How Do You Learn if the AI Does the Grunt Work?

The Verge has an interesting article about a history professor using a chat bot he helped build to processes source documents (parse their handwriting, etc.) and get source material appropriately tagged and sorted, a tedious, time consuming (as in years) task.

The professor had to build his own system because the OpenAI and other systems could not do the work he needed. They simply could not recognize the way people in the 18th century wrote without significant help. And, of course, they continued to have the hallucination problem that plagues all imitative AI. The professor provided that help, and he got a system that combined several different tools to help him get reasonable answers to his queries about the source documents.

The system was not perfect, but the professor claimed it was faster and more accurate than the graduate students he used to get to do this work. And, of course, cheaper (though, honestly prof? $16,000 a year to help with your research?) In theory this is an example where AI does mostly good — it allows for the correct tagging and understanding of source document in a much faster timeframe, and thus allows for more material — especially material that would not have been considered important for reasons of race, sex, or class bias — to be properly catalogued and exposed. Even the grad students who used to get paid for doing this weren’t making a living wage, based on the article, and could in theory be redeployed to more important work.

And yet something the professor said brought me up a little bit short. In the article he states that “In order to make a lot of this data useful, people are needed who have both the ability to figure out how to train models, but more importantly, who understand what is good content and what’s not. I think that’s reassuring.” And that can, on its face, be reassuring. What the professor built could be seen as an augmentation, not a replacement. Except how are people going to learn what good content is and what is not if they are never exposed to the kind of classification/research work that allows them to learn the distinction?

One of the dangers of outsourcing, and one of the reasons it is not taken over 100% of the industries that it is deployed in, is the knowledge drain. For a variety of reasons, if you want your systems to do what you want them to, then you need people in house who understand the systems. That is well and good at the start —you can outsource the junior people and keep the senior people in house. But how do you get new senior people if you never train up junior people? No one walks out of college an expert — you need to learn on the job in almost every job I have ever encountered to really know what you are doing.

In the case of history professors, if grad students don’t do the kinds of classification work that helps them learn how to process and understand source documents, then who is going to be positioned to correct the imitative AI systems when they inevitably screw up? You have to train the next generation to be able to understand the outputs, and that sometimes means being less then efficient in order to make sure the next generation gets the chance to actually learn what they need to know. That is a point that our focus on efficiency seems to too often miss. AI is going to make that tendency worse.

AI is wired for the people who think about the now, the next quarter. Unfortunately, that generally means the people in charge. The ones who know how to do things, who understand the nature of learning and of time, are not often consulted in these big shifts. I think it is becoming clearer that imitative AI could drive a reduction in expertise. That is a cost that I don’t think people are realizing we are going to pay and aren’t thinking about mitigating at all. As long as the line goes up the next quarter, who has to care about the next five years?

We all should, but I think that our economy and society are too short-sighted to understand this particular damage.