The impact of artificial intelligence on unemployment: a review

Abstract

Purpose

The aim of this paper is to summarise the state-of-the-art debate on impact of artificial intelligence on unemployment and reporting up-to-date academic findings.

Design/methodology/approach

The paper is designed as a review of the labour vs capital conundrum, the differences between industrial automation and artificial intelligence, threat to employment, the difficulty of substituting, role of soft skills and whether technology leads to the deskilling of human workers or favors increasing human capabilities.

Findings

Some authors praise the bright future developments of artificial intelligence while others warn about mass unemployment. Therefore, it is paramount to present an up-to-date overview of the problem, compare and contrast its features with what happened in past innovation waves and contribute to academic discussion about the pros/cons of current trends.

Originality/value

The main value of this paper is presenting a balanced view of 100+ different studies, the vast majority from the last five years. Reading this paper will allow to quickly grasp the main issues around the thorny topic of artificial intelligence and unemployment.

Peer review

The peer review history for this article is available at: https://publons.com/publon/10.1108/IJSE-05-2023-0338

Citation

Virgilio, G.P.M., Saavedra Hoyos, F. and Bao Ratzemberg, C.B. (2024), "The impact of artificial intelligence on unemployment: a review", International Journal of Social Economics, Vol. 51 No. 12, pp. 1680-1695. https://doi.org/10.1108/IJSE-05-2023-0338

Introduction

The debate about the impact of artificial intelligence (AI) on our society is fierce and getting more so.

The scientific community divides itself between optimists and pessimists. The optimists foresee a world where machines work and humans reap the benefits, others forecast a future of unemployment for those who do not possess the means of production, robots and systems, but knowledge as well.

Theoretical framework

This paper follows in the footsteps of two important reviews showing good methods of presenting the literature (Ozili et al., 2023; Martens and Tolan, 2018) and, in a similar way, it focuses on six main topics: (a) a comparison of optimistic and pessimistic views; (b) an analysis of the impact AI has on low- and middle-skill jobs and (c) on high-skill jobs; whether AI (d) increments or reduces the skills required of human workers (with all the consequences this has on motivation and effort). The final topic is a comparison of the current industrial revolution with those of the past.

The main purposes of this study are: (1) to provide a review of up-to-date research; (2) to assist researchers in understanding the state-of-the-art on the AI/unemployment issue; (3) to add to the literature with an analysis of the impact AI may have in the next future and (4) to contribute to policy decisions.

Methodology

Limitations of this review

Many papers used throughout this research display limitations, such as a narrow geographic focus or being too sector-specific, some present specific cases still at the pioneering level, others fail to discuss a wider reality. In order to overcome such difficulties, this study revised 106 sources, balancing some of these limitations by averaging out over a relatively large number of results, thus offering an ample horizon of the different opinions among researchers. The AI-related sectors this study refers to include the following: healthcare, automation, organizational processes, HR management, R&D, engineering, decision-making or management, services, clerical work, automotive, tourism, customer care, speech recognition, translation, manufacturing, ICT, software development, the legal sector, academic libraries, research, the public sector, consulting and repetitive jobs, among others. Moreover, the cited papers have been filtered from a larger base of over 300 items, with the purpose of presenting a review impartial and as representative as possible. About half of the references have been published in Scopus-indexed journals, 12 in books and 23 in reports/WP of prestigious institutions (NBER, OECD, MGI, Cornell University and Stanford University).

Criteria adopted in this study

The literature offers a large number of papers on the impact of AI to employment/unemployment. The selection criteria have been the following:

1. Relevance. A literature review should provide a broad view of the most debated topics. Sources have been selected based on their importance and impact on the current argumentation.

2. Balance. The literature presents both optimistic and pessimistic views of the future AI-led labour world, as there is no definitive conclusion about whether AI will create a jobless world or a business environment where more and more highly-paid, jobs will appear, as was the case of prior industrial revolutions. The goal of a review is to offer a balanced perspective on both sides.

3. Span. Although AI has a potential impact on several human activities, and in a paper, it is impossible to thoroughly present all of them, the selection focused on a significant sample of activities able to represent entire job categories.

4. Timeframe. Being AI a high-technology field and one in which innovations occur at a rapid pace, the focus had to be on the state-of-the-art, and the main choice was over the last few years. However, the paper also considers a broader temporal scope, including historical context and trends.

5. Quality of sources. Most sources have been published in journals indexed in Scopus or Web of Science, ensuring the high quality of the references.

