Are we overestimating Artificial Intelligence?

Article by: Nicola Costantino, Rector of the Politecnico di Bari, where he served as Full Professor of Management Engineering.

The debate over the employment impact of Artificial Intelligence (AI) is becoming increasingly intense, also because the available data are still partial and often contradictory. According to ISTAT, in 2025 in Italy, 15.7% of SMEs and 53.1% of large companies reported having integrated AI into their production processes, figures that were sharply higher than in 2024 (respectively +8% and +20.6%). Yet this does not seem to have had any significant impact at the aggregate level, at least so far, either on national GDP, which grew by only 0.5% compared with the previous year, or on unemployment, which actually declined. Once again, the contradiction between “reasonable” forecasts and actual outcomes seems to be emerging.

Even the views of leading experts on AI’s employment impact do not appear entirely aligned: on the one hand, Mark Zuckerberg states, “We are already seeing projects that once required entire teams being completed by a single talented person working alongside AI agents”; on the other, Andrew Ng, co-founder of Google Brain, says: “Job losses caused by AI have so far been overestimated.”

Employment forecasts that failed to materialize

In 2013, a study by C.B. Frey and M.A. Osborne of the University of Oxford attracted considerable attention. According to their analysis, as a result of automation processes, 47% of jobs in the United States would be at risk of replacement over the following 10 years. According to the 2016 World Bank Development Report, that share rose to 57% for OECD countries and even 77% for China, due to the lower initial level of automation in those countries. Of course, it was argued, the spread of new technologies also creates new professions: according to Forrester Research, during the 10 years following 2015, 13.6 million new jobs would have been created in the United States, against the loss of 22.7 million more traditional occupations, resulting in a negative balance of 9.1 million jobs. Why did none of these forecasts materialize in reality, at least within the expected timeframe?

In 1987, Robert Solow, who that same year received the Nobel Prize in Economics partly for his studies on the productivity effects of technological developments, stated: “You can see the computer age everywhere but in the productivity statistics,” probably because of the enormous complexity of the variables involved, not only technological but also organizational and social. And indeed, the history of the Third and Fourth Industrial Revolutions is marked by spectacular and unexpected successes, such as the internet, but also by more or less dramatic failures. We may recall the excesses of the so-called net economy around the year 2000, with the bubble of many dot-com companies reaching astronomical valuations before even testing their supposed target market; the short-lived, yet costly, trend of virtual real estate markets on Second Life; or, more recently, the enormous expectations surrounding the metaverse, which Mark Zuckerberg strongly promoted but which have so far gone largely unmet. Certainly, compared with these examples, AI is a far more structured and established reality, and one in which truly enormous investments are still being made.

Returns on profitability

The five companies with the highest market capitalization in the world are all listed on Wall Street: Nvidia, Apple, Alphabet (Google), Microsoft and Amazon. All are involved to varying degrees in AI, with a combined market value of more than 18 trillion dollars, a figure expected to grow as they continue to raise capital to invest in increasingly massive computing infrastructure. Whereas in the past we were accustomed to considering the environmental impacts of ICT relatively marginal, especially in terms of energy consumption, today’s enormous data centers, and even more so those of tomorrow, are beginning to have a significant impact on electricity consumption, as well as on water consumption for cooling, in the areas where they are located. This is to the point that technically and economically feasible versions of what once seemed like science-fiction scenarios are now beginning to be considered, such as placing data centers in space in order to optimize solar energy capture and heat dissipation.

These enormous investments rely on expectations, however justified they may be, of equally enormous profitability returns, generated by the development of Artificial General Intelligence (AGI) tools capable of understanding and applying human knowledge at least at the same level as, if not beyond, our cognitive abilities in any field. Will such returns actually materialize?

Their realization is essentially based on several assumptions, more or less explicit, which may not necessarily all prove correct. First, that the paths being pursued by industry giants such as OpenAI in the development of AGI are the most efficient and will therefore generate highly satisfactory returns. Second, that the growth of AI capabilities, especially in terms of Large Language Models (LLMs), is unlimited, because these systems are enriched every day by new contributions, meaning new digitized knowledge, making the models increasingly powerful through an exponential growth that appears to have no boundaries. Third, according to some, thanks to this unlimited growth, AGI is destined to evolve into Artificial Super Intelligence (ASI), a superintelligence capable of surpassing human intelligence even in creativity, wisdom and relational abilities.

The crisis of Large Language Models

These expectations, certainly reasonable, at least the first two, are nevertheless still subject to equally substantial uncertainties. Are we certain that AGI development could not be achieved through methods that are overall far less costly? The experience of DeepSeek, the Chinese AI system developed with significantly lower investment than its Western competitors while offering performance comparable to ChatGPT, appears to raise doubts in this regard. The history of industrial revolutions is full of examples of dramatic leaps in efficiency in which followers overtook the original innovators. If the same happens with AI, what will become of today’s industrial giants, which have concentrated investments on a scale unprecedented in human history?

Are we certain that the growth of digitized knowledge will continue at its current pace? It is expected that by 2030 virtually all human knowledge will have been made available to LLMs. After that date, further digitization will proceed at the pace of the actual “human” production of new knowledge, supplemented by autophagy, which is already taking place, namely feeding LLMs with documents produced by LLMs themselves at a speed vastly greater than that of humans. However, if we consider that LLMs produce documents not by “reasoning” but by “drawing” from the documents they have acquired, word by word, the most statistically likely semantic sequences, it is reasonable to expect that they may not be able to generate “new” knowledge, but only recombinations of previously acquired knowledge.

Moreover, the experience developed so far seems to show that repeated autophagy produces the so-called Model Autophagy Disorder (MADness), which causes the “collapse” of the computational output, meaning a progressive loss of meaning. If this is true, the growth of LLM capabilities is likely to slow dramatically within a few years, because the “new” knowledge made available to them will no longer, as happens today, come from the introduction of large volumes of data already produced in the past, such as when a text written 10 years ago is uploaded into a model, but only from the new and “original” production of human beings.

Will it be possible to develop Superintelligence? In 1905 Albert Einstein published four papers that completely revolutionized our physical understanding of the world. He reached these results without conducting new experiments, but “simply”, so to speak, by reasoning on the results of experiments and theoretical developments produced by others. The concept of the relativity of time, to give just one example, did not previously exist, and it is highly unlikely, if not impossible, that an LLM fed with the same knowledge available to Einstein would have been able to generate it, regardless of the amount of computational resources at its disposal.

Ultimately, it is possible to put forward a reasonable hypothesis, nothing more and certainly not a certainty: perhaps we are overestimating a new tool that is unquestionably very powerful and extremely pervasive, but that may never be able to replace human beings in the most significantly creative activities. Is that really the case? Niels Bohr, winner of the Nobel Prize in Physics in 1922, wrote: “Prediction is very difficult, especially if it’s about the future.”