Which human occupations will AI automate, and when?
MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) released a research report this morning to answer those three concerns.
Many attempts have been made to predict how AI technologies like massive language models will affect people’s livelihoods and economies in the future.
AI might automate 25% of the job market in the next several years, according to Goldman Sachs. McKinsey predicts AI will drive nearly half of labor by 2055. ChatGPT alone might affect 80% of employment, according to a Penn, NYU, and Princeton survey. According to Challenger, Gray, and Christmas, AI is replacing thousands of jobs.
But in their analysis, the MIT researchers intended to move beyond what they define as “task-based” comparisons and assess how plausible it is that AI will do particular functions—and how likely corporations are to actually replace labor with AI equipment.
The majority of vocations previously recognized as at risk of AI displacement aren’t “economically beneficial” to automate, at least not now, according to MIT researchers.
Neil Thompson, a research scientist at MIT CSAIL and study co-author, thinks the AI impact may be delayed and less significant than some pundits expect.
“Like much of the recent research, we find significant potential for AI to automate tasks,” Thompson emailed Eltrys. “But we can show that many of these tasks are not yet attractive to automate.”
Importantly, the study primarily examined visual analysis professions like assessing items for quality at the end of a production line. The researchers left it to future studies to determine how text- and image-generating algorithms like ChatGPT and Midjourney affect workers and the economy.
This study polled workers to determine what an AI system would need to do to totally replace their occupations. They then evaluated the cost of constructing an AI system that could perform all this and if “non-farm” U.S. enterprises would pay both the upfront and running costs.
Early in the investigation, researchers used a baker.
According to the U.S. Bureau of Labor Statistics, bakers spend 6% of their time monitoring food quality, which AI can automate. A bakery with five $48,000-a-year bakers may save $14,000 by automating food quality tests. The study indicates that a basic AI system may cost $165,000 to implement and $122,840 per year to maintain.
“We find that only 23% of human wages for vision tasks would be economically attractive to automate with AI,” Thompson added. “Humans are still the better economic choice for these jobs.”
The report does account for self-hosted, self-service AI systems supplied by OpenAI that just need to be fine-tuned to specific tasks—not taught from scratch. Even with a $1,000 system, the researchers found several low-wage, multitasking-dependent positions that would not make economic sense for a corporation to automate.
“Even if we consider the impact of computer vision just within vision tasks, we find that the rate of job loss is lower than that already experienced in the economy,” the study authors wrote. “Even with 20% annual cost decreases, computer vision tasks would take decades to become economically efficient for firms.”
To their credit, the researchers acknowledge the study’s limitations. It ignores scenarios when AI may complement rather than replace human labor (e.g., evaluate an athlete’s golf swing) or generate new activities and employment (e.g., manage an AI system). Pre-trained models like GPT-4 can save money, but it’s not considered.
MIT-IBM Watson AI Lab funding may have pressured researchers to reach particular findings. A $240 million, 10-year grant from IBM, which wants to make AI seem harmless, funded the MIT-IBM Watson AI Lab.
Researchers say otherwise.
“We were motivated by the enormous success of deep learning, the leading form of AI, across many tasks and the desire to understand whether this would affect human job automation,” Thompson added. Our findings emphasize the need for policymakers to prepare for AI job automation. Our studies also show that this transition will take years or decades, giving policymakers time to act. This effort shows AI researchers and developers the relevance of lowering deployment costs and expanding deployment options. These will let enterprises employ AI for automation economically.
