Will AI generate mass unemployment?

Abstract

The world is captivated  by the generative AI boom. Its rapid evolution has gone faster thanmost people could have imagined and has caused fears of substantial disruptions in labour markets. We believe that at the current juncture such fears should be qualified and nuanced, certainly with the medium-term impact in mind. The technology still needs to evolve, and its adoption is likely to be gradual. While generative AI is likely to find its way to the workplace and taker over specific (white collar) tasks, it is also likely to create a plethora of new jobs. Furthermore, the current macro-environment makes mass unemployment unlikely. Declining working-age populations will increasingly limit labor supply. Meanwhile, climate-change related investments, increased military spending, continued US-China decoupling and higher healthcare spending will support labour demand.. Generative AI will thus be able to counterbalance the negative effects of these trends on global growth. Nonetheless we should not ignore the sizeable transformation costs and disruptions that AI could bring to labour markets in the shorter term. 

Technology yet to reach its full potential

For one thing, NLP-tools are not close to having reached their full potential and still face major hurdles to overcome. One major problem is their accuracy. The widely reported story of a lawyer that asked GPT to help him prepare a court filing and ended up using a motion full off made-up cases and rulings is a high-profile example. Though these accuracy problems (aka hallucinations) are being tackled, it will take time before the technology is sufficiently accurate and reliable for certain industries such as the legal or medical industry.

Another problem is that AI replicates existing human biases against minorities and women, as the training of the AI modules relies on human-generated data batches². This makes it especially unfit for use in law enforcement and HR-processes, a.o. IBM recently released an open-source library, called AI Fairness 360, that enables AI programmers to test and mitigate bias in models and datasets. That is a great start. Yet more solutions need to be developed to eliminate all biases in generative AI tools. Many firms (though not all firms) will likely hold off before using the new AI tools in their HR departments.

Regulatory hurdles could also slow down AI-related technological advances. The recent EU Artificial Intelligence Act e.g. bans certain AI applications for privacy and ethical reasons. It also creates a regulatory framework for a.o. medical AI applications.

A final known problem with current NLP tools relates to copyright issues³. ChatGPT uses work of artists and writers without their permission. Many marketing departments a.o. will likely want to wait before using AI-generated content in their campaigns.

The tool’s capabilities are also likely to improve gradually. Hence, chatbots will initially only be able to tackle simple customer requests but will eventually be able to take over more complex assignments, thanks to their self-learning abilities.  Similarly, though generative AI programs are currently able to create an AI-generated image, their ability to create AI-generated short-films (let alone full movies) has yet to improve dramatically.

The technological development of AI might also be hindered by the so-called generative AI-collapse problem. If AI is trained on AI-generated content (which will increasingly populate the internet), defects are caused in the algorithm. Solutions still need to be found to remediate this.⁴

This is all to say that NLP tools will likely be usable for different jobs at different time horizons. Their impact on the job market is thus likely to be gradual.

Implementation of technology takes time

Even when the technology is mature enough to disrupt a certain industry, the disruption itself will not happen automatically. On the contrary, the implementation of new AI tools will often require the acquisition of new hardware, a prolonged integration into existing IT systems, the reengineering of existing processes and a long training and change management program. In nimble start-ups or disrupters, the AI-implementation can be quite rapid. However, large firms or entities operating in less competitive industries or in the public sector will likely take years or even decades before fully integrating AI in their processes. On top, the development of firm-specific AI environments can quickly encounter capacity problems as AI system require enormous amount of hardware processing capacity as well as data.

AI will also create many jobs

While companies and public entities around the world are integrating AI in their current processes, the AI-revolution is likely to generate many new jobs. AI is likely to also create a number of jobs that cannot be imagined today. A recent study showed that around 60% of current jobs did not exist in 1940⁵. Many important jobs such as solar panel installer or digital marketing experts were only created recently. Similarly, many new jobs will be needed to a.o. develop, train, test and run AI applications.

Furthermore, by lowering costs for goods and services, generative AI may boost demand for jobs or tasks that are harder to automate. Take management consulting e.g. Here, AI could dramatically lower the time it takes to aggregate data or make slide decks, allowing them to reduce their fees. This will make management consulting affordable for smaller and mid-sized firms and hence enhance demand for their services. A similar dynamic could eventually be at play for lawyers. According to a recent Goldman Sachs report, 44% of legal tasks could be executed by AI, in particular on research and the redaction of legal documents. Automating these tasks will make legal advice cheaper and thus accessible for a wider range of people.

Labor supply is gradually decreasing

Another reason why generative AI is unlikely to cause mass unemployment is the current demographic context. The A.I. revolution is namely happening in a context of declining working-age populations and rising dependency ratios in high and middle-income economies (see figure 1). Rather than adopting AI to replace active workers, firms and government might thus opt to use AI to replace retiring workers. This will ease the strains on the labor market somewhat (see also our research report on demographic change).

