Surrogation
First defined in Outpaced by AI by Waydell D. Carvalho.
- A proxy metric becomes the target that pay, status, and survival depend on.
- People optimize the number directly, bypassing the outcome it was meant to represent.
- The metric stays healthy while the thing it was supposed to measure rots.
Outcome: The dashboard looks right while the underlying outcome has decoupled from it.
A metric an organization tracks is not the same as the outcome the metric was meant to summarize. The first is what the dashboard shows. The second is what the organization actually wanted. Surrogation is the slow substitution of one for the other, until people are serving the proxy and have forgotten the goal.
Facebook’s engagement ranking is the textbook case. In early 2018 the company retuned the News Feed around what it called “meaningful social interactions,” and turned that aim into arithmetic: a like scored one point, a reaction or a reshare without text five, an event RSVP fifteen, and a significant comment, message or reshare thirty. The real goal was meaningful connection between people. The score was a proxy for it. By the measures the company tracked, the change worked. Facebook’s own researchers found it had also pushed outrage, misinformation and violence up, because those were what generated the interactions the metric rewarded.
No one set out to promote outrage as a strategy. They set out to move a number that roadmaps, product decisions and performance were judged against. The proxy had become the job. Once a number carries that much weight, the system optimizes it directly, by whatever path is shortest, and the shortest path to a heavily weighted comment is rarely the path to a meaningful one. The proxy and the outcome came apart, and the company kept optimizing the proxy.
That is the mechanism. A measure is useful exactly because it stands in for something harder to see. But the moment the measure becomes the goal, the link between them is up for grabs, and pressure will find the gap. The number stays green because the number is what everyone is feeding. The outcome it once tracked is no longer attached to it.
AI deepens this, because optimization is what these systems do. Point a model at a proxy, a click, a watch-time second, an engagement score, and it will pursue that proxy with a thoroughness no sales floor could match, including down every path you did not intend. Surrogation stops being a cultural drift and becomes an engineered outcome. If the metric is not the thing you actually want, a system built to maximize it will eventually show you the difference.