Will Artificial Intelligence Create More Jobs or Destroy Them?

 

Over the last ten years, artificial intelligence (AI) has progressed successively from the stages of machine learning and deep learning to the phase of generative AI, eventually reaching, in recent years, an industrial stage in which its use by businesses and public administrations has become increasingly widespread. Its application in scientific research significantly accelerates the processes of data collection and processing, as well as the generation of results. I have recently carried out several experiments illustrating this advantage of AI. In one case, I used a programme that I had developed some forty years ago, whose source code extended to twelve A4 pages and which required between three and four hours to produce each 50 × 50 matrix of results. The GEMINI platform very quickly converted the programme from the old QUICKBASIC language into modern Python, and within approximately one minute I obtained the first matrix of results. The approach in question, which at the time had been regarded as innovative, rapidly fell into disuse precisely because of the excessive computing time it required.

In businesses, AI makes it possible to increase the efficiency of production and management processes. In public administrations, it contributes to speeding up bureaucratic procedures and improving the quality of services provided to citizens. While AI generates considerable enthusiasm because of its benefits, it also gives rise to various concerns. One of these is the possibility that it may constitute an existential threat to humanity. This concern has prompted a multitude of debates regarding the future evolution of AI and its regulation.

Another concern relates to its consequences for the labour market. On this issue, there already exists a substantial body of academic research, as well as reports from various institutions that provide fairly conclusive findings regarding the employment effects of AI. If the use of artificial intelligence in business management and productive activities reaches a significant scale, it will affect many different professions in different ways. While some professionals will benefit from higher productivity and a beneficial reallocation of tasks, others will see their jobs threatened.

Whenever major and widespread technological changes have occurred in the past, they have generally been followed first by a period of job losses and only later by a phase of partial or complete recovery. This was the case in nineteenth-century England when mechanical looms were introduced and traditional artisans lost their livelihoods. Subsequently, large industries absorbed a substantial proportion of the population that had migrated from rural areas to the cities. A similar process took place in the early twentieth century when Taylorism and Fordism led to the creation of assembly lines. During the initial phase of these industrial transformations, the sharp rise in unemployment resulted in social unrest and labour conflicts.

The computer revolution of the 1980s, which introduced personal computers and large-scale information networks, also led to job losses in occupations such as typists, archivists, telex operators and bookkeeping clerks. A similar phenomenon occurred in banking when computerisation and the implementation of electronic payment infrastructures resulted in the dismissal of many bank employees whose work primarily involved repetitive paper-based operations. This initial phase was followed by another in which numerous new jobs were created for software engineers, systems administrators, hardware maintenance technicians, network analysts and programmers. What remains to be seen is whether the number of newly created jobs exceeds the number that have been destroyed or falls short of it.

The effects of technological change on economic activity and employment have attracted the attention of economists from various schools of economic thought since the nineteenth century. A survey of approximately four decades of literature on the impact of technological change on employment can be found in Technology and Jobs: A Systematic Literature Review, by K. Hotte, M. Somers and A. Theodorakopoulos, published in Technological Forecasting & Social Change in 2023.

The most influential contemporary study on this subject is Automation and New Tasks: How Technology Replaces and Reinstates Labor, by Daron Acemoglu (winner of the 2024 Nobel Prize in Economics and Professor at MIT) and Pascual Restrepo, published in the Journal of Economic Perspectives in 2019. In this study, the authors decompose the consequences of technological change into four effects: the productivity effect, the displacement effect, the reinstatement effect and the scale effect.

The productivity effect (positive) refers to the increase in aggregate output resulting from lower production costs brought about by automation. The displacement effect measures the negative impact on employment arising from the replacement of human labour by machines. The reinstatement effect occurs when innovation creates new and more complex tasks that must be performed by more highly skilled workers, thereby contributing positively to employment. Finally, the scale effect refers to the increase in labour demand resulting from the expansion of an industry or of the economy as a whole, driven by the higher level of capital accumulation made possible by innovation.

If the stock of traditional capital or the total supply of labour increases, the scale of the economy expands. As total output grows, the absolute need to perform all tasks—both automated and human—also increases. The scale effect acts as a buffer against technological unemployment and operates through two main mechanisms. The first is the absorption of displaced labour. When automation frees workers from certain tasks, the scale effect facilitates their reabsorption into the remaining human tasks that machines are still unable to perform. Consequently, if the expanding sector experiences a sufficiently large increase in scale, wages and employment may rise, mitigating the negative impact of the displacement effect.

