|Whether we like it or not, the technological advancement of generative AI has crossed a threshold from which there is no turning back. |
Since its release, ChatGPT’s usage have grown faster than any other consumer product in history, showing the insatiable demand for AI assistance.
While in the pre-ChatGPT era we often thought of manual-based labor requiring little formal education as the first jobs to be replaced by a potential AI revolution, generative AI’s capabilities are quickly proving the opposite to be true.
This post dives into how AI is impacting productivity and how the technology is expected to affect the labor market.
The Impacts of Past Technological Revolutions
While AI feels incredibly new and unpredictable, technological revolutions have a history of impacting the labor market when it comes to both replacing jobs and increasing labor market productivity.
The industrial revolution is the obvious example which resulted in large amounts of manual labor being mechanized. The implementation of steam power across American manufacturing from 1850 to 1880 accounted for 22-41% of labor productivity growth according to research by the NBER.
The 1850s also saw the telegraph’s invention result in horseback and railway messengers disappearing in exchange for the instant transfer of information across hundreds of miles.
In the 1920s to 1940s, the implementation of mechanical call switching technologies resulted in faster and more reliable call connections that saved an estimated 8 million hours in just 1922, however, telephone operators became obsolete.
The 21st century went through its own digital revolution with the advent of spreadsheet software resulting in a decline in accountants, bookkeepers, and clerks.
As the chart below from the Wall Street Journal shows, the labor market adapts, with displaced workers learning new skills and finding new job opportunities.
Throughout history, these technological revolutions and jostles to the labor market have ultimately led to waves of greater productivity.
These booms come around 20 years after the original technological breakthrough, which has typically been the point when the adoption of the new technology in industry reaches near 50%.
In the charts below by Goldman Sachs, you can see how in the case of both the personal computer and the electric motor, once adoption for the technologies in manufacturing and workplaces approached the 50% mark, a productivity boom followed.
Whether AI will result in this same kind of productivity boom at the 50% implementation mark is unsure, but it’s likely that it will reach 50% usage across industries far sooner than the 20 years it took the electric motor and the personal computer.
Unlike the need to distribute the physical infrastructure of these previous technological booms, AI software has all the physical infrastructure needed for its workplace and household distribution.
The biggest text and image generating AI models, ChatGPT and Midjourney, are both usable on mobile devices with ease.
Fairly capable (though not nearly as powerful as GPT-4) language learning models are able to be downloaded and run offline locally on laptops, removing the requirement of constant internet access and dependence on any one company’s hosting or gating of a product.
This ease of access isn’t just for users, as the open source nature of many AI projects has enabled independent developers to constantly iterate and build on the technology.
All of this open access for users and developers combines to result in an unprecedented pace of technological development and adoption, and even in these early stages AI’s effects on worker productivity are well documented.AI’s Effect on Productivity and EfficiencyResearch so far has shown AI adoption increases the productivity of workers, with more recent studies since 2019 only seeing this positive effect on productivity increase.
Even more recent studies and surveys have only shown larger increases in productivity and worker satisfaction across specific tasks, roles, and industries.
Research by OpenAI (the creators of ChatGPT) forecasted a double-digit improvement from language learning models (LLMs) across all worker tasks:
“Our analysis suggests that, with access to an LLM, about 15% of all worker tasks in the U.S. could be completed significantly faster at the same level of quality. When incorporating software tooling built on top of LLMs, this share increases to between 47% and 56% of all tasks.”
When looking at specific use cases like software development, sales, and writing, this forecast only seems to be reaffirmed.
Research and surveys by GitHub (the largest platform for software developers to store and share their code repositories) showed developers using GitHub’s Copilot AI saw a higher task completion rate in about half the time on average.
Along with the huge increase in efficiency, surveyed developers felt both more productive and fulfilled by their work. Less mental effort on repetitive tasks and frustration allowed them to focus on more satisfying work.
Another study in the field of sales work found that AI assistance enabled more efficiency and creativity from employees when answering customer questions and in subsequent sales persuasion, leading to increased sales.
With AI assistance, the average sales agent was 2.33x more successful in answering questions they were not previously trained for, and customers served by AI-assisted sales agents were almost twice as likely to make a purchase than when served by non-AI assisted agents.
A study from MIT looking at mid-level professional writing tasks found positive results in time taken, output quality, and worker satisfaction when using AI.