The optimistic view

The AI-optimists highlight the huge scope for collaboration between humans and machines. Ardon and Schmidt (2020) find that 64% of 1,721 respondents to a questionnaire distributed among clinical lab employees have a positive attitude towards AI, with 24% highlighting “reducing errors” and 16% “saving time” as main benefits, whereas only 27% showed concern about having their job replaced by AI. By surveying 487 pathologists in 54 countries, Sarwar et al. (2019) report almost 75% of respondents displaying “excitement” in AI as a diagnostic tool. Raj and Seamans (2019) also sustain that AI may be a tool at the disposal of humans rather than a competitor for their job, a stance recognised also by Cao et al. (2021) that weighs pros/cons of AI adoption among 269 managers. Yang (2022) is positive about AI increasing employment and Hamaguchi and Kondo (2018) share the same opinion, yet recognising that only workers with appropriate skills will enjoy the upcoming opportunities. Even though in the short-term, jobs may be temporarily lost, they will soon be recreated with interests (Ma et al., 2022). This statement matches the historical view that every industrial revolution has initially spread anguish for the jobs lost, but over the entire process period, it created more jobs and better paid than the lost ones (Kelso and Adler, 1958). This is a crucial point: according to Hancock et al. (2020), by 2030, more than one-third of all employees will need significant retraining to keep their job. Analysing EU data over 1999–2010, Gregory et al. (2018) find that the spillover effect outweighs job reduction and that the net balance is positive. Research on the German market allowed Dauth et al. (2018) to reach similar conclusions: even though jobs may be lost in manufacturing, even more jobs are created in services. As stated by Frey and Osborne (2017), only low-level jobs are at risk; high-skill tasks benefit in both headcount and salary. Also, Kaplan and Haenlein (2020), using a PESTEL framework, notice how, despite the huge technological innovations in place over decades, the market is increasingly hungry for skilled labour. In a similar fashion, Abdullah and Fakiek (2020) carried out 250 interviews among healthcare personnel in Saudi Arabia, and most of them boasted confidence in the irreplaceability of their job by AI. This confirms an earlier result by Oh et al. (2019) observing positive attitudes of medical personnel towards technology. Bhargava et al. (2021) find that professional level and acceptance of AI innovations are directly correlated. Similarly, Jaiswal et al. (2022) trust that humans will be relieved from executing routine tasks and free to undertake more creative activities. Plastino and Purdy (2018) indicate the beneficial role of AI in supporting humans and enhancing their capabilities rather than substituting them.

The pessimistic view

AI is also regarded as a potential threat to human employment. Cognitive tasks, once thought to be immune to automation, now face serious challenges. At the turn of the century, Mincer and Danninger (2000) observed that technology appeared to reduce unemployment in the long run, at least in the years before they wrote. A few years later, Beaudry et al. (2013) documented an increase in unemployment since the year 2000. This time, things actually look different. Schwab (2016) discusses how the three biggest Silicon Valley companies had roughly the same 2014 revenue as the 1990 revenue of Detroit’s top three manufacturers (about $250 bn), but with one-tenth as many employees. In 2015, WhatsApp had just 50 employees to service about one billion users (Metz, 2015). Ford (2015:xi) observes how AI innovations involve a widespread reduction in wages. “As of 2013, a typical production or nonsupervisory worker earned about 13% (inflation-adjusted) less than in 1973, even though productivity rose by 107% and the bill for major expenses, like housing, education, and health care have soared” Vieira and Giannoccaro Von Huelsen (2020), noticing that 80% of jobs lost are linked to technological innovation, doubt that employment recovery could make up for the losses. According to Brynjolfsson and McAfee (2011), jobs created by AI would not suit those workers affected by technological unemployment, as they may not possess the required skills for being re-employed in the short term. Frey and Osborne (2017) state that AI may put at risk about 47% of jobs in the USA and Nedelkoska and Quintini (2018) raise it to around 50% unless significant retraining programmes are carried out. Indeed, several forecasts seem to converge towards technology adoption, resulting in serious unemployment. Mahroof (2019) notices how AI continually breaks new frontiers and the decision-making potentials of AI are increasingly recognised (Duan et al., 2019). As a consequence, managerial jobs are now at risk since, differently from previous industrial revolutions, AI may also impact high-skill jobs (Webb, 2020). According to Georgieff and Hyee (2021), new technologies mainly target low- and middle-skill tasks, whereas Jaimovich and Siu (2012) find that only medium-level jobs are at risk. On the contrary, Ma et al. (2022) interpret panel data from 30 Chinese cities over the period 2003–2017 as a sign of a weakened effect of AI on middle-skill jobs and Yang (2022) concludes that high-skill employees experience a rise in demand to the detriment of middle-skill ones, with a neutral effect on low-skill workers. There seem to be quite different opinions among scholars.