The second mechanism is that the scale effect depends heavily on the elasticity of demand in the sector undergoing automation. In sectors where demand elasticity is high, the scale effect is particularly strong and tends to offset the displacement effect. Historical examples include the textile industry in the nineteenth century and electronics in the twentieth century. In such cases, automation reduces production costs, leading to substantial price declines that stimulate demand. Under these circumstances, employment increases because the scale of production expands dramatically.

In sectors where demand elasticity is low, such as agriculture, the scale effect is much weaker. Even if automation lowers food prices, people do not consume twice or three times as much food simply because it becomes cheaper. As a result, production expands only marginally, and workers tend to be permanently displaced from the sector.

Acemoglu and Restrepo demonstrated that until approximately 1980 the labour market experienced a balanced cycle in which the destruction of tasks was broadly compensated by the creation of new ones. Since then, however, a profound structural shift has taken place. The long-run equilibrium share of labour income depends on the relative speed of two technological forces: the rate of automation and the rate at which new tasks are created.

During the late nineteenth and early twentieth centuries, major technologies such as electricity and the internal combustion engine developed in parallel with industrial expansion. The employment impact of a technology also depends on the magnitude of its productivity effect. If a technology is sufficiently revolutionary, the increase in wealth it generates is large enough to compensate for the jobs it destroys.

Around one hundred years ago, when tractors automated agricultural production and agricultural workers lost their jobs, the complexity of the emerging industrial economy simultaneously created engineering, accounting, logistics and public service occupations in roughly equivalent proportions, allowing labour’s share of GDP to remain relatively stable.

From 1980 onwards, however, the revolutions in software, industrial robotics and artificial intelligence have dramatically accelerated automation, while the creation of genuinely new tasks has gradually declined. The result has been a net displacement effect that reduces aggregate demand for labour.

According to Acemoglu, much of the automation introduced since the 1980s can be characterised as “so-so automation”. The situation differs markedly from that of the early twentieth century, when replacing horses or human physical labour with electric motors generated dramatic productivity gains and sharply reduced production costs, thereby expanding the scale of the economy.

In contrast, many contemporary automation technologies—such as supermarket self-checkout systems, automated telephone services and basic administrative software—offer only marginal efficiency improvements relative to human workers. Since the resulting cost savings are limited, the productivity effect is correspondingly small. Without a substantial increase in aggregate wealth to stimulate economic expansion, the negative displacement effect tends to dominate.

Using their task-based model, Acemoglu and Restrepo showed that automation since 1980 has contributed relatively little to improvements in labour productivity. Instead, it has altered the content of tasks in ways that disadvantage workers. Examining US data from 1947 to 2016, Acemoglu found that prior to 1987 the displacement and reinstatement effects largely offset one another and wage growth closely tracked productivity growth. Between 1987 and 2016, however, the displacement effect intensified significantly as industries specialised increasingly in routine tasks.

According to his estimates, between 50% and 70% of all changes in wage structures and income inequality in the United States since 1980 can be attributed directly to the displacement effect generated by automation. More recent studies largely confirm that the conclusions reached by Acemoglu and Restrepo remain highly relevant today.

In order to weigh the benefits and social costs of artificial intelligence, it is first necessary to estimate the scale of AI investment over the coming years. Secondly, it is necessary to identify the professions that are likely to benefit from this innovation and those that are likely to be adversely affected.

Since becoming counter-cyclical after 2020, as highlighted in the report No US-Style AI Investment Boom to Drive EU Growth, published by Oxford Economics in January 2026, investment in AI has acquired particular importance in the current context of prolonged economic stagnation affecting both Europe and the United States.

Since 2020, annual investment in AI has amounted to approximately 1.9% of GDP in the European Union, with a somewhat higher figure in the United States. According to the 2026 Federal Reserve Bank of St. Louis publication Tracking AI’s Contribution to GDP Growth, investment in AI has outperformed almost every other category of technological and intellectual property investment.

In the United States, most AI investment has been directed towards physical infrastructure, particularly data centres and advanced semiconductors, which account for between 60% and 70% of total expenditure. The remainder has been devoted to the development, training and fine-tuning of large language models (LLMs).

In Europe, by contrast, AI investment has focused primarily on human capital formation. This policy choice reflects the large number of workers requiring reskilling for the digital transition. In Portugal alone, it is estimated that approximately 300,000 professionals per year will require retraining until 2030.

The financial services company Morgan Stanley forecasts that AI investment will grow at an annual rate of between 7% and 8% over the coming years, with growth expected to be somewhat faster in the United States than in Europe. This difference is largely explained by more favourable financing conditions in the United States, particularly the substantial financial resources of Corporate Venture Capital firms. Between 2020 and 2024, these institutions allocated approximately 34% of their investment resources to AI-related companies. Their European counterparts possess considerably lower financial capacity and allocated only around 18% of their resources to AI enterprises. As a result, the AI sector in the European Union is estimated to suffer from an annual funding shortfall of approximately €60 billion, which can only be compensated through European Union and national government programmes.

