In a 1987 article in the Time book criticismRobert Solow, Nobel’s winner economist at MIT, said: “You can see the age of the computer everywhere but in productivity statistics.” Despite a massive increase in computing power and the growing popularity of personal computers, government figures have shown that global worker production, a key determinant of wages and standard of living, had stagnated for more than a decade. The “paradox of productivity”, as it is known, has persisted in the years nineteen and beyond, generating a huge and uncompromising whole of literature. Some economists have blamed mismanagement of new technology; Others argued that computers pale in economic importance compared to older inventions such as the steam and electricity machine; Still others blamed the measurement errors in the data and argued that once it has been corrected, the paradox has disappeared.
Almost forty years after Solow’s article, and almost three years since Openai published his Chatbot Chatgpt, we can be faced with a new economic paradox, it involving generative artificial intelligence. According to a recent survey carried out by economists in Stanford, Clemson and the World Bank, in June and July this year, almost half of all workers – 45.6%, to be precise – used AI tools. And yet, a new study, of a team of researchers associated with the MIT media laboratory, reports: “Despite 30 to 40 billion dollars of business investment in the Genai, this report results in a surprising result in the fact that 95% of organizations do not obtain any return.”
The study authors examined more than three hundred IA public initiatives and ads interviewed more than fifty business leaders. They defined a successful investment of AI like that which had been deployed beyond the pilot phase and had generated a measurable financial return or a marked gain of productivity after six months. “Only 5% of integrated AI pilots extract millions of values, while the vast majority remain stuck without measurable P&L”-profit-and-loss-“impact”, they wrote.
The interviews of the investigation aroused a range of responses, some of which were very skeptical. “The media threshing on LinkedIn says that everything has changed, but in our operations, nothing fundamental has changed,” said the director of operating a medium -sized manufacturing company. “We are dealing with some contracts faster, but that’s all that has changed.” Another respondent commented: “We have seen dozens of demos this year. Maybe one or two are really useful. The others are packaging or scientific projects. ”
Admittedly, the report stresses that some companies have succeeded in AI investments. For example, it highlights the efficiency created by personalized tools targeting back office operations, noting: “These first results suggest that systems compatible with learning, when targeted on specific processes, can offer real value, even without major organizational restructuring.” The survey also cites certain companies noting “an improvement in customer retention and sales conversion through automated awareness and intelligent monitoring systems”, which suggests that AI systems could be useful for marketing.
But the idea that many companies find it difficult to obtain substantial yields in peak with another recent survey, by Akkodis, a multinational consulting company. After having contacted more than two thousand business leaders, the firm noted that the percentage of CEOs which are “very confident” in the implementation strategies of the AI of their business went from ninety-two percent in 2024 to forty-nine percent this year. Confidence had also fallen among the directors of business technology, but not as much. These developments “can reflect the disappointing results of previous attempts of digital initiatives or AI, delays or implementation failures as well as concerns about scalability,” said the Akkodis investigation.
Last week, the media accounts of the study of MIT Media Lab coincided with a reduction in highly appreciated actions associated with AI, in particular Nvidia, Meta and Palant. Correlation is not causality, of course, and recent comments by Sam Altman, the director general of Openai, may have played a more important role in the sale, which was surely inevitable at a given time, given the increases in recent prices. During a dinner with journalists, Altman said that the evaluations were “crazy” and used the term “bubble” three times in fifteen seconds, CNBC reported.
However, the MIT study attracted a lot of attention, and after the initial series of research reports, a report emerged that the media laboratory, which has links with many technological companies, discreetly restricted access. The messages I left with the organization’s communications office and two of the authors of the report have not returned.
Although the report is more nuanced than a media coverage has not been made, it certainly raises questions about the great economic story that has underpinned the technological boom since November 2022, when Openai published Chatgpt. The short version of this story is that the dissemination of generative AI on the economy scale would be bad for workers, in particular knowledge workers, but ideal for businesses and their shareholders, because it would generate a big productivity jump and, by extension, profits.
A possible reason why this does not yet seem to have happened to recall the suggestion that management failures constituted the advantages of computers’ productivity in the nineteen years and in the early 90s. The study of the media laboratory revealed that some of the most successful AI investments were made by startups that use highly personalized tools in narrow fields of work. On the other side of the “Genai divide”, the study underlined less successful startups which “built either generic tools or trying to develop internal capacities”. More generally, the report indicates that the divisions between success and failure “do not seem to be motivated by the quality or regulation of the model, but seems to be determined by the approach”.
In theory, the novelty and complexity of generative AI can retain certain companies. A recent study, by the consulting firm Gartner, revealed that less than half of the CEOs are convinced that their main information officers are “warned”. But there is another possible explanation for the disappointing file highlighted in the report of the media laboratory: for many established companies, a generative AI, at least in its current incarnation, quite simply is not all that it was cracked. “It is excellent for brainstorming and the first drafts, but it does not keep the knowledge of customer preferences or does not learn from previous changes,” said a respondent to the investigation into the media laboratory. “He repeats the same mistakes and requires an extensive context for each session. For high issues work, I need a system that accumulates knowledge and improves over time. ”
Of course, there are many people who find useful AI, and there are academic evidence to support this: in 2023, two MIT economists found that Chatgpt exhibition allowed participants in a randomized trial to finish “professional writing tasks” more quickly and to improve the quality of their writing. The same year, other research teams identified results improving productivity for IT programmers who used the Github co -pilot and for customer support agents who had access to owners’ tools. Media laboratory researchers have discovered that many workers use their personal tools, such as GPT or Claude, to their work; The report refers to this phenomenon as “the shadow economy on the shadows” and comments according to which “it often offers a better return on investment” than the initiatives of the employer. But the question remains, and this is the one that the senior executives of the company will surely require more frequently: why have more companies have not seen these types of advantages which feed on the results?
Part of the problem may be that generative AI, however remarkable, has a limited application in many regions of the economy. Taken together, leisure and hospitality, retail, construction, real estate and the care sector – the children’s attachment and dealing with old or infirm people – employees of around fifty million Americans, but they do not resemble immediate candidates for an AI transformation.
Another important thing to note is that the adoption of AI throughout the economy may well be a long process. In Silicon Valley, people like to move quickly and break things. But economic history tells us that even the most transformative technologies, that economists call technology for general use, cannot be exploited with a maximum effect until the infrastructure, skills and products that can supplement them are developed. And it can be a long process. Scottish inventor James Watt invented his cylindrical steam engine in 1769. Thirty years later, most cotton factories in Great Britain were still powered by water wheels, in part because it was difficult to transport coal for use in steam engines. This did not change before the development of steam railways at the beginning of the 19th century. Electricity has also spread slowly and did not immediately lead to an economic scale of productivity growth. As Solow noted, the development of computers has followed the same scheme. (From 1996 to 2003, productivity growth across the economy finally increased, which many economists attributed to the delayed effect of information technology. Subsequently, however, it collapsed.)