Whatever the skill class most affected by AI, worries are not unfounded. Jiang and Zou (2018) doubt that newly created employment may offset jobs lost to technology. According to Acemoglu and Restrepo (2018a), new jobs would replace labour-intensive tasks, whereas Susskind and Susskind (2016) forecast a further weakening of demand for technical employment. The real question may no longer be whether a computer is able to pass the Turing test (Turing, 1950), but whether humans can be as effective as computers are. The challenge for humans is to differentiate themselves to protect their jobs from AI-powered robots. A question raised by many authors (Ford, 2015; Davenport and Kirby, 2016) is whether it will still be possible to find a job given the growing impact of AI not only on the manufacturing industry but also in highly cognitive industries.

AI threats to low-/middle-skill jobs

Capital-for-labour substitution

The first Industrial Revolution exploited automation to replace an entity in finite supply, labour, by another, capital, “which is in unbounded supply” (Aghion et al., 2019:150). Not surprisingly, the economic growth that followed dismissed workers no longer required or unable to adapt to the new technology, the mechanical loom. A similar scenario was repeated in the 1970s, when industrial automation was the new technology that created unemployment in manufacturing (Dauth et al., 2017). A common feature of all prior industrial revolutions is the increment of high-skilled labour and, to some extent, also of low-skilled labour while shrinking demand for the middle-level. Dong et al. (2020) research over 97 firms, finding that technology can widen the gap between differently skilled jobs and create a “polarization of employment”. This can be explained by the augmented need for workers able to manage new technologies and for low-skilled workers whose salaries are lower than the investment in automation required to replace them. Categories in the middle are at risk of heavy unemployment. As pointed out by Ballestar et al. (2020), this is strictly related to productivity increases. Innovation fatally lays off workers unable to ride the new wave with the promise to hire many more who are able to adapt. Nevertheless, the World Trade Organization (WTO, 2017) is doubtful about the belief that growth following an employment slowdown would reverse the negative trend. Increasingly, since the late XX century, capital has started to take on the lion’s share of production (Karabarbounis and Neiman, 2014), mainly because of its technological content. Dosi and Mohnen (2019) and Staccioli and Virgillito (2021) display similar opinions by finding that technology-led industrial innovative processes raise unemployment levels. Martens and Tolan (2018:4) state how “recent evidence points to a declining labour share in total income”.

Automation vs AI

Despite some similarities, the AI revolution presents strong differences from industrial automation. The former tends to displace middle-level jobs while boosting higher level and sparing lower-level employment. As pointed out by Georgieff and Hyee (2021), technological workers do not necessarily need to possess knowledge of the production process that can be regarded as a black box; the only indispensable knowledge is the operation of technology without digging into it. This leads to hiring a completely different type of employee, making most “old” workers unsuitable and therefore redundant. Thus, the second step of automation was the computerisation of all activities that could be standardised – and therefore suitable to be converted into a sequence of programming steps. Huang et al. (2019) establish a theory according to which initial AI replacement occurs at the task level and subsequently transfers to the job level, a result confirmed by field observations (Acemoglu and Restrepo, 2018b). Most researchers find that wages are being reduced in the process, even when productivity raises (MGI, 2017). However, the scenario is not clear-cut. According to Arntz et al. (2016), only a small percentage of US jobs are at risk of being replaced by AI, whereas Felten et al. (2019) do not find any impact of AI on employment and Bughin (2020) even finds an increase in employment driven by stepping up investments. Damioli et al. (2021) also argue that an increase in productivity-per-capita rises employment. Obviously, in this scenario, increasing employment concerns the educated/experienced quantile. The big advantage of AI versus automation is that the former just needs free-of-charge replication of algorithms rather than the costly purchase of new automation equipment. No wonder its adoption is welcomed by firms. Nevertheless, Shaukat et al. (2020:51) state that at present time “[a] production robot might thus cost cheaper than a worker in China”. The negative impact on employment is a logical consequence, as explained by de Vries et al. (2020).

