MIT model suggests that only one-quarter of potential tasks–and jobs–will be displaced by AI near term because of high costs
One of the biggest questions–and fears– about commercial use of AI tools is how many jobs will be lost and in what fields? Until now, that’s been very hard to calculate.
A newly published paper, Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision? co-written by MIT IDE research group lead Neil Thompson, offers a way to quantify the effect of smart devices and automation of tasks. The proposed AI task automation model considers several factors including: The level of performance needed to do a task; the technical capabilities of the AI system needed to reach that performance, and the economics of whether to build such a system.
When these factors are considered, the result is a model that quantifies which tasks are technically feasible and economically attractive to automate — and which are not. Overall, the findings suggest that AI job displacement “will be substantial, but also gradual.”
While the study focused narrowly on tasks that could adopt computer vision, the findings were surprising and hopeful: Only 23% of worker wages now paid for vision-related tasks would be attractive to automate with computer vision AI “because of the large upfront costs of AI systems.”
This slower roll-out of AI could accelerate if costs fall rapidly or if it is deployed via AI-as-a-service platforms,” the paper noted, but in general, far fewer tasks will be automated than previously suggested — and in turn, that signals far less labor disruption.
“AI will deliver on its promise of greater productivity and it also poses a threat of worker displacement,” Thompson said. “Both will probably occur, but it depends on how quickly the technologies are adopted.” However, “even with rapid decreases in cost of 20% per year, it would still take decades for computer vision tasks to become economically efficient for firms.”
The paper was co-written by Maja S. Svanberg and Wensu Li of MIT, Martin Fleming of The Productivity Institute, and Brian C. Goehring from IBM’s Institute for Business Value.
MIT IDE Content Director, Paula Klein, asked Thompson to describe some of the highlights of the research in the following interview.
IDE: It seems that you are offering a more nuanced scenario — it’s not entirely doom nor is it a rosy optimism. What was the most surprising finding of the research?
Neil Thompson: We do find a middle ground: There is significant automation that will occur in the next few years, but much of it could easily take a decade or so — similar to how many technologies have spread throughout the economy.
The most surprising finding is the big difference that results based on how you analyze the data. When when you analyze AI at the level of broad technical compatibility– what jobs are vulnerable to replacement by AI, usually called AI Exposure — and when you require that deployments be economically attractive, the findings vary widely. The former assumes that all tasks that can be automated, will be. But we challenge that view, finding that only 1 in 4 of the tasks (23%) with broad technical compatibility are economically attractive to deploy today.
Framing it another way, today’s firm-level computer vision only has an economic advantage in 23% of vision tasks and barriers to AI-as-a-service deployments exist, therefore, we will most likely need to see a sharp reduction in costs for computer vision to replace human labor near term.
[See Figure 1]
IDE: How is the AI task automation model “end-to-end?” and how does it differ from previous models and assessments?
Thompson: Our model starts with the performance of real tasks in the economy, asks what AI system would be needed to automate them, and then finally considers the adoption decisions by businesses. The ability to model this entire process results in notably different answers than prior research.
IDE: Why the focus on computer vision tasks? Does that limit the scope of your study?
Thompson: In some ways, yes. But that limitation also means that we can analyze things more deeply than we could if we tried to consider all types of automation at once. In computer vision — tasks such as checking products for quality at the end of a factory assembly line or scanning medical imagery for anomalies — cost modeling is more developed, so we can estimate the cost of vision systems, which is central to our analysis. We can next ask how applying our model to automation with other technologies (e.g. large language models) would differ and how it will be similar.
IDE: Can you describe which kinds of tasks/jobs will be easy to automate and which won’t?
Thompson: Whether it’s economically attractive to automate a task with an AI system depends both on the cost of deploying the AI and the benefits to the firm of avoiding it. AI systems that are cheaper are easier to deploy, so we predict more automation when the accuracy needed from the system or the complexity of the task being done are lower.
The benefits to firms of replacing workers comes primarily from avoided labor costs;tasks done by more workers, or those with higher wages, tend to be more attractive to automate.
[See Figure 2 for the impact of computer vision on certain business sectors]
IDE: AI rollouts and ramp ups are so rapid these days, how confident are you that your model will hold — say 12 months from now? What if ChatGPT and other machine learning languages accelerate beyond your models and computing is able to keep pace?
Thompson: While AI systems are certainly rolling out quickly, their improvements are remarkably predictable, as work in our lab and others demonstrate. So, unless AI progress takes a dramatically different turn, we can be confident that our model will still be a useful guide in the coming years.
IDE: How can business leaders use these results in their AI planning efforts?
Thompson: Our results can help business leaders understand both the scope and timing of when tasks will become economically attractive, which can help them plan both production and HR decisions. The research also points to another important adjustment that businesses will need to make: Many tasks are only attractive to automate in the near term if platforms emerge that sell to many participants in an industry. This means that businesses need to start preparing themselves for the possibility that tasks that have traditionally been done inside the firm will be outsourced.