AI layoff announcements are a poor proxy for AI’s impact on jobs

This post was co-written with Sam Manning.

There have recently been several high-profile cases of companies announcing that they were laying off employees because of AI. A common response to these cases has been to treat them as direct indications of the effect of AI on the economy, and even to tally them up to measure total job loss due to AI.

However, company reports of AI-driven job loss are a pretty bad proxy for the actual effect of AI on employment, for four main reasons:

1. Self-reporting is unreliable. Companies have incentives to both over- and under-report AI-driven layoffs, so self-reported figures of AI-driven layoffs shouldn’t be taken at face value.

2. Job losses don’t capture the net effect within a company. Even if self-reports were reliable, they miss new job creation and changes in hiring (including hiring slowdowns due to reduced labor demand).

3. AI adoption in one company can affect jobs in other companies. Even if you could accurately measure the net effect of AI adoption on jobs within a company, changes in one company’s productivity can affect employment at other companies, such as when an early adopter outcompetes a rival and forces them to downsize.

4. AI can create entirely new jobs and businesses. New firms and roles can emerge, which are not captured in data from existing companies.

Accurately assessing AI’s impact on jobs matters: both overestimating and underestimating labor market effects can lead to misguided labor and AI policy. Company layoff announcements are an unreliable and incomplete basis for that assessment.

Company announcements miss a lot of detail

Self-reporting is unreliable

There’s no particular reason to expect self-reports of AI-driven job loss to be accurate. This is in large part because companies don’t have strong incentives to accurately represent the reason for job losses.

For example, companies might overstate job loss attributable to AI in order to cover up things like declining revenue or previous overhiring with something that makes them sound more productive and technologically advanced (AI integration).

Companies might also be inclined to understate AI job loss – for example, to avoid negative public backlash to automation.

It’s also hard to verify a company’s claims about what is causing job loss. Firm self-reports vary considerably in estimated adoption rates, and it’s not clear that managers and executives have full insight into how their workers are adopting AI in their workflows. Even if you could confirm that a company both increased its AI use and had layoffs around the same time, the relationship between the two isn’t obvious. For example, those employees might have been laid off anyway for reasons unrelated to AI use.

In practice, because it is hard to externally verify the relationship between AI adoption and layoffs, it is difficult to know how much to trust company claims of AI-induced firing. Indeed, many company-reported cases are pretty murky.

Job losses don’t capture the net effect within a company

Even if you had reliable estimates of the number of people in a company fired because of AI, it’s not obvious what the net effect on jobs would be in that company. There are two potential effects that layoff numbers would miss:

1. Productivity effects. While the jury is still out on AI’s firm-level productivity impacts, companies presumably adopt AI because they expect it will boost productivity. If productivity gains allow firms to expand output or enter new markets, they may hire more workers. For example, a company might replace some of the workers in a production plant with AI, but increase the number of production plants. Not accounting for productivity effects overestimates the negative effect of AI on jobs in the firm.

2. Changes in hiring patterns. Companies might change how they hire employees based on AI adoption, in ways that could either increase or decrease jobs within the company (holding changes in productivity constant).

   1. Companies adopting AI might be driven to hire new types of workers (or move existing workers into different jobs). For example, companies might hire systems engineers or cybersecurity experts to support their AI integration. Not accounting for hiring in new types of jobs overestimates the negative effect of AI on jobs in the firm.

   2. Companies might slow hiring (in addition to or instead of firing employees). For example, the CEO of Klarna announced that the company had mostly stopped hiring since 2023 because of AI (though, as noted, self-reporting of AI’s causal impact shouldn’t be treated as particularly accurate). Additionally, one recent study found that after firms hired workers into “AI-integrator” roles (signaling investment in AI adoption), they also slowed hiring for junior workers in AI-exposed positions. Not accounting for reduced hiring underestimates the negative effect of AI on jobs in the firm. It’s plausible that reduced hiring will account for a larger share of employment loss from AI than outright displacement, making this a potentially important gap missing from layoff-focused self-reports.

