The Economic Incentives of Automation

The Economic Incentives of Automation

Hey Everyone,

With the Generative AI trend of 2023 and general purpose robots having a good start to the year in 2024, automation and its adoption is on my mind a lot.


There have been some interesting papers in recent times around the future of work and incentives to automation. I asked Tobias Jensen of the publication Futuristic Lawyer to take a deeper look into this paper by MIT.


Working Paper by MIT, 45 pages.

Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision?

Mit Study

1.07MB ? PDF file

Download

The paper was a collaboration between MIT, IBM’s Institute for Business Value and the Productivity Institute.

If you enjoy the piece, take a look at his Newsletter.

Publication Specs: "Futuristic Lawyer"

Writing about the intersection of business, law and tech with a focus on social issues.

By Tobias Mark Jensen


Discount for Paid Yearly Subscriptions to get all of our Deep Dives


Get 15% off forever


By Tobias Jensen



AI’s Labor Market Impacts Are Slower Than Expected

One of my favorite topics to speculate and write about is AIs impact on work. What will the short-term implications of AI in various jobs and sectors be? A new working paper born out of a ten-year partnership between Massachusetts Institute of Technology (MIT) and IBM Watson AI Lab,? backed with a gift of $240 million from IBM (TechCrunch), casts a new light on the matter.?

In general, I think we should be skeptical about scientific research conducted or funded by big technology companies concerning subjects the companies have a vested interest in. Nowadays, genuine academic research and advanced corporate marketing are unfortunately becoming hard to distinguish from each other.?

In my post about AGI, I mentioned how a paper by Microsoft Research, Sparks of Artificial General Intelligence: Early experiments with GPT-4, hinted not so subtly that maybe, just maybe, OpenAI was on the path to developing truly intelligent AI with GPT-4. Conveniently for Microsoft, that would be an earth-shattering win, since Microsoft is rumored to attain 75% of OpenAI’s profits and will have a future ownership stake in the company of 49% (Semafor).? I am not saying that outside interests in fact did play a role in the research, but the dangers to objectivity are hard to overlook.?

The new study by MIT and IBM strikes me as thoroughly researched and satisfyingly documented with no signs of private corruption. Also, the results are important. In this post, we will look deeply into the authors' method, the study's results, and consider its implications. But first, let's briefly understand why other studies on AI's labor impacts have fallen short.?

Your Take ??

Loading...

The Challenge of Assessing AIs Labor Market Impact?

The working paper Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision? (from here on: Svanberg et al.). was published on January 22, 2024, by Neil Thompson, Maja S. Svanberg, and Wensu Li from MIT FutureTech, Martin Fleming from The Productivity Institute, and Brian C. Goehring from IBM's Institute for Business Value.?

Previous studies on AI’s labor market impacts have assumed that any work task an AI system is capable of performing is exposed to automation without considering the technical feasibility or the economic viability of automating such tasks. This is quite an important detail. Just because a work task can be automated does not mean that it is desirable or feasible to do so from an economic and/or technical perspective.?

Past research has made predictions about labor market impacts but has typically been vague about timelines. For example, two researchers from OpenAI, Tyna Eloundou and Pamela Mishkin, along with Sam Manning from OpenResearch, and Daniel Rock from University of Pennsylvania, were behind a study on GPTs potential impacts on the US labor market (Eloundou et al. 2023). It was published a week after GPT-4’s release in March 2023. The study found that 80% of the U.S. workforce could have at least 10% of their work tasks affected by GPTs, while around 19% of workers may see at least 50% of their tasks impacted.??

However, Eloundou et al. 2023 did not specify a timeline that describes when these predicted labor market impacts would occur. The timeline is crucial. If the predicted labor market impact happens slowly over many years, the market would have time to adjust, adapt, and absorb the changes, whereas if the impacts occur rapidly over a year or a few years, the effects would be enormously disruptive.?

A report by McKinsey, (Ellingrud et al. 2023) predicts the impact of automation on the US labor market, and includes a timeline, but is vague on the rate and the impact of AI adoption. The report predicts that by 2030, activities accounting for 30% of hours currently worked across the US economy could be automated. However, it estimates AI adoption in a very wide range between 4% to 55% so AIs future impacts are not clear from the study.?

As we shall see, Svanberg et al predicts that changes to the labor market caused by the impacts of AI will happen slowly over many years and that the impacts will not be as disruptive as predicted by other papers.?


The Methodology of Svanberg et al.?

Svanberg et al. only concerns vision-based work tasks because, according to the abstract, cost modeling in computer vision is more developed. However, the researchers believe that the economics of AI described in the paper applies to AI more broadly, including to foundation models and generative AI for language. Future research could look closer into this based on the method applied in Svanberg et al. that outlines what the researchers call “the first end-to-end AI automation model”. It includes three parts:

  1. Surveys of workers with domain expertise about what it would require of an automated system to perform a specific task.
  2. A cost model of building AI systems capable of reaching that level of performance.
  3. A model for the decision about whether adopting AI to automate a task is economically attractive.?


A simple example of how the model works from the paper: A small bakery is considering whether it should adopt an AI system to check if ingredients it uses are of a good enough quality. The bakery could theoretically adopt a computer vision system by adding a camera and training the system to detect if food has gone bad.? Data from O*NET (more on that right below) implies that checking food quality comprises roughly 6% of the duties of a baker. A small bakery with five bakers is making typically $48.000 each per year. That means the labor savings from automating this task is roughly $14.000 per year (6% of the five bakers' combined salary) which is far less than it would cost to train and adopt a computer vision system. In conclusion:? automating this task would not be profitable.?