Substitution of technology for human jobs

Technology, in the form of robots, computers or algorithms, has already carried out a heavy substitution of human labour, and it is expected to follow a similar trend over the next few decades. Acemoglu and Restrepo (2017) find that every robot in US manufacturing leads to losing 6.2 jobs and declare 0.18–0.34% more unemployed per every robot, whereas, taking into account change in employment in six European countries, Chiacchio et al. (2018) report a decrease of 0.16–0.20%. Aghion et al. (2019) observe 0.37% in the USA, nearly twice as much. At the same time, robotization resulted in a 0.5–0.73% increment in wages and 0.36% productivity growth (Graetz and Michaels, 2018). On their part, Jiang and Zou (2018) notice that technological substitution does actually create new jobs, although the number is insufficient to balance the lost ones. Gu et al. (2022) report that any change in employment following technological innovation tends to penalise the demand for medium skilled workers in favour of high-skilled employment. In this race to competence, firms gain the highest income share improvement, then come the highly skilled workers, to the detriment of lower-skilled workers, who experience wage penalization if they are so lucky not to lose their job altogether (Jaimovich and Siu, 2012). Also gloomy are the predictions of Korinek and Stiglitz (2017) and Autor and Salomons (2018), whose studies result in a net negative effect of technology on all types of employment. Ma et al. (2022) argue that in the medium term, even a highly skilled workforce may experience a decrease. Actually, it is sensible to accept the existence of a “physical” limit to the skills that can be acquired in a human lifetime. On the enterprise side, Koo et al. (2021) recognise the huge advantages of adopting technological innovation. Against a high initial investment, enterprises enjoy lots of benefits, such as higher execution speed and higher efficiency. Lewicki et al. (2019) argue that one robot may execute the tasks of up to seventy full-time human workers. Moreover, firms would also likely consider other benefits such as 24/7 shifts, lower error rates, no work safety concerns, no absenteeism, annual, sick or maternity leaves, no salary rise, no HR management or health insurance costs and no strikes, among others. Aum et al. (2018) demote the productivity myth since manufacturing, historically enjoying high productivity growth, is gradually declining in importance to services. Other sullen forecasts are higher productivity with less human employment, a viewpoint authoritatively supported, especially in industrialised countries (BLS, 2015). Glenn and Florescu (2015) are definitive in predicting mass unemployment within a few decades.

AI threats to high-skill jobs

Perceived irreplaceability

The danger of technological replacement is not equally felt by all categories of workers. This viewpoint is shared by Bhargava et al. (2021), according to which high-level jobs enjoy a certain security face to threat of technology, although the study recognises that managerial and executive positions may be affected to some extent. Kospanos (2018) identifies in human decision-making capabilities and in “human touch” the unique feature that makes us irreplaceable by AI. Lloyd and Payne (2019) do not notice job losses due to technological advancement.

Difficulty to substitute

Although perhaps not irreplaceable altogether, many jobs are widely considered difficult to replace. Contrary to the findings of Alsharqi et al. (2018), Abdullah and Fakiek (2020) report that 65% of respondents are sceptical about AI being able to critically handle unexpected events. In the view of Cao et al. (2021), advanced technology should be a supporting instrument for human decision-making activities and certainly not a replacement thereof. Ghosh and Kandasamy (2020) explain the reason: since humans do not understand the “mental process” of AI in making clinical decisions, relying on the unknown is not rational and non-ethical. Indeed, in the case of inappropriate adoption of unsuccessful therapy or in the case of a plain mistake, who will be liable for it? Since machines do not have a conscience, the ultimate responsible person must be the human operator/supervisor. But if he/she cannot understand all the details, would it be just to assign him/her the responsibility? On the other side, according to Black and van Esch (2020), algorithms only replicate what humans, either directly or indirectly, instruct them to do. If teaching is biased, can we blame the teacher for making the wrong decision? Huang et al. (2019) differentiate between capabilities that are successfully aggregated into algorithms on a daily basis and those that are more difficult to emulate, involving such soft skills as reasoning, empathy-building, relationship-building and communication. This leads to recognising different categories of skills, not necessarily based on a ranking of difficulty. Indeed, theorists developed the concept of general reinforcement learning (Martin, 2019a, b). The AlphaZero algorithm, by simply being given the game rules, achieved stellar performance in chess, shogi and Go by starting from scratch and learning by playing against itself, completely avoiding any human bias. This is the most promising way ahead for solving the problem in other fields of AI, demolishing another barrier against substitution. Leyer and Schneider (2021) identify a major bias in risk-avoidance, a widely appreciated feature of business behaviour that often misses to make optimal, yet risky, decisions. Businesspeople should perhaps learn from chess engines.

