Friction, or, Why AI Will Not Replace White-Collar Jobs
Thinking about AI adoption through the lens of Clausewitz’s concept of friction.
“Everything is very simple in War, but the simplest thing is difficult,” wrote Carl von Clausewitz in his seminal treatise On War.1 He used the concept of friction as a metaphor to describe how a series of small problems can add up to significantly delay and interfere with the execution of a general’s plans.
He illustrates this with an example of a traveller riding on post-horses.2 Judging by the distance he has to travel, he expects to make it to his destination by the end of the day, but, upon arriving at the next inn, he finds that they have no horses for hire, or ones that are already tired. He then runs into some bad roads, or difficult terrain, or bad weather, and so on, all of which add up to delay his journey. “So in War, through the influence of an infinity of petty circumstances, which cannot properly be described on paper, things disappoint us, and we fall short of the mark.”
Friction creates the gap between expectation and reality. A general can never predict the outcome of his actions with certainty because everything in war is “brought into contact with chance.” A fog might hide the enemy, heavy rainfall might prevent reinforcements from arriving on time, cavalry might get mired in mud, and so on. No matter how much time a general spends planning, he can never account for the delays and setbacks caused by friction.
The excitement generated by new AI models is leading pundits to speculate about these models taking over white-collar jobs. The idea is that, just as mechanical automation has replaced certain manual-labor jobs, AI will now replace many administrative and even creative jobs. The flaw in this line of thinking, however, is that it assumes a world without friction.
The new AI models are good at performing tasks, but not doing jobs.3 The distinction is crucial. A task is a defined unit of work, while a job is a system of processes. For example, right now AI can summarize a lengthy text or generate an image based on a prompt, but it does not do all of the other work that makes its outputs useful—i.e. where is this summary or image going to be used, what will it be used for and how? Using the output of a task to achieve an actual goal requires human input. Can AI be trained to do this? Maybe. But this other work involves two types of friction that will make it very difficult in the context of business.
The first is that your typical corporate infrastructure is not a neat, perfectly structured machine, but rather a mass of cumbersome enterprise systems and spreadsheets, cobbled together over many years by different people at different stages of an organization’s life. Human employees are the glue keeping it all together. The reason why these systems are old and cumbersome is that it is often cheaper to hire people to deal with their quirks than to pay significantly more to update them. And because people are used in place of software as the interface between a medley of spreadsheets and enterprise systems, getting the AI to access the data trapped within will prove difficult.
The second type of friction is simply the fact that a lot of administrative work involves dealing with exceptions. What makes human employees especially valuable is their initiative and judgement, which allows them work around a system when it breaks down. For AI to be able to replace human workers, it would have to be able to not only recognize exceptions and mistakes, but also to fix them. It would have to work so well that it would no longer need constant supervision.
Certain fields with clearly defined deliverables—i.e. tasks rather than jobs—such as illustration, will be and are already being disrupted. But more generally, rather than replacing jobs, AI will most likely manifest in a series of tools to help solve specific problems. And these tools will still need people to operate them.
Carl von Clausewitz, On War (Book 1, ch. 7)
Post-horses were horses kept at inns along post routes for use by post riders, or hired out to travelers. This allowed for tired horses to be changed for fresh ones at the next inn, increasing the distance one could travel in a day.
This distinction has similar implications for manual labor. A factory worker on a conveyor belt is typically performing a task, which can be mechanized, while someone like a plumber is doing a job, which cannot be (at least not in the near future).