Hitting the target and missing the point
The moment a number becomes a target, people start working on the number rather than the thing it was meant to measure. It's one of the oldest ideas in management and one of the easiest to forget when you're the one building the dashboard.
The idea has a few names
Economist Charles Goodhart made the point about monetary policy in the 1970s. The version most people quote is anthropologist Marilyn Strathern's paraphrase: "When a measure becomes a target, it ceases to be a good measure." It's known as Goodhart's law. Around the same time, social scientist Donald Campbell argued that the more a quantitative indicator is used for social decision making, the more it gets corrupted and the more it distorts the process it was meant to watch.
They were describing the same thing from different directions.
Health services are a well studied case
Some of the clearest evidence comes from the English NHS in the early 2000s, when the government set hard targets for things like waiting times and ambulance response times, with real consequences for missing them. Gwyn Bevan and Christopher Hood studied what happened in a paper with a very good title: "What's measured is what matters: targets and gaming in the English public health care system" (Public Administration, 2006).
Their reading was balanced, which is part of why I like it. Reported performance did improve, and some of that improvement was real. But they also found signs of gaming: data that was hard to believe, effort moved toward whatever was being counted and away from what wasn't. LSE's later summary of the research put it well, linking strong incentives to targets brings both improvement and attempts at "hitting the target and missing the point".
Their conclusion wasn't "don't use targets". It was that targets work best when they measure something that matters directly and can't easily be manipulated, and that gaming gets worse the more a target is only a proxy.
Why this matters for data teams
Most of us aren't setting national targets. But anyone who builds dashboards is choosing which numbers people look at every morning, and that's a quieter version of the same thing.
A few things I try to keep in mind:
- Ask what the number is a proxy for. If the answer is "something we can't measure directly", expect it to drift away from that thing once people are judged on it.
- Pair metrics. A speed metric with a quality metric. A volume metric with an outcome. It's harder to game two numbers that pull in different directions.
- Watch for numbers that improve too neatly. A sudden cluster of results just under a threshold is worth a look.
- Keep some measures for learning only. Not everything needs to be a target. Some numbers are more useful when nobody's bonus depends on them.
- Revisit metrics on purpose. A good measure two years ago might be the most gamed one today.
AI makes this sharper
This applies to models too. Anyone who has trained a model has seen it find a shortcut that improves the score without doing the task. And as more work gets evaluated by automated checks, the same law applies to people and to the agents they build. If the check is the target, the check is what gets optimized.
The takeaway
Choosing a metric is a decision about what people will spend their time on. Make it as carefully as any other decision, and check every so often whether the number and the thing you care about are still moving together.