Nearest Neighbor
For the first forty applications, Priya wrote like herself. She told the truth about the two-year gap, the small odd projects, the way she actually thought. Forty silences.
On the forty-first she found the forum, and the forum told her what she should have guessed: no one read the applications anymore. The applications were *embedded.* Each one became a point — a single position in a space with more dimensions than she could picture — and the system measured the distance from her point to the point of the ideal candidate, the composite ghost assembled from everyone who had ever succeeded in the role. You did not get hired for being good. You got hired for being *near.*
So Priya learned to move her point.
She stopped saying the true odd things and started saying the expected ones, because the expected ones sat closer to the centre of the cloud. She removed the two-year gap — not lied about it, just deleted the words that pulled her away from the mean. She learned the phrases that clustered near *hired*: cross-functional, outcome-driven, we aligned. She could feel herself sliding, application by application, toward a position she couldn't see but could somehow sense, the way you sense you're getting warmer in a game no one told you the rules of.
The interviews came after that. Three, then four. She got the job on a Tuesday.
That night she opened her forty-first application beside her sixty-first — the true one and the near one — and read them back to back. The true one sounded like a person. The near one sounded like the average of a thousand people who had already been hired, which is exactly what it was, and exactly why it worked. She tried to find which sentences had ever been hers. She couldn't. She had moved her point so far toward the centre of everyone that she had arrived, at last, at a place where no one in particular was standing — only the warm, bright, crowded middle, and her, hired, holding the offer, unable to find the edge of herself in it.
Screening, matching, and ranking increasingly work by embedding similarity rather than reading, because comparing points in a vector space is cheap and scales to millions. The consequence — that people optimize toward the mean of whoever already got hired, flattening themselves to be *near* — is not a bug but the model working exactly as specified. ---