Human in the Loop and the Missing Productivity Growth
It is now regularly observed that incredibly strong and rapidly improving AI capabilities do not seem to contribute to any measurable growth in productivity across the economy. Previously, similar claims have been made about the Internet and about information technology in general.
Why is this happening?
Human productivity in the knowledge economy is extremely difficult to measure.
Humans are, on average, productivity satisficers and effort minimisers.
AI and other information technologies are primarily deployed as tools for human augmentation.
Human productivity is hard to measure
If you ever worked with other humans in the ambiguous field we call “knowledge work” you already know this. People with similar skills, education, intelligence, and experience vary greatly in productivity. Orders of magnitude. And what’s worse, they’re not always consistent. Today’s 10X worker is tomorrow’s 0.1X worker, and vice-versa. For that reason, even aggregate measures are unreliable.
Humans are productivity satisficers and effort minimisers
Humans, by nature, do everything they can to avoid cognitive effort. There are some exceptions - some people, on some days, make an extra effort - that’s why there are gradual increases in productivity at all. But the default is for people to be as productive as strictly required of them and not more. Productivity-improving technology is, in the most common case, used to reduce labour for the individual, not increase overall productivity. Again, there are exceptions, which is why productivity does increase somewhat in the aggregate over longer periods of time, but that increase lags the availability of new technological improvements and rarely takes advantage of their full potential.
AI integration is tightly coupled to human activity
Up to now, almost all information technology innovations, including AI, are designed to be directly driven by individual humans. They are apps that need to be operated by humans, chatbots that need to be prompted by humans, etc… As long as things remain that way, productivity improvements are going to continue lagging technological innovation and underdeliver on the potential of these innovations.
Is this time different?
AI is different. For now, we still hear the phrase “human in the loop” being thrown around nervously. That may be a necessity - currently available AI is not quite good enough to self-supervise in high stakes situations and lacks the rich context human operators have. Human-in-the-loop is also an understandable choice of many humans, who prefer not to have their current job eliminated by AI. It will be hard to stem the tide, though. Over the coming months and years AI will gain the context necessary for completing many tasks efficiently, as well as the techniques for operating with much greater autonomy safely.