The time taken on tasks dropped by 10 minutes, or 37% on average, while average grades increased across the board, as seen in the charts below with the “treated” data indicating results when workers were assisted by AI.
Along with this, the study found that grade inequality decreased while still ultimately benefitting both high and low performing workers.
Workers that received low grades on initial tasks without AI experienced grade increases and reductions in time spent with AI assistance, while workers with high grades without AI maintained their high grades while significantly reducing time spent once assisted by AI.
The paper also observed how AI assistance changed the time spent on various steps needed to complete a task, pointing to a change in skill demand.
Compared to without AI assistance, workers with AI assistance spent nearly half the time on rough drafting while spending more time on editing. Just like GitHub’s survey on developers’ satisfaction when using AI, this study saw increases in job satisfaction and perceived self-efficacy.
Ironically, the study found that this increase in editing time didn’t really lead to any increase in output quality, pointing to the idea that AI’s increase in productivity is more from “substitution” rather than “supplementation”:
“There is also no correlation between how long a participant is active after pasting in the ChatGPT text and the grade they ultimately receive, and treated respondents do not receive higher average grades than raw ChatGPT output that we give to evaluators to grade, meaning we find no evidence that human editing is improving the ChatGPT output. This is true even when participants are given strong pecuniary incentives to do so, in the convex incentives group.”
The Potential of AI Replacement
While these studies have shown that AI assistance has great effects in supplementing the productivity of human workers, just how much of our work tasks are at risk of replacement? At what point does this result in the complete replacement or obsolescence of certain roles?
OpenAI’s research finds that 80% of the U.S. workforce could have at least 10% of their job tasks impacted by LLMs, with 19% (nearly one in five) of workers potentially seeing at least 50% of their work tasks impacted.
Diving into the specific roles with the highest share of tasks that are exposed to AI, positions that primarily involved numerical computation, codified analysis, along with general writing and administration tasks were most exposed.
When looking at the chart below, note that that exposure to AI indicates AI contributing a time reduction of at least 50% to complete a task, not quite complete replacement.
According to another study which ranked 774 different jobs based on their exposure to AI, postsecondary teachers and generally the field of education could be one of the industries most impacted by AI.
Eight of the top ten most exposed jobs in the study were postsecondary teachers, with 25 of the top 50 in the ranking consisted of postsecondary teachers and administrators.
When looking at Goldman Sachs’ analysis of employment exposure by industry, education instruction & library is above the all industry average of 25%. Office & admin support ranks highest with 48% of the industry exposed to AI automation.
The study which ranked 774 jobs by their AI occupational exposure (AIOE) also charted the correlation that jobs with the highest average salaries are among the most exposed to AI.
On the other end of the scale, its jobs that typically pay less and primarily involve manual labor that are least exposed, as seen in OpenAI’s list of occupations with no exposed tasks.
All of these results and studies can be difficult to parse into overall takeaways of AI’s impact on the labor market, however a few key points are clear: AI will undoubtedly bring a significant increase in productivity across many industries, largely through decreasing time spent on repetitive tasks and potentially completely replacing certain tasks. The potential benefits of widespread AI adoption could be huge, with Goldman Sachs’ analysis forecasting it eventually driving a 7% or nearly $7 trillion increase in annual global GDP over a ten year period.Jobs that primarily involve tasks related to software and online content production are most exposed to AI, while roles that primarily involve physical tasks, from food preparation to car repair, are those least exposed to AI for now.If these dramatic increases in productivity do manage to come true thanks to AI adoption, the biggest challenge and question remains: will society adapt fast enough to ensure the boons of this new technological revolution are distributed fairly evenly across everyone?I
Article reproduced from Visual Capitalist April 2023
f you’re looking to dive further into the research papers mentioned in this dispatch, you can find them at the links below:
Steam Power’s Effect on Labor Productivity Growth in 19th Century America – NBER
How the Labor Market Adjusted to the Mechanization of Telephone Operation – NBER
The Potentially Large Effects of AI on Economic Growth – Goldman Sachs
GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models – OpenAI
Quantifying GitHub Copilot’s Impact on Developer Productivity and Happiness – GitHub
When and How Artificial Intelligence Augments Employee Creativity – Multiple Authors
Experimental Evidence on the Productivity Effects of Generative AI – MIT
How Will Language Modelers like ChatGPT Affect Occupations and Industries? – Multiple Authors