Irreplaceability of soft skills

Frank et al. (2019) follow the opinion of many other scholars in identifying repeatable tasks as the easiest ones to replace, whereas those requiring emotional intelligence and communication skills will survive. In fact, R&D employees working in highly creative departments do not feel pressure from AI. Soft skills are the hardest to replace: human “touch”, decision-making capabilities (, interpersonal relationship-building, emotional intelligence, instinct (Lichtenthaler, 2018), meaning of speech, cultural subtleties and persuasion (Guenole and Feinzig, 2018). AI is not capable of understanding; all concepts required for simulating a conscience need to be hard- or soft-wired into chips or algorithms. It is true that some hotels and restaurants have tried to provide their customers with a robotised-caring experience, but the attempts are just at the beginning. Long before AI became a possibility in customer interaction, Choi et al. (2004) highlighted the impact of customer service as the difference between expectations and perceptions of service quality. Moving ahead in time, Klie (2013) identifies AI-based customer interaction as a powerful tool for delivering a pleasant customer experiences whereas Muro and Andes (2015) quantify the cost of human customer care service at more than $10 per interaction, while AI response would only cost slightly more than two dollars. Robots proved to be good at analytical and cognitive tasks, yet less so when emotional intelligence is required (Zhou, 2017). The big bet is whether technology will continue to develop and pile up successes on its way or whether there is a ceiling made by intrinsic limits to technology. Yet, AI is progressing fast in several arenas that, until recently, were considered “for-humans-only”. Stock (2018) is confident that robots meet the requirements necessary for providing customer service with the appropriate care and “understanding” of counterparts. Arora et al. (2023) discuss customer experiences so successful as to lead humans to feel job insecurity (Li et al., 2019). On the other side, Koo et al. (2021) believe that AI is not able to imitate soft skills typical of human employees at the same level. Different from technicians, who are more easily replaceable, doctors and nurses directly interact with patients and therefore, feel to be in a safe position facing the advancement of AI in medical care (Krittanawong, 2018). However, scholars report several activities where AI medical applications are far superior to those of their blood-and-flesh colleagues. According to Esteva et al. (2017), diagnostic algorithms surpass professional dermatologists in spotting out skin cancer and Arsene (2019) states that AI surpasses humans in pinpointing precancerous signs (91 versus 69%). Gastroenterological patients benefited from deep learning, which proved successful in suggesting the most appropriate therapies (Ruffle et al., 2018).

Threats of substitution

At the beginning of the century, Autor et al. (2003) ruled out technological substitution for activities as varied as medical care, driving, writing and selling. Although ChatGPT casts some doubts about writing and selling, which do not seem, at the moment, within the capabilities of AI, medicine is definitely within its scope, as is self-driving. Two out of four predictions of non-replaceability must be dismissed after just 20 years and wise people would not bet on writing as a human-only activity for long. A few years later, Brynjolfsson and McAfee (2011) clarified that automation was no longer restricted to mechanical or repetitive tasks and after another handful of years, Frey and Osborne (2017) estimated a gloomy future for more than 700 human jobs. Whether or not the number is overestimated, the path seems crystal clear. Aghion et al. (2019:158) suggest to “relax the assumption according to which automation could not threaten nonroutine jobs”, Mahroof (2019) argues that some processes earlier considered unsuitable for AI now fall within its scope, whereas other authors warn against underestimating the potential AI has in decision-making processes. Not surprisingly, Ransbotham et al. (2018) detect fear among managers because of the threat to their jobs coming from AI. HR recruitment is an example of an activity requiring soft skills and therefore allegedly being an exclusive domain for humans. Nevertheless, AI is providing threats to its blood-and-flesh counterparts in that arena as well. Upadhyay and Khandelwal (2018) highlight that AI has already proven able to screen resumes more impartially than human recruiters, Kuncel et al. (2014) show how AI is 25% better at screening applicants and Chen (2022) recognises the higher level of accuracy achievable through deep learning models. According to Ahmed (2018), AI is able to analyse emotions, body-language, tone of voice and speech nuances, resulting in better decisions about candidate selection (Friedman and McCarthy, 2020). In recent years, AI has proved able to significantly advance in many areas, such as “image and speech recognition, natural language processing, translation, reading comprehension, computer programming and predictive analytics” (Georgieff and Hyee, 2021:8). These capabilities allowed algorithms to expand activities they can accomplish and business tasks they can replace; creativity, social intelligence and abstract reasoning are among such activities, considered outside the scope of AI until recently (Acemoglu and Restrepo, 2020). It is pretty clear that the number of jobs outside the reach of AI gets smaller every day. As found by many researchers, blue-collar workers are not the only ones at risk, even white-collar employees, managers and executives are not guaranteed to keep their job for long. Advancement in language translation makes AI substitution a matter of “when” rather than “if”. Neural Machine Translation shows an error reduction compared to previous non-AI-based translators of nearly 60% to Chinese and nearly 90% to Spanish (Wu et al., 2016). Lack of text comprehension has been addressed by Gröndahl et al. (2018), who, analysing texts on social networks, successfully demonstrated AI being capable of tracking down and banning hate speech. Not only do investments in AI have huge potential but also they are deemed to bring substantial savings (Miyashita and Brady, 2019). How can profit-oriented companies resist the lure of the bottom line?

