We are finding ourselves in a strange place this Labor Day: the impact of AI is simultaneously being overestimated and underestimated. There is still massive hype surrounding self driving cars, which will almost certainly take much longer to arrive than the public appears to believe. At the same time, there is mostly quiet but massive AI progress on tasks such as reading radiology images.

The key thing to understand – if you want to form an opinion on AI progress – is to distinguish between “open” and “closed” domains. The difference revolves around how much knowledge is required in adjacent domains to be effective in the domain itself. In an open domain you need to know a lot about other domains as well. 

Driving is an example of an open domain: to be an effective driver, you don’t just need to know about the road and the car and the physics of acceleration and breaking, but you also need to know about people and their intentions (eg is this person about to cross the road or not?), stuff that can be found lying on a road (eg a plastic bag is nearly weightless and you can drive over it but a piece of metal might puncture your tire). 

Reading a radiology image, by contrast, is an example of a closed domain. The image will have a part of the human body on it and you are supposed to find something that’s wrong, e.g. a broken bone or a tumor. In order to find that you do not need to know anything beyond that; for instance, you do not need to know how the bone was broken or how the tumor might be treated.

AI is making rapid progress on closed domains, such as playing a game or reading a radiology image. As it turns, out much of what humans do in the workplace today operates on closed domains. That’s in fact so by design and has been the hallmark of Industrial Age organization (by this I don’t just mean manufacturing but also services, such as banking, insurance, etc.) We have over a long time carved up jobs through specialization so as to reduce the knowledge required for carrying out the task.

The hype around self-driving cars will likely disappear and that could happen any day. Don’t let that fool you about the prospects of AI to radically change how much and what kind of labor will be required going forward.