Even when layoffs are credibly attributed to AI, the overall employment effect within a firm is unclear and likely varies widely across companies.

AI adoption in one company can affect jobs in other companies

Some of the largest effects on jobs might flow through other companies. As noted, firms presumably adopt AI because they expect it to increase productivity. Greater productivity in company A can affect the number of jobs in companies B and C through a few mechanisms:

1. Within-market effects. Productivity gains in firm A can change market share among competitors in the same industry. This could decrease employment at competitor firms or push them to differentiate into a different niche (including a higher-employment one). Some specific examples:

   1. Example 1: Carmaker A’s adoption of AI allows it to produce cheaper/better cars. Other car companies lose market share. As a result, they leave the market or downsize, reducing employment in their companies.

   2. Example 2: Hotel chain A automates away a lot of their management structure, allowing them to offer cheaper budget-friendly hotel rooms. Hotel chain B, now less competitive in the budget hotel market, leans into luxury hotels, hiring more people for service jobs (housekeepers, restaurant workers, attendants, etc.), increasing employment in company B.

2. Cross-market effects. Productivity gains in firm A might change how consumers allocate spending across other industries. There are two effects here which can occur simultaneously, and which push in opposite directions for job loss in other industries:

   1. Income effects: Higher productivity in firm A might make its goods relatively cheaper. Lower prices free up purchasing power for consumers, who then might spend more money on other goods.

      1. Example: Carmaker A’s adoption of AI allows it to make cheaper cars. People spend less money on cars, and so have more money to spend on other things, like leisure activities (e.g., going to the theater). Employment in the performing arts rises.

   2. Substitution effects: If firm A’s goods are relatively cheaper, that might lead consumers to switch their spending towards buying more of A’s products, and spending less on other goods.

      1. Example: AI lowers the cost of producing streaming content, improving its price or quality relative to in-person entertainment. This could lead households to spend more on streaming subscriptions and less on attending plays or concerts. Employment in the performing arts falls.

In short: firms probably don’t report direct AI job loss very accurately, but even if they did, it’s not clear how a firm’s adoption of AI – even in a way that leads to AI-driven layoffs – affects jobs on net.

AI can create entirely new jobs and businesses

Even a perfect accounting of AI’s effects on employment at existing firms would miss one of the potentially largest channels of AI’s overall impact on the labor market: the creation of businesses and job categories that don’t yet exist.

New general-purpose technologies have historically spurred entirely new industries. In the late 19th and early 20th centuries, electricity may have changed existing jobs significantly, but it also gave rise to appliance manufacturers, electrical utilities, and eventually consumer electronics, employing workers in roles that had no pre-electricity analog. AI may do the same, creating industries that employ workers in roles current occupational classifications don’t capture.

There’s also some early evidence that AI may be lowering the cost of starting new businesses, particularly for first-time and resource-constrained founders. If this pattern holds more broadly, AI-enabled new firm formation could generate meaningful employment in both existing and emergent sectors.

We should have considerable uncertainty about what new businesses and jobs will emerge due to AI, but any estimate of employment impacts based solely on what existing companies report will be structurally incomplete.

Getting this right matters

It’s worth being cautious when estimating AI-driven labor market impacts. Estimating AI job loss and new job creation is hard, and there are consequences to getting estimates wrong.

Underestimation might lead policymakers to prepare policies that don’t adequately meet the challenges workers face.

Overestimating negative effects could lead to miscalibrated policies that, for example, ban automation, slow growth, reduce demand, and cause labor markets to stagnate. If predictions of AI-driven job loss don’t play out, it could also lead to a negative backlash against policies aimed at supporting workers. If this “crying wolf” effect happens, it could make it harder to garner public support for well-designed policy interventions later when downside risks from automation may be larger.

While the goal of better understanding AI’s labor market impacts is laudable, it is not clear that company reports are the best instrument for achieving it. At best, self-reported data (from, say, firm-level surveys about AI’s impacts) should be taken as one source of information among others in order to understand the full picture.

Thanks to GovAI staff for very helpful feedback on an earlier draft.