At the beginning of 2023 the world witnessed a historic turning point for Artificial Intelligence (AI). Several generative AI platforms went viral in a matter of weeks. ChatGPT, a large language model (LLM)-based AI designed to emulate human conversation reached 100 million monthly active users in less than three months. According to a report from Swiss bank UBS, no social media platform, internet services or consumer apps had ever achieved such rapid growth before.
The imagination of half the world was captured by the possibilities of the new machine that were impossible to imagine a few months ago, whether due to its infinite applications for the generation of ideas or businesses, its instant capacity for the automatic creation of analytical or creative content or its skills to pass complex professional exams with grades in just a few seconds that require years of preparation for humans. The shock was even more disconcerting when, when inaugurating the new version of the algorithm, GPT-4, Sam Altman, the CEO of Open AI – the company that developed ChatGPT – declared that “we are just at the dawn of these revolutionary technologies.”
While some have seen on ChatGPT “sparks of General Artificial Intelligence” – a form of higher-than-human technological intelligence imagined by California technologists – others have announced the definitive loss of control over the machine and the imminent arrival “of a Jurassic Park moment for Artificial Intelligence”.
But, a year after the phenomenal emergence of this AI, what effects is ChatGPT really having? In this text and the next I will analyze two groups of important effects. The first is the impact on productivity and employment, which I discuss here. The second is the effect on democracy, which I will discuss in the next installment.
The first effect that seems clear in the experimental literature so far is the very substantial improvement in productivity across a wide variety of tasks. In one of the first studies released, two MIT students found that ChatGPT increases productivity in a series of professional tasks associated with writing typical of the daily work of any office worker, such as writing small reports, delicate emails or press releases. To measure productivity, the authors calculate the time spent performing tasks and measure their quality relative to a control group. The most impressive thing about the results is the size of the improvements. For example, in the treatment group the average time doing homework was reduced by almost 40%.
A second effect on which there is beginning to be quite a consensus is that the productivity improvements achieved by this AI are concentrated in the lower part of the distribution. That is, AI reduces inequalities in a multitude of tasks. For example, in a recent post Authored by researchers from several American universities, the impact of ChatGPT is analyzed on a group of 750 strategic consultants from the Boston Consulting Group in 18 different tasks including economic analysis, idea generation or persuasion. The improvements in the results of the ChatGPT-supported consultants versus the control group are also enormous. However, the impact is much more positive for workers who had worse results in the previous test (without access to AI). These workers show improvements of 43%, while those in the upper half only improved 17%. Results along the same lines are found in other experimental studies with customer services, lawyers either creative writing exercises.
A third relevant conclusion regarding the impact on productivity of this AI is the transversal nature of the tasks and sectors it affects. While the impact associated with the development of robots or previous AI families seemed relatively limited to certain industries and a relatively limited profile of automatable tasks – think of robotic arms on automobile assembly lines, for example -, these AIs are going to clearly impact creative and intellectual jobs of all kinds. Edward Felten, professor at Princeton University, and co-authors develop an index of tasks and sectors potentially affected by generative AI. They conclude that the most affected sectors tend to be those with higher salaries and in particular the professional services, legal or financial advice or teaching industries.
What can we conclude from all this? Automation processes can be of two types: rather complementary to the tasks that humans do, favoring workers (and companies and consumers), improving their productivity and freeing up time and talent to dedicate them to more stimulating activities, or rather substitutes for those same tasks, eliminating jobs that robots take on and generating unemployment. From a social point of view, many of the improvements provided by technologies of the second type are very limited, since they provide very small improvements in efficiency and erode resources in the hands of workers to increase the profits of the owners of capital, as argued Acemoglu and Robinson in their latest book, Power and Progress. An example that the authors use is that of ATMs in supermarkets: they improve business margins slightly, but they do not improve the customer experience and they do eliminate many jobs.
It is still too early to assess what the effect of ChatGPT will be on employment. Not only because little time has passed, but also because of the characteristics of these AIs. These AIs are “black boxes” (opaque models of deep knowledge that use massive data to predict the most likely association of words, code or images) and their usefulness varies greatly in tasks within the same types of work and in cognitively demanding tasks. very different. For example, ChatGPT may be very good at generating creative business ideas, writing articles on populism, or solving complex legal problems, but it may not be good at doing basic elementary math problems.
But there are reasons to be optimistic. Anyone who uses ChatGPT regularly (I do, also to write this text) will have realized that the fun is in the interaction with the machine. In many of the cases mentioned, ChatGPT works as an assistant or a support tool that to be useful must be constantly guided, tuned and corrected by a human. The MIT students elaborate on this idea in their research: they show that ChatGPT helps replace worker effort and reorient tasks towards idea generation or editing rather than the more cumbersome tasks of initial writing. They also find that exposure to chat improves worker satisfaction… but also their concern about the future of AI.
Follow all the information Five days in Facebook, x and Linkedinor in our newsletter Five Day Agenda
_
#It39s #year #ChatGPT39s #impact #employment