AI deskilling human workers

AI not only surpasses humans in many tasks they compete in but also it has a fiendish effect of leading to deskilling of human employees. In workplaces where heavy AI is implemented, humans see their cognitive workload diminish. Employees learn to trust AI and to delegate some tasks of theirs; therefore, their dexterity in such tasks gets reduced. They tend to lose capabilities and make their functions more and more dependent on silicon solutions (Chen, 2022). Technological advances occur at a faster rate than humans can get acquainted with them, let alone learn about them in depth. Keynes (1931) labelled the rapid advancement of innovations as technological unemployment. He was probably the first to worry that even cognitive workers were at risk of headcount cuts. Yet, in the 1930s, innovation was not running as fast as it is nowadays and, for many decades, there was the chance to learn about new emerging jobs, provided enough time and commitment were put in place. However, a faster and faster pace of innovation arrival may jeopardise even the best will because of too long learning time with respect to business needs. Agrawal et al. (2018) suggest deskilling is a real possibility, with joblessness as the natural consequence. Faraj et al. (2018) foresee de-skilling as a result of technological innovation. The more the complexity of algorithms grows, the less cognitive workers will be able to grasp their behaviour and keep their jobs. Leyer and Schneider (2021) find that 42% of all managers interviewed decided to switch to AI systems after they made a wrong decision. This can be viewed as surrendering to the superiority of the machine; it may be difficult to go back afterwards. Delegating knowledge and reasoning to the chips and essentially losing human domain over such capabilities may be forever.

AI increasing skill of human workers

A common misconception about prior industrial revolutions is the net positive effect on headcount. This takes the average into account and as all averages are true in general, it is false in each particular case. It is certainly true that, over some time, the total number of employees increased with respect to the beginning of the period. Alas, all those workers that were unable to adapt to the new technological environment lost their jobs forever, and it was no help thinking that one-and-a-half young engineers replaced each blue-collar worker dismissed by automation. The fittest survive and the others die; the youngsters advance in their careers and the oldies retire with the minimum package. As Huang et al. (2019) recommend, keeping up with technology and acquiring new skills is crucial to hope preserving own employment. Many scholars categorise employees’ attitudes towards corporate changes; among them, Lund et al. (2020) identify innovators, early adopters, the majority and laggards. Gu et al. (2022) find that if attitudes are positive and people commit themselves to learning about AI, total employment may even increase. Hamaguchi and Kondo (2018) notice a correlation between years of education and the probability of safeguarding one’s own job and Ma et al. (2022:258) conclude that training workforce is the first way “to reasonably deal with the impact of technological innovation such as AI on employment skill structure”. Hancock et al. (2020) are even more clear-cut when they recognise that between 30 and 40% of the total workforce would need to undergo upgrading or substantial retraining over the next decade to keep competing with AI. Humans need to learn how to collaborate with algorithms, although, according to Daugherty and Wilson (2018), workers cooperating with AI should be warned that they are actually teaching algorithms to eventually replace them: learning new tasks is precisely what machine learning (ML) is all about. Raisch and Krakowski (2021) point out that AI-based task augmentation leaves decision-making power in humans’ hands, contrary to automation that yields it to machines. The former allows for employment but requires the capability to interact with AI, a goal that costs time and effort for those willing to undertake the task. But those who welcome the challenge are compensated with higher wages and career opportunities (Agrawal et al., 2018). In the largely sensible opinion of Verma and Singh (2022:3), “[i]f AI-enabled jobs require high-tech professionals to perform different tasks, professionals might feel more intellectually challenged”, while Kaplan and Haenlein (2020) stress the demand and psychological reward given by emotional, cognitive and intellectual requirements to top-level employees. For them, continuous learning is a must, but the key question is: What about the rest? Here we are talking about the 375 m people that, according to McKinsey (MGI, 2017) will be required to develop AI-related skills or risk losing their job.

