The narrative many of us have taken in goes as follows: once a machine is capable of performing your job, it will take over your role. It seems like an unyielding rule of nature. If the software can detect the flaw, sort the package, or evaluate the produce, then the human currently handling that task is on borrowed time.
This notion overlooks an important aspect. Being able to execute a task does not equate to it being financially viable. A machine may be entirely competent yet still be considerably more expensive than the individual already fulfilling that role, once one factors in the expenses involved in constructing, installing, and maintaining it. When a group at MIT calculated that total cost, the scenario appeared significantly different from the mass unemployment narrative.
The study we are referencing here examines just one facet of AI, hence it is not the final statement on the entire technology. However, it is thorough and reshapes a conversation that requires reshaping.
What the MIT group evaluated
In January 2024, MIT researchers released a working paper with an unexciting title: “Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision?” The crucial term is “beyond.” Most previous studies posed a more limited question: could an AI system, theoretically, carry out this task? That informs you of what is feasible. It reveals nothing about what a business would genuinely opt to implement.
The group, directed by Neil Thompson of MIT’s FutureTech initiative, simulated the genuine decision a company encounters. As Thompson outlines it, the model begins with the execution of real tasks, queries what AI system would be necessary to accomplish them, and then considers whether a business would actually pursue it. They concentrated on computer vision, the type of AI that analyzes and interprets images, partly due to its costs being simpler to calculate than those of other kinds.
Thompson expressed the headline conclusion straightforwardly. “In many instances, humans represent the more cost-effective method, and a more economically enticing approach, to perform work at present,” he told CNN. That “at present” is significant, and we will revisit it.
The statistics that undermine the alarm
“We discover that merely 23% of worker compensation ‘exposed’ to AI computer vision would be cost-effective for companies to automate due to the substantial initial costs of AI systems,” the report states. Approximately three-quarters of the tasks that computer vision could technically address are less expensive to retain with humans.
If we broaden the view to encompass the entire US economy, the proportion diminishes. Computer vision could technically automate tasks valued at approximately 1.6% of US worker salaries, excluding agriculture. However, when you tally the costs, merely around 0.4% of salaries would actually be more economical to automate at this time. A fragment of a fragment.
This disparity is evident at the job level as well. One analysis notes that while approximately 36% of US non-farm positions have at least one task a camera could manage, only about 8%