How will AI impact jobs?

Since the public launch of ChatGPT in November 2022, a familiar question has resurfaced with renewed intensity: will artificial intelligence (AI) destroy jobs? The concern itself is hardly new, but generative AI gives it a distinct edge. Industrial machines mainly threaten physical labour. Software mainly disrupts routine tasks. However, generative AI models now impact language, analysis, coding, images and elements of intellectual work. This shift explains why today’s debate feels different: for the first time, technology potentially directly affects a broad share of skilled services employment.

For investors, this issue goes beyond social or political considerations. It cuts directly into productivity, profit margins, wages, consumption and, ultimately, potential growth. An AI shock that destroys jobs without creating alternative income streams would weigh on demand. An AI-driven productivity boost without widespread job losses would, conversely, support corporate earnings. Between these two outcomes lies a more subtle scenario: aggregate employment may remain broadly stable, while labour reallocates across sectors.

Simply put explores this hypothesis, examining whether sectoral employment data has already been affected by the arrival of large language models (LLMs) such as ChatGPT.

Read also: Markets look beyond the Iran crisis as AI momentum and earnings build

How technological progress shapes work

Economists have long studied the relationship between technological progress and employment. A classic framework is Alfred Sauvy’s spillover theory¹, also developed by Jean Fourastié, which describes how productivity gains eliminate jobs in some activities while creating others elsewhere. Rather than disappearing altogether, employment shifts across sectors – historically from agriculture to industry, and from industry to services. This intuition is not uniquely French. In the Anglo‑Saxon tradition, Allan Fisher and Colin Clark offered a closely related view through the three‑sector theory, which shows that as economies develop, labour gradually moves toward services. The core idea is the same: technological progress does not necessarily reduce the amount of work; it reshapes where work is done. This framework is particularly useful when thinking about AI today. The key question is not just about how many jobs will be destroyed, but which sectors will absorb labour next.

More recent work reframes this logic in terms of tasks. Daron Acemoglu and Pascual Restrepo distinguish between a displacement effect, where automation replaces labour in specific tasks, and a reinstatement effect, where new tasks emerge and create fresh demand for labour. Generative AI complicates this balance, as it not only affects repetitive work but also cognitive activities. Estimates suggest that around 80% of the US workforce could see at least 10% of their tasks being affected by AI tools, while roughly 20% could see half of their tasks being exposed. These figures measure exposure rather than job losses, but they underline the scale of a potential shift.

Read also: How vulnerable are US corporate earnings to an oil supply shock?

Early empirical evidence is more nuanced. Studies show that AI tools can raise productivity, particularly for less‑experienced workers, acting as a complement rather than a substitute. International institutions echo this mixed picture. The IMF, for example, estimates that around 40% of global employment is exposed to AI, rising to nearly 60% in advanced economies. Some jobs may benefit from productivity gains, while others may face weaker demand for labour, lower wages or slower hiring.

Three conclusions emerge:

1. Task destruction should not be confused with job destruction
2. Sectoral dynamics matter more than aggregate employment figures
3. While it remains too early to establish full causality, it is already possible – and meaningful – to search for AI-related impacts on employment trends.

AI-related impacts visible across most sectors

To do this, we analysed 866 US sectoral employment series drawn from the establishment survey. For each series, we estimated the pre‑ChatGPT trend and tested whether it has changed significantly since December 2022 (the first full month following ChatGPT’s release). To limit contamination from pandemic‑related distortions, the period from March 2020 to November 2022 was excluded.

Our conclusion is unambiguous: sectoral employment trends have shifted.

Across levels of aggregation, around three-quarters of all series show a statistically significant change. These changes are widespread and are not confined to a narrow set of technology‑related industries. Currently, most are negative, but a meaningful minority – around 20% – are positive. However, these changes do not imply that AI is already destroying employment in aggregate terms. Total employment still depends on the business cycle, interest rates, demand, public policy and firms’ adjustment speed. What the data does suggest is that the labour market has entered a phase of active reallocation.

Some sectors – often linked to digital workflows, administrative functions or post‑pandemic normalisation – show weaker trends. Others, more closely tied to physical presence, healthcare, infrastructure or local services, show relative improvement. Importantly, the sectors benefiting from this shift are neither small nor marginal.

FIG 1. Share of US sectoral employment series showing a trend break since the launch of ChatGPT²

Where is change most visible?

Looking more closely at detailed industry groups yields a clearer picture.

The greatest negative changes tend to appear in sectors where employment growth was particularly strong before 2020 or closely tied to post‑COVID-19 dynamics. Warehousing, couriers, temporary help services, business support, computer systems design and data infrastructure are prominent examples. Many sit close to digital value chains or administrative functions exposed to automation.

By contrast, positive breaks are more frequently observed in sectors tied to physical needs, healthcare, local transport, infrastructure and certain public services. Nursing and residential care, social assistance, passenger transport and selected infrastructure‑related activities stand out. The common trait is that they are harder to automate fully, as they rely on physical presence, human interaction, real‑world logistics or regulatory constraints – such jobs may eventually be disrupted by robotics, but this is not part of the current technological revolution.

Not all negatively affected sectors are those instinctively associated with AI. Some reflect cyclical normalisation or shifts in demand. This highlights an important limitation: this analysis identifies employment shifts since the launch of ChatGPT, but not pure causality from ChatGPT. Still, the breadth and consistency of the results suggest that AI is interacting with a labour market already in transition, and may be amplifying that transition rather than initiating it.

The most reasonable conclusion, therefore, is not that AI is already destroying jobs on a large scale, but something more nuanced. Employment trend changes are widespread and more often negative than positive, yet positive sectors clearly exist and are meaningful. This is precisely what historical theories of structural change would predict: work does not disappear, but employment growth is redistributed across activities.

FIG 2. Sectors with the largest positive and negative employment trend breaks since ChatGPT launch³

Simply put, artificial intelligence is not yet eliminating jobs at the macro level, but it is already reshaping where employment growth occurs.

Macro/nowcasting corner

The most recent evolution of our proprietary nowcasting indicators for global growth, global inflation surprises, and global monetary policy surprises is designed to track the recent progression of macroeconomic factors driving the markets.

Our nowcasting indicators currently show:

• Our world growth nowcaster has increased this week, with a surge in China signals coming from improved export data. The nowcaster is currently situated in a low but rising regime
• Our inflation indicator has increased globally, with a particularly strong rise in the eurozone as costs rose
• Similar to growth, our monetary policy nowcaster has increased, mainly reflecting the rise in China amid stronger export data.

World growth nowcaster: long-term (left) and recent evolution (right)

Reading note: LOIM’s nowcasting indicator gather economic indicators in a point-in-time manner in order to measure the likelihood of a given macro risk – growth, inflation surprises and monetary policy surprises. The nowcaster varies between 0% (low growth, low inflation surprises and dovish monetary policy) and 100% (the high growth, high inflation surprises and hawkish monetary policy).