Is this time different?

There is no one type of AI; there are many. Weak-AI focuses on one specific task, artificial general intelligence is indistinguishable from human intelligence (it passes the famous Turing test) and finally, artificial super intelligence displays more intelligent behaviour than humans. For the first time in its history, mankind faces an entity superior in the field so far considered its own exclusive domain. This is no little novelty; no longer have we to compete for a job with a mechanic loom, an electronic system, a dumb computer or a smart phone. In the current competition, humans lose at chess, at surgical operations, at business decisions and at painting, sculpturing or producing music. This time everything seems to be different. Very few are the activities not under threat from AI, whereas all previous industrial revolutions focused on one or a few, specific fields: the textile sector, transport, computing or communications. For the first time, technology is being charged with moral questions: who is accountable for decisions made by a robot? This is not akin to prior technological innovation, like the nuclear bomb; in such a case, the responsibility was deeply enrooted in human judgement. Not so when, on May 7, 2016, a self-driving car killed a person. Had the car had a driver, the answer would be easy. But blaming a person through an algorithm is incommensurably harder. As difficult as it is to decide who is on the driving seat when AI collaborates with humans. Leyer and Schneider (2021:711) ask whether “managers will benefit from enhancing their abilities with AI-enabled software or become powerless puppets”, a question never asked with mechanic looms. Despite noticing that suggestions from the past go towards reassuring evidence, Martens and Tolan (2018:4–5) conclude that “the nature of AI/ML is different from previous technological change”, although “there is no empirical evidence yet to underpin this view” and that “the ‘this time is different’-syndrome keeps stirring concerns”. For the first time, the technology-led displacement effect may overwhelm productivity and Aghion et al. (2019) notice that AI algorithms and not only human researchers produce innovative ideas.

Euphoria about real or expected bursts of production already occurred in the decade ending in the 1929 Great Crash or the 2001 dot.com crisis. Will this time be different? Reis et al. (2021) point out how technology adoption displays shorter and shorter cycles. Vieira and Giannoccaro Von Huelsen (2020) observe that in order to reach 50 m users/subscribers, it took analog phones 75 years, no more than 38 were needed by radio, just 14 by TV, a mere four by the Net and a flash of 50 days by mobile technology. Innovations no longer are allowed time to gain acceptance as new offers appear at short intervals. Learning at the required pace is beyond human capability.

Conclusion

Implications

Many authors are rather pessimistic about the future of human employment due to the AI revolution. Policymakers are being advised to act very carefully in this respect, and appeals in this sense are mushrooming. Nevertheless, a balanced understanding of the real situation is made difficult by the heat surrounding the debate. Despite the academic style displayed by all studies, conclusions are rarely grey; they tend to represent black or white scenarios, whereas past experience taught that big changes take time to deploy and pockets of the old-fashioned world will long continue to cohabit with new tendencies. The categorization of attitude towards innovation made by Lund et al. (2020) (innovators, early adopters, early majority, late majority and laggards) is a constant of previous waves and is being observed by most researchers in the AI case as well. A balanced approach is warmly recommended. Wise words may help to face this situation.

That full employment is more desirable than increased production combined with unemployment would be clear alike to the most sophisticated and the most primitive politician. Galbraith (1998).

Recommendation

According to many authors cited, the only feasible way ahead is to start in all more advanced economies a huge programme of reskilling not only in AI but also in big data, the Internet of Things, robotics and related disciplines. The recommendation from Schwab (2016:8) is clear: “governments, business, academia, and civil society […] have a responsibility to work together to better understand the emerging trends.” In the light of recent developments, we may suggest to focus on governing such trends rather than merely riding them. It is expected that even an untrained eye will be able to recognise this need over the next few decades. If we believe this is a long time ahead, it may be an illusion to the whole of mankind. Perhaps the last one.