audience it could see: the village, the boss, the camera. This one performs for a reader it cannot — the model that will summarize its emails, rank its résumé, score its code and, in aggregate, learn from everything it leaves behind. The audience is no longer watching. It is training.
You can already watch the instincts form. People write for retrieval — front-loading conclusions, structuring documents the way parsers like them. They keep a private register of what never goes in writing. From this week, EU law obliges companies to tell them when the counterparty is synthetic [1]; the etiquette of talking to machines is becoming a legal category before it becomes a social one.
Legibility as labor
Being machine-legible is starting to function like a job skill — the way being camera-ready became one in the last century. The professional who structures their work so agents can act on it compounds faster than the one who does not. The cost is a subtle self-editing: when everything you produce is potential training data, spontaneity becomes a decision rather than a default.
The counter-instinct is forming too. Off-record spaces — group chats, unrecorded calls, paper — are acquiring the cachet that unplugged vacations had a decade ago. The next status divide may not be who has access to the best models, but who can still afford to be illegible to them.