AI and the Uncertain Future of Jobs: An Impossible Prediction?

The Uncertain Impact of AI on Jobs

In a world where artificial intelligence (AI) is advancing at a breakneck pace, it would be extremely useful to determine which jobs, companies, and sectors are most exposed to this technology. The idea would be to assign scores, create charts, and link them to the advancements of large language models. Historically, every major technological wave has led to the destruction of certain jobs while creating new ones. But the question remains: which ones will be affected this time? For the past three years, many researchers have focused on analyzing census data, creating tables, and generating viral charts in an attempt to answer this question.

However, this predictive exercise seems destined to fail. Indeed, it involves trying to forecast the unpredictable.

Lessons from Past Technological Revolutions

To understand the complexity of the current situation, it is helpful to look at the major technological upheavals of the past. Some industries that seemed destined to suffer ultimately thrived, while others, apparently safe, were hit hard. Take accounting as an example: over the last century, this profession has been at the heart of automation with the invention of calculators, punch cards, mainframes, databases, PCs, spreadsheets, ERPs, and the cloud. Despite this massive automation, the number of accountants has continued to rise.

Large-scale survey data confirms this trend, but it is also observable at a more micro level. A specific chart shows that 50 years of financial automation have not harmed the market for accountants. If an analysis had been conducted on professions exposed to computer automation, accounting would have topped the list. Dan Bricklin, in the late 1970s, already mentioned how accountants used VisiCalc to accomplish in a few days tasks that previously took a month. Yet, the profession continued to grow.

The Paradoxes of Automation

Three key elements emerge from this analysis. First, technology is not the only variable at play: regulatory changes have created new accounting requirements, leading to an increase in the hiring of accountants. Second, the Jevons Paradox, often discussed in the context of automation, plays a crucial role: making a task less costly can lead to doing more of it for the same cost, or investing more for a greater return on investment. For example, if a discounted cash flow (DCF) calculation goes from taking a week to 30 seconds, it is likely that more of these calculations will be performed. Thus, exposure to automation could mean more work, not less.

Third, automating a costly and time-consuming task can unlock new opportunities. When analysis becomes cheap and easy, it is performed more frequently, often in new forms. Today's accountants do not perform exactly the same work as those in 1970 or 1980, even if their title remains unchanged. Technology often starts by improving old methods but ultimately transforms the work fundamentally.

Evolution of Jobs and Titles

Examining census data reveals that the category "accountants and auditors" remains stable, but other financial jobs appear and disappear over time. For instance, the job of "billing, publishing, and calculating machine operator" emerged for a decade before disappearing. This could represent individuals who started as stock clerks, became "publishing machine operators," and then returned to their original position when software absorbed their tasks. Similarly, there is still a category for "data entry clerk," but not for "ERP operator." Thus, the same person may hold different titles over time, while "accountants" retain their title while performing different tasks.

The Impact of AI on Businesses

Another issue arises when examining the data: the work may remain unchanged, but the business can evolve. The internet did not fundamentally change what it took to be a good journalist or an A&R scout, but it transformed the underlying business model. Journalism, for example, was funded by a manufacturing and distribution operation with a local monopoly on classified ads, while record labels relied on the manufacturing and shipping of physical media. These structural changes would not be captured in an analysis of jobs.

AI is likely to have a similar impact. Many people hold jobs that are minimally exposed to AI, but whose companies depend on other positions heavily affected by this technology. AI will make some previously costly tasks very cheap or free, potentially unlocking new opportunities while destroying others.

The Unpredictable Effects of Technologies

Continuing the analysis of the unpredictable effects of past technologies, how do we integrate innovations like Uber into an analysis? In the 2000s, while I was working in the mobile sector, we talked a lot about location data, but no one anticipated that it could disrupt the taxi industry. A more efficient distribution was conceivable, but not a radical change in the nature of work. If we had assessed "exposure to the internet" by profession in 1995 or "exposure to smartphones" in 2005, would we have included taxi drivers?

The problem with using resources like O*NET to analyze job automation is that they do not capture the complex and unpredictable transformations of work. They do not reveal how work may evolve with automation, nor how it can be altered by innovations in other fields.

The Complexity of Job Descriptions

A more fundamental problem lies in the difficulty of creating a complete and useful job description. O*NET descriptions remind us of the failure of expert systems, where it was thought that logical steps could be used to build an AI capable of recognizing images or translating languages. Theoretically, one can describe how a machine recognizes a cat or what an associate in a law firm does, but in practice, these tasks are too complex to be described that way. Sometimes, work boils down to a task that can be transformed into a button, but that is rare. Work is often a complex entanglement of tasks that are difficult to articulate.

Aaron Levie, CEO of Box, referred to this as a variant of Gell-Mann amnesia. We understand the complexity of our own field and how AI might not fully address it, but we forget this in other areas. We see a Claude model for a PowerPoint or a legal project and think that consultants and lawyers are threatened. Yet, when we hire Bain, BCG, or McKinsey, we get a few slides, but that is not what we are paying for, just as when we buy software, we receive code, but that is not the product.

The Limits of Predictions

The argument against this analysis would be to say that, despite exceptions, it is directionally correct to think that jobs involving many repetitive office tasks are the most exposed. But we do not know if the exceptions are more significant than the rule. If we had analyzed the internet in 1995, we could have said it would destroy the value of physical distribution for media, which was directionally correct, but this meant different things for record labels, newspapers, television networks, and movie studios.

The question is whether we are trying to make a prediction with real value or simply observing a truth. On average, we are all dead. Half of the jobs analyzed could be completely unaffected, and other sets of jobs could be transformed without being detected.

Uncertainty About the Future

Some time ago, someone criticized my work by saying that I always ended up saying "it depends." But at such an early stage of a fundamentally new technology, any specific prediction about a particular field will only be correct by chance: it really does depend. As Yogi Berra said, "It's tough to make predictions, especially about the future." We can certainly point to frameworks and mental models about how this might work, and we can highlight what has happened the last six times we have gone through this type of change. We can even say things that are probably directionally correct. But as soon as you try to quantify this, and model it job by job and industry by industry, and make pretty radar charts, you are mistaken, because you do not actually know what those jobs are today, and you do not know how they will change. At a minimum, you need to ask.