The researchers fetched an overview of different work tasks on the O*NET Database. O*NET is a free and public database sponsored by the US Department of Labor that contains an overview of different occupations, tasks, skills, worker requirements, and more across the US economy. Included in O*NET was a list of 2.087 different "direct work activities" (DWA) out of which the researchers identified 420 tasks related to vision. This could, for example, be, “Assess skin or hair conditions,” “Examine patients to assess general physical condition,” “Inspect items for damage or defects,” or “Monitor facilities or operational systems”.?


To fill in the scientific literature gaps on AI’s labor market impacts, the researchers in Svanberg et al., centered their work around two questions:

  1. Exposure: Is it possible to build an AI model to automate this task?
  2. Economically-attractive: Would it be more attractive to use an AI system for this task than to have human workers continue to do it?

To answer the first question regarding a task’s exposure to automation, the researchers gathered information from domain experts recruited from an online crowd-sourcing platform, Prolific. The respondents were guided to choose the job they had familiarity with and answer questions about the vision tasks involved in the job. The researchers collected 9 responses per vision task on average.?

To answer the second question regarding if the automation of a task is economically attractive, the researchers examined the benefits and costs of using computer vision systems, as compared to the compensation of human workers currently doing those tasks. More specifically, they mainly looked at the costs of fine-tuning and deploying a computer vision system to perform a task, including fixed costs, performance-dependent costs, and scale-dependent costs in a given time span vs. the marginal cost of compensation per worker.??


Share


Key Findings and Observations

Now, the key finding in Svanberg et al. is that only 23% of visual-based tasks would be cost-effective for firms to automate. This is because of the large upfront costs of adopting AI systems. The job loss from AI computer vision will actually be smaller than the existing job churn seen in the market.?

Svanberg et al. does a very good at taking different considerations into account.?

Instead of building and adopting AI systems in-house, companies could buy AI-as-a-service from vendors in the market.? However, designing an AI system for a particular task that can be generalized to the needs of many different firms is technically challenging and expensive. Also, typically, companies would have to fine-tune a more generalized model to their specific needs. By using proprietary data to fine-tune a third-party vendor’s model, companies are exposing themselves to privacy and confidentiality risks.? Overall, the researchers find it unlikely that any third-party vendor could capture more than a fraction of the total market.

Another consideration is that the deployment of AI systems could decrease sharply in costs over the next few years. But even with a 50% annual decrease in costs, the researchers estimate that it will take until 2026 before half of the vision tasks have an economic advantage. By 2042 there will still exist tasks where human labor has an advantage over computer vision. With a more modest cost decrease of 20% per year, it would take decades for most computer vision tasks to become economically efficient for firms.

More broadly, according to data from the U.S. Census Bureau (2021), the researchers highlight that on average 11% of jobs in the private sector were destroyed annually between 2017 and 2019. On the other hand, many new jobs were created, leaving a net average job gain of 1.6% during the period. Although 23% of vision tasks are economically attractive to automate today, and we may be in for an initial shock to the labor market, the impact will likely be flattened out over time? as depicted in the figure below.

Even if the annual costs of computer vision systems drop by 50%, and if we assume that all vision-based tasks that are economically attractive for companies to automate will be automated in the same year, the percentage of vision task compensation that is lost every year will still not exceed 6-8% in the peak years.?

This seems to be very good news for wage earners and for humanity! And bad news for Sam Altman’s nefarious-sounding obsession to “replace the median human with AI”.??

What do you think?

Let me know in the comments.?


Read the Paper


Leave a comment

Kajal Singh

HR Operations | Implementation of HRIS systems & Employee Onboarding | HR Policies | Exit Interviews

4 个月

Well said. In 2018, Nedelkoska and Quintini expanded on Frey and Osborne's methodology to estimate the risk of job automation across 32 OECD countries. Their study found that 14% of jobs were highly vulnerable, 32% were somewhat less vulnerable, and 56% were not very vulnerable to automation. With a total workforce of approximately 628 million in OECD countries, their analysis suggested around 200 million jobs could be lost to AI and automation, but no specific time frame was provided. Additionally, the World Economic Forum (WEF) conducted surveys in 2016 and 2018, predicting the displacement of 75 million jobs by automation by 2022, with 133 million new roles emerging. However, counterarguments in the text challenge the immediacy of these predictions, asserting that the job loss and new job creation are unlikely to happen in the specified timeframe. The subsequent sections suggest that, by 2050, more global job losses due to automation and AI are expected, with the WEF's prediction of 133 million new jobs becoming plausible by 2045 as AI becomes ubiquitous. More about this topic: https://lnkd.in/gPjFMgy7

回复

"Absolutely, the advent of AI and robotics is thrilling! ?? As Albert Einstein once said, "Imagination is more important than knowledge." It's this imagination that drives our innovations and pushes us into the future. ???? #AI #TechInnovation"

回复
Brook Martin

"Business Operations Strategist – I collaborate with organizations to refine their processes, using consultative approaches that drive growth, improve efficiency, and enhance overall performance.

8 个月

fantastic article and review! well done!

要查看或添加评论,请登录

社区洞察

其他会员也浏览了