Labor demand is increasing

While labour supply is currently constrained, labour demand is generally increasing. Additional sources supporting high demand of labour include the climate change transition, increased military spending, US-China decoupling and increased healthcare demand. These four evolutions amongst others will cause important increases in government (labour-intensive) spending.

On climate change, The Economist estimates that governments will need to spend 0.2% of GDP annually on decarbonization in the coming decades if they wish to reach net zero⁶. Other institutions such as the IEA estimate that the costs might even be higher. The military build-up adds to that. Western-European countries only spent 1.6% of GDP on their militaries in 2021, well below the 2% NATO pledge. Several of them have promised to ramp up spending dramatically. Other countries such as China and Russia are likely to also dramatically increase military spending as geopolitical tensions mount.

Geopolitical tensions are also leading to a costly technological trade war between China and the Western world. To reduce its dependence on China, the US passed the Chips and Science Act, which includes $280 billion in new funding to boost domestic research and manufacturing of semiconductors in the United States. Europe is taking similar steps. A recent IMF paper estimates that the cost of technological decoupling on a 10-year horizon would be around 8% for China, 4% for the US and 6% for the euro area in worst case scenarios⁷. Productivity growth, which has been sluggish in recent decades in the Western world (see figure 2), is likely to be further negatively impacted by this evolution. AI could revert this negative trend (see further).

Finally, ageing populations are not only reducing labor supply, but also increasing labor demand in the (labor-intensive and hard to automate) health care sector. In the US e.g., the Bureau of Labor Statistics projects that jobs in healthcare and social assistance industries will increase at 2.8% annually, adding 2.6 million jobs from 2021 to 2031. It will also put significant pressure on government finances. The Congressional Budget Office projects that spending on Medicare (the public healthcare program for the US elderly) will grow from 10.1% of the federal budget in 2021 to 17.8% in 2032 and could thus raise debt to GDP levels. As AI is likely to raise GDP significantly, it could counterbalance this budgetary pressure.

Impact on the economy

The full impact of generative AI on the economy is still a matter of fierce debate among economists. Some economists, such as Tom Davidson of Open Philanthropy,  estimate there is a decent chance that AI will cause explosive double-digit growth in the coming decades. Given the need for further technological improvement and slow adoption process we discussed above, this seems – in our view - less likely. A recent McKinsey report paints a more realistic picture. It estimates that the deployment of generative AI and other technologies could provide the global economy with an annual productivity boost of 0.2 to 3.3 percentage points from 2023 to 2040, with generative AI contributing 0.1 to 0.6 percentage points of that growth. It estimates that generative AI could add the equivalent of 2.6 trillion USD to 4.4 trillion USD to global GDP over time (global GDP was estimated at 100 trillion USD last year).

Goldman Sachs is slightly more bullish and estimates generative AI could increase global GDP by 7 trillion USD in the next decade. They estimate that AI could boost labor productivity by 0.27 percentage points to 2.93 percentages points, depending on the speed and scope of adoption and the technological evolution. The report also estimates that around 63% of current jobs will be partly automated by AI. That said, less than 5% of jobs would see more than 50% of their workload of automated. Needless to say, both reports stress the extreme uncertainty surrounding their impact assessments. 

Conclusion

Generative AI is certainly going to transform the global economy. That said, we see the risks of mass unemployment on the medium term as rather limited, especially in the current context of lower labour supply and higher labour demand. The further need for technological evolution and the likely slow adoption process of AI will allow the economy to adapt. Rather than causing mass unemployment, AI could provide the global economy a welcome productivity boost! Nonetheless, we should not ignore the sizeable transformation costs and disruptions that AI could bring to labour markets  in the shorter term.

1 “The economic consequences of artificial intelligence : an overview”, C. Piton, 2023, NBB Economic Review

2 Bias in AI: What it is, Types, Examples & 6 Ways to Fix it in 2023 (aimultiple.com)

3 'New York Times' considers legal action against OpenAI as copyright tensions swirl : NPR

4 “The Curse of Recursion: Training on Generated Data Makes Models Forget”, Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, Ross Anderson, 2023, Cornell University

5 “New Frontiers: The Origins and Content of New Work, 1940–2018”, David Autor, Caroline Chin, Anna M. Salomons & Bryan Seegmiller, 2022, NBER

6 Adding up the fiscal drag from ageing, energy and defence (economist.com)

7"Sizing Up the Effects of Technological Decoupling", Diego A. Cerdeiro, Johannes Eugster, Rui C. Mano, Dirk Muir, en Shanaka J. Peiris, 2021, IMF

Terug naar boven