Here’s an idea: let’s say AI value can be conceptually divided into two buckets: automation and new value creation. The vast majority of work in AI today is probably on automation, even though most prefer to believe they work on new value creation. Let’s let “automation” refer to work currently done by people that can now be done by machines more cheaply. Value from automation thus derives from decreased production costs, rather than quality increase. By contrast, let’s say “new value creation” refers to everything else - so new economic activities or any quality improvements over current processes.
For example, if I launch an AI assistant for doctors that improves their medical diagnoses so that doctors make fewer mistakes - that would create new value. Whereas if I create a chatbot that you can talk to instead of going to mundane doctor’s appointments - that would be automation.
Don’t get me wrong, automation is hugely impactful. The effect of automating an economic activity is analogous to increasing the supply of individuals performing that economic activity. Automation may decrease costs and thus increase the accessibility and ubiquity of expensive services. For example, automating mundane doctor interactions could allow for medical services to reach far more people and to reach people far more often. People may also get more time and attention from each doctor interaction than they might with the current shortage of medical personnel. And there may also be a massive backlog of medical activities that aren’t performed today or are understaffed that may finally see the light of day.
Why does this matter? First and foremost because automation will obviously eliminate many jobs. The important question is how we can estimate the rate and number of jobs that will be created in the process.
I suspect that job creation will likely come from three places:
AI new value creation will lead to new jobs; for example, AlphaFold may unlock many scientific jobs by solving the protein folding problem.
Automation itself will create some jobs, for example, for the “automators.” But (almost definitionally) fewer jobs than it’s replacing. For instance, robots automating factory jobs may need people to manage the robots - but undoubtedly far fewer people than the number of displaced factory workers. Data labelers probably fit in this camp too.
Gaps in the job market, where certain professions and roles remain unfilled or don’t even exist yet due to our limited workforce capacity.
Our current US unemployment rates are incredibly low, revealing a large number of unfilled jobs. Beyond our truck driver shortage, there is likely greater job demand than we can see in certain areas. For example, it’s well established that having fewer teachers per student in a classroom is good for a student’s education; perhaps with an increasing supply of workers in the economy, we could see continued demand for more teachers until we reach ratios as low as four students for one teacher. And similarly with nurses and other caring professions.
I suspect that the vast majority of folks in AI today think of themselves as working in new value creation rather than in automation. It’s easy to think that GitHub Copilot doubles a single engineer’s code output and quality, rather than halving the number of engineers a company needs to achieve the same results. You just can’t sell Copilot to engineers with the promise of automating away jobs, even if that is its net effect.
My claim: job creation will likely come from the three above sources: new value creation, automation, and current gaps in the job market. If automation far outpaces new value creation, the only thing preventing high unemployment is the unseen gaps in the job market. If you believe, as I do, that current gaps are massive, that might buy us some time. But filling in gaps is not a permanent solution, and we’ll likely need new value creation for any meaningful number of new jobs to be created.
Finally, there is some reason to doubt that new value creation can be infinite - that lots of people are pretty happy with the amount of stuff they have today [1]. Economic growth is no longer driven only by new innovations like the airplane and the iPhone, but by marketing campaigns that encourage us to believe that we need to buy an air fryer or an Instant Pot. Many of us are actively turning away from consumerism and discovering happiness from other sources, preferring minimalism over maximalism. It’s possible that new value creation from AI is just less exciting to us than the prospect of working less, i.e. automation.
Why I might be wrong:
New innovations might just be really hard to imagine. Let’s all remember that in Back to the Future 2, no one imagined the internet as something that could change the world. Perhaps new innovations will create value in ways we just can’t imagine.
Relatedly, value creation has often come from automation unlocking higher levels of abstraction. For example, compilers automated the work of low-level engineers adapting software for different machines, opening a new category of engineers operating at the software level.
I’m curious to hear how significant the gaps in the job market really are. Are they 3x the total number of jobs that exist today or 30x? If you’re an investor, are you more excited about investing in automation or new value creation? For anyone, are you more excited about automation from AI (i.e., working less) or new value creation?
Also, let me know what gaps you see in my logic.
[1] Obviously, poverty exists, but I’m hopeful that it is a solvable problem
Interesting destinction between "new value creation" and "automation". I wrote a related post differentiating between augmentation and automation: https://scalingknowledge.substack.com/p/why-job-displacement-predictions