The Limits of Data
I once sat in a room with a bunch of machine learning folks who were developing creative artificial intelligence to make “good art.” I asked one researcher about the training data. How did they choose to operationalize “good art”? Their reply: they used Netflix data about engagement hours.
The problem is that engagement hours are not the same as good art. There are so many ways that art can be important for us. It can move us, it can teach us, it can shake us to the core. But those qualities aren’t necessarily measured by engagement hours. If we’re optimizing our creative tools for engagement hours, we might be optimizing more for addictiveness than anything else. I said all this. They responded: show me a large dataset with a better operationalization of “good art,” we’ll use it. And this is the core problem, because it’s very unlikely that there will ever be any such dataset.
Right now, the language of policymaking is data. (I’m talking about “data” here as a concept, not as particular measurements.) Government agencies, corporations, and other policymakers all want to make decisions based on clear data about positive outcomes. They want to succeed on the metrics—to succeed in clear, objective, and publicly comprehensible terms. But metrics and data are incomplete by their basic nature. Every data collection method is constrained and every dataset is filtered.