By directing a substantial share of AI investment towards training and workforce reskilling, in accordance with the principles established by the EU Artificial Intelligence Act, the European Union seeks to reduce employment losses resulting from technological innovation. However, this strategy also increases Europe’s dependence on imported American technology to sustain its productivity performance.

According to estimates produced by the International Monetary Fund (IMF) and the International Labour Organization (ILO), approximately 60% of the workforce in advanced economies, including the European Union and the United States, will be directly affected by artificial intelligence. The effects are expected to be distributed almost equally between positive and negative outcomes. The remaining 40% of workers are considered to occupy positions that are largely neutral with respect to AI.

Highly skilled professions that are likely to benefit substantially from AI account for approximately 30% to 35% of total employment, representing around 50 million workers in the United States and 60 million in the European Union. This group includes professionals with higher education qualifications, such as physicians, managers, senior engineers and university lecturers, who will be able to use AI as a high-performance assistant for analytical and administrative tasks, allowing them to focus on diagnosis, strategic planning and decision-making.

Among the professions most vulnerable to disruption are those whose core activities consist of routine administrative work, standardised text processing or basic programming. This category includes administrative staff, call-centre assistants, routine accountants and junior software developers. According to the ILO, these occupations disproportionately affect women in advanced economies, owing to their historical concentration in clerical and secretarial roles. This group represents approximately 25% to 30% of all occupations and includes around 45 million workers in the United States and 55 million in the European Union.

The ILO estimates that between one quarter and one third of existing jobs are exposed to significant transformation as a result of current AI investment. Similarly, the IMF argues that in advanced economies approximately 30% of jobs are likely to benefit from AI, while another 30% may be negatively affected. In developing economies, the picture differs somewhat: around 40% of jobs are expected to benefit, while only 26% are likely to experience adverse effects. This suggests that the negative employment consequences of AI may be less pronounced in emerging economies.

Finally, there is a group of occupations that remain largely unaffected by AI because they rely heavily on physical activity, manual dexterity in dynamic environments or direct human empathy—characteristics that neither algorithms nor current robotic technologies are capable of replicating effectively. This group accounts for approximately 35% to 40% of all occupations and includes around 65 million workers in the United States and nearly 80 million in the European Union.

Examples include plumbers, electricians, heavy goods vehicle drivers, carers for the elderly, nurses and healthcare assistants, firefighters and police officers. These occupations are expected to remain relatively immune to technological disruption in both the short and medium term because of the complexity of their interactions with people and physical environments.

To appreciate the scale of future investment in artificial intelligence, it is sufficient to note that global AI expenditure is projected to exceed US$500 billion annually. According to Goldman Sachs, the United States is expected to maintain a comfortable lead over the European Union owing to the scale of its private capital markets and technological infrastructure.

With regard to employment, projections presented in the European Commission report The Future Employment Impact of Artificial Intelligence and Emerging Digital Technologies in Europe indicate that AI will ultimately have a positive net effect on employment, potentially increasing total employment by up to 1.4% in the long run. In the short term, however, significant job losses are expected among junior and administrative positions.

In Portugal, Amazon Web Services examined AI investment trends in the report Unlocking Portugal’s AI Ambitions in the Digital Decade. According to this report, 35% of Portuguese companies had already adopted AI technologies by 2023, while AI investment grew by 25% compared with 2020 levels. If this adoption rate is maintained, the total economic impact of AI in Portugal could reach €61 billion by 2030.

The less encouraging aspect is that 71% of Portuguese firms report difficulties in recruiting employees with strong digital skills, compared with a European average of only 44%. Given the relatively low level of digital preparedness among the Portuguese population, the difficulties many companies face in recruiting appropriately qualified workers, and the substantial employment risks facing numerous occupations, the balance between jobs created in expanding AI-related professions and jobs lost in declining occupations is likely to remain negative for some time.

The process of investing in innovation is complex and heavily dependent on access to finance. It also faces other obstacles, including the lack of preparedness among many employers, which may discourage participation in technological innovation. Consequently, there is no guarantee that the pace of AI innovation will continue indefinitely at the rate observed during the initial surge of recent years, which has been concentrated primarily among large corporations.

Given the positive effects that AI investment can have on the economy through its counter-cyclical properties and its capacity to increase productivity, it is desirable that such investment be encouraged through appropriate public policies. These include the integration of AI into public administration, thereby allowing the state itself to become a driver of technological innovation, and the development of large-scale retraining programmes aimed at workers whose occupations are likely to decline as a consequence of technological change.