AI can already perform many tasks that once required people. But a new study from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) argues that technical ability is only one part of the automation story. For many employers, replacing workers with AI may still be too expensive to make sense.
The study does not claim that AI will have little effect on work. Instead, it suggests that the shift may happen more gradually than some forecasts imply, especially when companies must pay to build, deploy and maintain systems that can reliably take over real workplace duties.
Why the study looks beyond task lists
Many projections about AI and employment begin by asking which human tasks current AI systems can perform. The source article cites several examples of broad forecasts: Goldman Sachs estimates that AI could automate 25% of the entire labor market in the next few years, McKinsey projects that nearly half of all work will be AI-driven by 2055, and a University of Pennsylvania, NYU and Princeton survey finds that ChatGPT alone could impact around 80% of jobs.
The MIT researchers took a different route. Rather than only comparing tasks with AI capabilities, they asked whether it would be economically attractive for businesses to replace workers with AI in practice. That distinction matters because a task can be technically automatable while still being too costly, narrow or awkward to automate profitably.
Neil Thompson, a research scientist at MIT CSAIL and a co-author on the study, summarized the point directly: "Like much of the recent research, we find significant potential for AI to automate tasks," Thompson told TechCrunch. "But we're able to show that many of these tasks are not yet attractive to automate."
The focus is computer vision, not every kind of AI
The study has an important boundary: it examined jobs requiring visual analysis. That means work involving tasks such as inspecting products for quality at the end of a manufacturing line. It did not study the labor impact of text- and image-generating models such as ChatGPT and Midjourney.
That makes the findings narrower than the headline debate over AI and jobs. The research is not a full assessment of every workplace use of large language models, image generation or other systems. It is a focused look at whether computer vision tasks are currently worth automating from a business cost perspective.
To build that assessment, the researchers surveyed workers to understand what an AI system would need to do in order to fully replace their jobs. They then modeled the cost of building a system capable of those duties and considered whether U.S.-based "non-farm" businesses would pay the upfront and operating costs.
The bakery example shows the cost problem
The study uses a baker to show why automation is not simply a matter of whether AI can perform a task. According to the U.S. Bureau of Labor Statistics, a baker spends about 6% of their time checking food quality. That kind of quality check can be, and is being, automated by AI.
But the economics are less straightforward. A bakery with five bakers earning $48,000 per year could save $14,000 by automating food quality checks. The study estimates that a bare-bones AI system built from scratch for that task would cost $165,000 to deploy and $122,840 per year to maintain, even on the low end.
That mismatch between potential savings and system cost is central to the study's conclusion. Thompson said: "We find that only 23% of the wages being paid to humans for doing vision tasks would be economically attractive to automate with AI." He added: "Humans are still the better economic choice for doing these parts of jobs."
Lower AI costs still may not erase the gap
The researchers also considered self-hosted, self-service AI systems sold through vendors such as OpenAI. These systems may only need to be fine-tuned for a specific task rather than trained from the ground up. Even then, the study finds that many jobs would not make economic sense to automate.
This is especially true for work that is low-wage or depends on multitasking. If a person performs many different duties, automating one small visual task may save too little to justify the system. In those cases, businesses may have a technical option but no strong financial reason to use it as a replacement for labor.
The researchers write that, even when looking only at computer vision within vision tasks, "the rate of job loss is lower than that already experienced in the economy." They also write that "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."
What the limits of the research mean
The study's conclusions come with several limits. It does not examine situations where AI augments workers instead of replacing them, such as analyzing an athlete's golf swing. It also does not account for new tasks and jobs that may appear, including maintaining AI systems.
The research also does not include all possible cost savings from pre-trained models such as GPT-4. And because it centers on visual analysis, it leaves the economic impact of text- and image-generating models to later work.
The study was backed by the MIT-IBM Watson AI Lab, which was created with a $240 million, 10-year gift from IBM. The source article notes that IBM has an interest in how AI is perceived, but the researchers say their motivation was to understand what deep learning's success means for job automation.
For policymakers, Thompson said the findings reinforce the need to prepare for AI job automation while also showing that the process may take years, or even decades. For AI researchers and developers, the work points toward lowering deployment costs and expanding how AI systems can be used, because those factors will influence when automation becomes economically attractive for firms.