Over the past few weeks, several notable executives from large employers, such as Ford and JP Morgan Chase, have provided predictions that AI will result in massive white-collar unemployment.
Some technicians, including people from Amazon, Openai and Meta, have admitted that the latest wave of AI, known as agent AI, is closer to radically transforming the workplace than previously expected.
Dario Amodei, CEO of AI Business Humanity, said nearly half of the entry-level white-collar jobs in technology, finance, legal and consulting could be replaced or eliminated by AI.
Christopher Stanton, Associate Professor of Business Administration at Marvin Bower at Harvard Business School, studies AI in the workplace and teaches MBA courses. In this compiled conversation, Stanton explains why the latest generation of AI is evolving so quickly, and how it shakes up white-collar work.
Currently, several top executives predict that AI will eliminate a large number of white-collar jobs much faster than previously expected. Does that sound accurate?
I think it's too early to tell. If you were pessimistic in the sense that you are worried about labor market disruption and the depreciation of skills and human capital, then if you see what we think AI can do with the tasks that white-collar work workers can do, that overlap affects about 35% of the tasks we see in labor market data.
“My personal trends – this isn't necessarily based on a deep analytical model – that means policymakers are very limited in their ability to do something here unless they're through subsidies or tax policies.”
The optimistic case is that if you think a machine can perform several tasks rather than all, the task that a machine can automate or perform is to free people to focus on different aspects of the job. You may see 20 or 30% of the tasks that a professor can do with AI, but the other 80 or 70% complement what AI generates. They are both extremes.
In fact, it's too early to tell how this is shaking, but we've seen at least three or four things that we might suspect are reasonable to believe that the view that AI has a more destructive effect on the labour market.
One of these is that computer science alumni and STEM alumni generally struggle to find work today than they have in the past. This may be consistent with the view that AI does a lot of the work that software engineers have done before.
For example, if you look at reports from Y Combinator, or from other high-tech sectors, it appears that much of the code for early-stage startups is written by AI. Four or five years ago, that would have been completely untrue. So we are beginning to make sure that the intake of these tools is consistent with the stories from these CEOs. That's one part.
The second part is that even if you don't necessarily consider displacement, AI can still be thought of as a result of wages.
There are two ways to compete about where this goes. Some of the early evidence, such as AI rollouts and contact centers and frontline work, suggest that AI reduces inequality among people by lifting the performer's lower tail.
Some of the best papers on this look at randomized rollouts of conversational AI tools or chatbots and frontline call center work, showing that low-performing workers or workers at the bottom of the productivity distribution will benefit from the AI deployment tool. If these workers have knowledge gaps, AIS fills in for knowledge gaps.
What is driving the acceleration speed used by this generation of AI has evolved and driven by the companies?
There are a few things. I have a paper with Microsoft researchers looking at the impact of AI adoption and AI rollouts in the workplace. Our tentative conclusion was that it would require a lot of adjustments to actually see some of the productivity effects of AI, but it immediately affected individual tasks, such as email.
“Our tentative conclusion was that it would require a lot of adjustments to actually see some of the productivity benefits of AI, but it immediately affected individual tasks like email.”
One of the messages of that paper that is not necessarily widely spread is that it is probably part of the most rapid technology.
In our sample, half of participants who accessed this tool from Microsoft used it. And the feature was incredible.
My guess is that one of the reasons executives didn't predict this is that it is a very fast spreading technology. You're looking at different people from different teams running their own experiments to understand how to use it, and some of those experiments will generate insights that were not expected.
The second thing that accelerated the usefulness of these models is a type of model called the Thinking Chain Model. The earliest versions of the generator AI tools tended to provide hallucinations and inaccurate answers. Inference of the type of thinking is intended to do error correction on the spot.
So, rather than providing answers that could be affected by errors or hallucinations, the model itself is provided with a prompt saying “Are you sure about that? Double check”. Models with thoughtful inferences are far more accurate, far more accurate and unaffected by hallucinations, especially for quantitative and tasks that involve programming.
As a result, there is a considerable amount of penetration in early stage startups that use natural language queries to code, or today called “atmosphere coding.” These vibe coding tools have built-in error corrections that allow you to create practically usable code as a result of these feedback mechanisms built by model designers.
The third thing driving major adoption, especially in the world of technology, is that model providers have built tools to deploy code. Anthropic has a tool that allows you to write code based solely on queries or natural language, and then you can deploy it with the tools of humanity.
There are other tools like Cursor and Replika, and ultimately you can tell your machine to create technical software with limited technical background. It doesn't necessarily require specific technical tools, making deployment much easier.
This is back to what I had told you before. This means you've seen a lot of experiments and you've seen a huge spread. And one of the reasons you saw huge spread is that you have these tools and these models that allow people who don't have domain expertise to build things up and understand what they can build and how they can do it.
Which type of work is most likely to see the change first? You mentioned writing code, is there anything else?
I have never seen any immediate data suggesting job losses, but it is easy to imagine that employment benefits could be seen in any knowledge work, at least in theory.
In reality, looking back at the history of predictions about AI and unemployment makes them extremely difficult.
There was a lot of debate in 2017, 2018 and 2019 about whether training radiologists should be stopped. However, radiologists were as busy as ever, and we didn't stop training them. They're doing a lot more, and one reason for this is that the cost of imaging has decreased. And at least some of them have some AI tools at their fingertips.
So, in a sense, these tools could potentially employ some tasks that humans have done, but they also reduce the costs of doing new things. Therefore, that net net is extremely difficult to predict. Because doing more things that complement what people in those professions do, more people may need to do a slightly different task.
So I think it's too early to say that we don't necessarily see net displacement in any industry or in the whole.
If AI suddenly removes the majority of middle-class Americans from work or makes education and skills much less valuable, it can have a devastating impact on the US economy, politics and generally quality of life. Are there any policy solutions that lawmakers are thinking today to help them get ahead of this ocean change?
My personal trends – this isn't necessarily based on a deep analytical model – is that unless policymakers are through subsidies or tax policies, they have very limited ability to do anything here. Whatever you do to support employment, you will see competitors who are more agile, less costly and probably dynamically compete with people, at low cost, without the same legacy labor stack.
If you don't necessarily understand technology at this point, it's not so clear that policy interventions should be. My guess is that policymakers' relief will be a relief for ex-posts, not the original relief. My doubt is a better safety net policy, and a better retraining policy will become the tool that is playing, rather than trying to prevent the adoption of technology.
Source link
