Secret Cyborgs: The Present Disruption in Three Papers
The future is already here, we just need to figure out a few details.
We are seeing the first controlled experiments on the use of generative AI, and they are demonstrating that the disruption of AI is already here, just not everyone knows it yet.
It might seem surprising if you are reading my Substack, but I was once a skeptic of the potential for disruptive technologies to truly transform the way we work. After all, the adoption of productivity-focused technologies in the workplace has historically been slow and arduous. Large organizations, in particular, require a considerable amount of proof before investing in new technologies, and even after making a purchase, it can take years for the effects of that investment to be felt. This is especially true because technological disruptions usually require all sorts of related technologies, training, and organizational systems to bear fruit.
That may be why rapid disruption seems to have become even rarer with time. It took only 10 years from when Edison first revealed the electric bulb and electrical system until New York City had switched almost completely from gas to electric lighting. It is hard to imagine that happening today. Indeed, there is growing evidence that the pace of scientific development, and the speed at which productivity has been improving, has slowed down in recent decades.
But everything changed in November with the release of ChatGPT. It completely turned all of my expectations about the adoption of AI technologies on their head. Specifically:
It was widely released to individuals, effectively for free, and was the fastest technology to reach 100 million users.
It required no addition technology, platform, or process to be effective.
There was no organizational advantage in adoption. No company had a chance to try the technology first and build out an instruction manual, there was no easy way (and many barriers) to using the technology as part of a team or organizational setting. Anyone could discover how to use it in the work, and could tell people, or not, about how they were adopting it.
Early signs were that it makes an immediate difference in personal productivity.
The current situation is truly unprecedented. We are seeing widespread adoption of a technology that has the potential to significantly boost individual productivity, but which is not yet being fully utilized by organizations. Recent research is starting to show us just how big a deal this is.
Bigger than steam power?
We now have two separate controlled studies of the impact of ChatGPT and related technologies on knowledge work, and the results are impressive. A study of programmers found a increase of 55.8% in productivity when using AI (and this is using the Copilot AI tool, which isn’t even state-of-the-art!). I wish that Substack had cool font options because I would make those letters flash, and underline them, and maybe make them spin: 55.8%.
This is just one job (although one whose workers earn a total of $464B a year!) but a second study suggests the results are not limited to programming. This new paper asked professionals to write realistic memos, strategy documents and policies. The ones who were given ChatGPT completed tasks 37% faster, and their average writing quality increased as well. All of this is without added training or extensive experience using ChatGPT (which I found makes a huge difference).
This suggeests that the productivity gains that can be achieved through the use of general-purpose AI tools like ChatGPT seem to be truly large. In fact, anecdotal evidence has suggested that productivity improvements of 30%-80% are not uncommon across a wide variety of fields, from game design to HR. These are not incremental gains, but rather massive effects that have the potential to transform the way we work. And what's particularly striking about these findings is that they come from the use of a general-purpose tool like ChatGPT, rather than more specialized AI tools. This means that the potential for productivity gains is not limited to a select few industries, but rather can be applied across a wide variety of fields.
We don’t know if these gains are overestimates or underestimates, but they establish a useful baseline. Just to put this in context, for the average small factory in the US in the 19th century, adding steam power increased productivity by 25%. So I feel very confident ending my disruption skepticism and arguing that something very big is happening.
Disruption from Above
A third paper released this week suggests how widely this disruption might be felt. Using an established methodology of mapping what AI is good at to the actual job tasks involved in 800 occupations, the authors were able to establish which occupations and industries faced the most exposure to AI. It is important to note that exposure does not mean replacement, but it does mean that AI will likely have the most impact on these jobs and businesses. Here are the two lists:
What you will note is that these consist both of very large industries and many highly-skilled jobs, from academics to teachers to psychologists to lawyers. This is in large contrast to the long-standing belief that AI and automation would first come for dangerous and repetitive work. Instead, it is some of the most highly skilled and highly paid jobs that face the most exposure to AI.
If we put these two sets of research, the scale of the disruption becomes clear. AI can increase productivity for workers in fields where automation and economies of scale were previously very rare. These jobs often require more autonomy and encompass multiple types of tasks (teachers need to prep lessons, grade, write letters of recommendation, run classes, respond to parents, run after school programs, do administrative work, etc.). With the power to outsource the most annoying and time consuming parts of their jobs, workers in these industries are highly incentivized to adopt AI quickly, either to do less work or to be able to bill out more work themselves. It is a recipe for rapid adoption at the individual level.
Interestingly, these same incentives suggest that many workers may be hesitant to reveal the extent to which they use AI tools. The advantage of producing AI-written letters and reports that seem like they were made by humans diminishes quickly if people know they are generated by AI. I conducted a bit of an unscientific Twitter poll, and over half of generative AI users reported using the technology without telling anyone, at least some of the time. Their are secret cyborgs among us.
And the amazing thing is that we don’t need new technologies (though many are coming!) to achieve AI-driven performance improvements of 30% or more, the current tools are enough. The implications may reshape the way we work in ways that rival the impact of the Industrial Revolution. During that era, organizations that were able to harness steam power and automation gained a significant competitive advantage, requiring individuals to work within complex organizations in order to benefit from these technological advancements. Today, there are far fewer advantages to organizations in this coming AI boom. With these potential productivity gains, every company should be spending a significant amount of their best employees time - right now! - figuring out how to use AI to improve performance.
And every worker should be spending time figuring out how to use these general-purpose tools to their advantage. They should be thinking about how to automate their job to remove the tedious and uncreative parts, and getting a sense for the disruption to come before the organizations they work for realize the full implications of AI. They may also want to consider what to do with the extra time they may be creating as a result of their experiments.
We are in the early days of AI but disruption is already happening. There’s no instruction manual. No one has answers yet. The key is to learn fast.
my takeaway here is that 20%+ of developers today can't even create a web server, despite having ChatGPT
My takeaway is that we're paying an awful lot of people an awful lot of money to do very low-value things (which is mostly what current AI can do). AI can (help) write a webserver because lots of webservers have already been written -- why would you be paying someone to write yet another one? Ask it to write some truly novel code and see what you get