Will artificial intelligence automate human jobs? Those are three questions a new study released this morning from the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) attempts to answer. There have been many attempts to extrapolate and predict how today's artificial intelligence technologies, such as large language models, will impact people's lives and the economy as a whole in the future.
Goldman Sachs estimates that artificial intelligence will automate 25% of the entire labor market in the next few years. According to McKinsey, by 2055, nearly half of all jobs will be driven by artificial intelligence. An investigation by the University of Pennsylvania, New York University, and Princeton University found that ChatGPT alone could affect about 80% of jobs. A report from employment agency Challenger, Gray & Christmas shows that artificial intelligence has already replaced thousands of workers.
But in their study, the MIT researchers sought to go beyond what they call "task-based" comparisons to assess how feasible it is for AI to fill certain roles and how likely it is that companies will actually replace workers with AI technology.
Contrary to expectations, MIT researchers have found that most jobs previously thought to be at risk of being replaced by artificial intelligence are not actually "economically beneficial" to be automated - at least not yet.
Neil Thompson, a co-author of the study and a research scientist at MIT CSAIL, said the study's main implication is that the coming AI disruption may happen more slowly and less dramatically than some critics suggest.
"Like many recent studies, we found that artificial intelligence has great potential for automating tasks," Thompson said in an email interview with TechCrunch. "But we were able to show that automating many of these tasks is not yet attractive."
It is important to note that this study only looked at jobs that require visual analysis, i.e. jobs that involve tasks such as checking product quality at the end of a production line. The researchers did not investigate the potential impact of text and image generation models such as ChatGPT and Midjourney on workers and the economy; they left that question to follow-up research.
In conducting the study, researchers surveyed workers to understand what tasks would be needed for an AI system to completely replace their jobs. They then modeled the cost of building an AI system capable of doing all of these tasks, and modeled whether businesses - particularly "non-farm" businesses in the United States - would be willing to pay the upfront and operating costs of such a system.
Early in the study, the researchers gave the example of a baker.
According to the U.S. Bureau of Labor Statistics, bakers spend approximately 6% of their time checking food quality, and AI can (and is) automating this task. A bakery with five bakers earning an annual salary of $48,000 could save $14,000 if it could automate food quality testing. But the study estimates that a simple, scratch-made AI system to accomplish this task would cost $165,000 to deploy and $122,840 per year to maintain... and that's just the low end.
"We found that only 23% of the wages paid to humans performing vision tasks are economically attractive for AI automation," Thompson said. "Humans are still the better economic choice for these jobs."
The research now takes into account self-hosted AI systems sold through vendors such as OpenAI, which only need to be fine-tuned for specific tasks rather than being trained from scratch. But even if a system costs as little as $1,000, there are many jobs -- albeit low-wage ones and ones that rely on multitasking -- that it doesn't make financial sense for businesses to automate, according to the researchers.
"Even if we only consider the impact of computer vision on vision tasks, we find that job losses are lower than those already seen in the economy," the researchers wrote in the study. "Even if costs decline at a rapid rate of 20% per year, it will still take decades for computer vision tasks to become economically viable for businesses."
The researchers acknowledge that the study has some limitations. For example, it does not consider situations in which AI can augment rather than replace human labor (such as analyzing an athlete's golf swing), or create new tasks and jobs that did not previously exist (such as maintaining AI systems). Additionally, it does not take into account all the cost savings that pre-trained models like GPT-4 could bring.
We can't help but wonder whether the researchers felt pressure from the study's sponsor, the MIT-IBM Watson AI Lab, to reach certain conclusions. The MIT-IBM Watson Artificial Intelligence Laboratory was established by IBM with an investment of US$240 million for a period of 10 years. But researchers assert that this is not the case.
"We were motivated by the overwhelming success of deep learning, the dominant form of artificial intelligence, in many tasks, and we wanted to understand what this means for the automation of human jobs," Thompson said. "For policymakers, our findings should reinforce the importance of preparing for the automation of jobs with AI... But our findings also reveal "This process will take years or even decades to unfold, so there is time for policy measures to be put in place. For AI researchers and developers, this work demonstrates the importance of reducing the cost and scope of AI deployment to make AI economically attractive for enterprise automation."