Machines will steal our jobs” is a fear often expressed in an era of rapid technological change. This anxiety has strongly resurfaced with the advent of large language models (e.g. ChatGPT, Bard, GPT-4) that demonstrate remarkable abilities in tasks where only humans were previously proficient. A recent MIT study, titled “ Beyond exposure to AI: What tasks are profitable to automate with computer vision?” » found that around 50% of tasks could be at least partially automated with large language models (LLM). If automation on this scale were to happen quickly, it would represent significant disruption to the workforce. Conversely, if this level of automation were to occur slowly, the workforce may be able to adapt as it has during other economic transformations (e.g. moving from agriculture to manufacturing industry).
The degree of “exposure to AI”
Therefore, making good policy and business decisions requires understanding how quickly AI will automate tasks. Although there is already evidence that AI is changing job demand, most of the concerns surrounding it come from predictions about exposure to AI that rank tasks or skills based on their potential for employment. automation, measured by various proxies. It is important to note that almost all of these predictions are vague about the timing and extent of automation, as they do not directly consider the technical feasibility or economic sustainability of AI systems, but rather use similarity measures between tasks and AI capabilities to indicate exposure.
The only exception in the literature that we are aware of is a McKinsey report that estimates AI adoption between 4% and 55%. With such imprecise forecasts, it is difficult to know what conclusions should be drawn. AI exposure models also confound predictions about full task automation, which is more likely to displace workers, with partial automation, which could increase their productivity. Separating these effects is essential to understanding the economic and political implications of automation.
What does the MIT study contain?
In this study, MIT addresses three important gaps in AI exposure models to construct a more economically robust estimate of task automation. First, workers familiar with the final tasks were interviewed to understand what performance would be required from an automated system. Second, the cost of building AI systems capable of achieving this level of performance was modeled. This cost estimate is essential for understanding AI deployment, as technically demanding systems can be extremely expensive. Finally, the decision “whether AI adoption is economically attractive” was modeled.
The result is the first end-to-end AI automation model.
The example of the bakery
A simple hypothetical example clearly illustrates these considerations. Take the example of a small bakery that is evaluating whether to automate with machine vision . One of the bakers' jobs is to visually check ingredients to ensure they are of sufficient quality (e.g., not spoiled). This task could theoretically be replaced by a computer vision system by adding a camera and training the system to detect spoiled food. But even if this visual inspection task could be separated from other parts of the manufacturing process, would it be cost effective to do so?
O*NET data from the Bureau of Labor Statistics suggests that checking food quality makes up about 6 percent of a baker's duties. A small bakery with five bakers earning a typical salary ($48,000 each per year) could therefore realize potential labor savings by automating this $14,000 per year task. This figure is much lower than the cost of developing, deploying and maintaining a computer vision system and therefore we would conclude that it is not economical to replace human labor with an AI system in this bakery. .
The conclusion from this example, that human workers are economically more attractive to businesses (especially those that are not large-scale), turns out to be widely held. We find that only 23% of workers' compensation claims "exposed" to AI computer vision would be profitable for companies to automate due to the high upfront costs of AI systems. The economics of AI can be made more attractive, both by reducing deployment costs and increasing the scale at which deployments are made, for example by launching AI platforms as a service and.
Overall, the model put together by MIT shows that job loss due to AI computer vision, even just in visual tasks, will be less than the workforce turnover already present in the market, suggesting that the replacement of labor will be more gradual and sudden .
The MIT study is structured as follows: Section 1 presents frameworks for estimating which tasks are economically attractive to automate; Section 2 presents the results; Section 3 examines how job replacement AI could proliferate; Section 4 discusses the importance of computer vision automation for other parts of AI; finally, section 5 brings together the conclusions.
Comments
Post a Comment