AI automation is likely to be shaped as much by cost as by capability, research from MIT FutureTech suggests, with firms expected to automate work only when deploying the technology is cheaper than employing people to do it.
The finding challenges measures of “AI exposure” that focus on tasks systems could theoretically perform. Researchers examined whether businesses would actually pay to build, deploy and maintain those systems, rather than simply whether the technology was capable of doing the work.
In a study covering about 1,000 computer-vision tasks across roughly 800 US occupations, only 23% of the wages associated with those tasks were judged worth automating at the costs at the time.
A key obstacle is the fixed cost of developing an automated system. The expense can be broadly similar whether the technology replaces a small part of one worker’s day or the work of hundreds of employees.
A large bread factory, for example, might find it worthwhile to use computer vision to identify spoiled ingredients because of the scale of its production. A small bakery would be unlikely to recover the same upfront cost through labour savings.
Generative AI lowers the cost barrier
Generative AI presents a different calculation because general-purpose models can be accessed through payments based on usage, rather than requiring every business to commission a bespoke computer-vision system.
MIT FutureTech’s newer research tested the quality of the work produced by such models. Experienced workers assessed more than 6,000 text-based tasks drawn from the US Department of Labor’s occupational database, producing more than 60,000 evaluations.
The most capable models generated work that could be used without changes about two-thirds of the time. However, that overall success rate does not show whether a system performs some tasks consistently while failing at others, or produces acceptable results only intermittently.
Businesses must therefore factor in the cost of checking and correcting AI-generated work, as well as the expense of completing tasks the system cannot handle. The gap between work that is “usually right” and work that is dependable can be particularly costly.
For many tasks, the research suggests, AI is more likely to act as a collaborator than an outright replacement. Its capabilities are improving across a wide range of activities, but that progress does not necessarily mean an entire occupation will disappear suddenly.
Automation can still alter the value of the work left for people. When GPS became widely available on basic mobile devices, the expertise required by taxi drivers fell, along with their wages. In contrast, when software took over routine checking by proofreaders, the remaining work required greater expertise and wages rose.
The effect of AI on workers will therefore depend on which part of a job is automated. Employers will need to assess the full cost of using the technology, including supervision and correction, rather than relying on demonstrations of what it can do.
For policymakers and workers, the research points to the importance of tracking reliability and cost alongside technical capability. The tasks automated first will be those for which AI is cheaper than a person, and that pattern will determine whose remaining work becomes more valuable and whose becomes less so.