Five Paradoxes of Artificial Intelligence

Author: Yan Deli, Senior Expert at Tencent, Tencent Research Institute

This is the era of artificial intelligence, filled with paradoxes—prediction paradox, employment quantification paradox, productivity paradox, data value paradox, industrial revolution paradox... We move forward amidst these paradoxes.

Prediction Paradox

When it comes to the future of artificial intelligence, people seem to never predict accurately, whether they are Turing Award winners, Nobel laureates, entrepreneurs, or startups. They always make unrealistic or overly conservative judgments. This is the prediction paradox of artificial intelligence. Historically, pioneers like Marvin Minsky, Allen Newell, and Herbert Simon have made astonishing predictions. For example, Marvin Minsky pointed out in 1970: "In three to eight years, we will have a machine with human-level intelligence." We are still striving for that today. Geoffrey Hinton, one of the most influential figures today, known as the "Godfather of AI," predicted in 2016: "We should stop training radiologists immediately. It is clear that deep learning will outperform radiologists within five years." This did not come true; in reality, the opposite has occurred, with the number and income of radiologists in the U.S. significantly increasing over the past decade. Demis Hassabis, who is on par with Hinton, stated (2025): "In the next decade, AI may help cure all diseases." As the most successful AI entrepreneur, Amodei (2025) believes that AI could double human lifespan within 5-10 years. We await the outcome of these recent predictions. The above are the predictions of six of the most famous figures. Regarding the topic, AGI is currently the most sought-after. There is a clear divide in the judgment of the timeline for AGI realization between entrepreneurs and scholars. Entrepreneurs are radically optimistic, even believing it will happen soon. There are mainly four categories: ① Already here (unaware), ② Coming soon (one to two years), ③ Coming in the short term (three to five years), ④ Coming in the long term (five to ten years, or even longer). Some of these judgments have been falsified, while others still require time for verification. As shown in the table below. Scholars tend to be generally pessimistic, even considering AGI a false proposition that will never arrive. For example, Iris van Rooij pointed out (2024): "Creating AGI with human-level cognitive abilities is impossible."

Table: Predictions of AGI Arrival Time by Notable AI Company CEOs
Source: Compiled by Tencent Research Institute, May 2026
People love to make predictions. Some do it for publicity, to motivate themselves, some for papers and projects, and others purely for verbal disputes. In fact, no one knows what the future holds. The future does not follow a predetermined script; it is shaped by the choices and actions of many, marked by a complete break from the past, filled with coincidences and paradoxes. Like a pig living in tranquility, it cannot predict the black swan of the Spring Festival.

Employment Quantification Paradox

The impact of technology on employment is a well-known subject, both an established field and a mysterious one. It is an ancient topic, like a ghost, wandering for hundreds of years. It is well-known, discussed by scholars and accessible to the layman. Its lower limit is extremely low, while its upper limit is extremely high, leading to endless debates and varied opinions. In recent years, scholars and institutions have turned to "data-driven" approaches, attempting to prove their points using methods that appear scientific, rigorous, and sophisticated. International organizations such as the OECD, IMF, World Economic Forum, UNCTAD, ILO, World Bank, as well as consulting firms like Goldman Sachs, McKinsey, and Pew Research Center have all released reports estimating the impact of AI on employment. As shown in the table below.

Table: Some Quantitative Estimates of AI's Impact on Employment
A single report can carry significant weight, even being regarded as a standard. However, when these reports are viewed together, it becomes apparent that the estimated results vary greatly, generally ranging from 0.4% to 67%, making them almost incomparable. This is disillusioning. The prerequisite for quantifying the impact of AI on employment is the ability to accurately grasp the pulse of technological development. As mentioned earlier, tech giants are still unable to accurately predict the future development of AI technology; how can economists quantify AI's impact on employment? They can only assume that technology is either static or progresses at a predetermined rate, which is not the case in reality. On the other hand, AI is not an independent influencing factor; multiple factors such as economic cycles, industrial economy, technological development, demographic structure, employment preferences, employment policies, globalization, and unexpected events all jointly influence employment. It is impossible to neatly separate AI from these interrelated and interacting factors. In this sense, quantifying AI's impact on employment is a paradox.

Productivity Paradox

Artificial intelligence is a new general-purpose technology characterized by universal applicability, continuous improvement, and the generation of innovation, serving as the engine for future economic growth. The term artificial intelligence has been around for 70 years, and the machine learning revolution has been ongoing for 14 years. The current wave of AI is in full swing, from large language models and multimodal systems to world models, intelligent agents, and physical AI, emerging one after another. However, the growth rate of productivity has not significantly accelerated and is even facing a productivity crisis (Rogers, 2024). Since the release of ChatGPT, the hourly labor productivity growth in the EU has fluctuated around 0%, with a growth of 0.1% in the first quarter of this year. In the 14 quarters from Q4 2022 to Q1 2026, only three quarters saw productivity growth exceed the long-term average level since 1999 (1.0%). As shown in the figure below. The U.S. has shown strong growth, with an average annual growth rate of 2.2% in labor productivity in the non-farm business sector from Q4 2022 to Q2 2026, standing out among Western countries, but this is only equivalent to the long-term average level since 1948 (Source: U.S. Bureau of Labor Statistics).

Figure: Recent Labor Productivity Growth in the EU (Source: Eurostat)
This simultaneous existence of "rapid technological innovation and disappointing productivity growth" is the productivity paradox. It is not the first time this has occurred. Nobel laureate Robert Solow wrote in 1987: "You can see the computer age everywhere but in the productivity statistics." This is regarded as the most classic expression of the "productivity paradox" or "Solow paradox." Regarding the productivity paradox, there are mainly three explanations: erroneous expectations, measurement errors, and time lags. Brynjolfsson believes (2017) that the lag explanation is the most persuasive and is the main reason for the productivity paradox, summarizing the lag effect of general-purpose technology on productivity as the "J-curve." General-purpose technologies require multiple rounds of secondary innovation, complementary innovation, and organizational change before they can have a substantial impact on productivity. Historically, steam engines, generators, and computers took 118 years, 91 years, and 49 years, respectively, to begin significantly driving productivity improvements after their invention, and 54 years, 40 years, and 21 years, respectively, after commercialization. Therefore, in the long run, the productivity paradox is not really a paradox. AI's significant enhancement of productivity still requires time.

Data Value Paradox

Data is the food of AI, requiring both quantity and quality, with a vast quantity and quality that can be described as "garbage in, garbage out (GIGO)," or in other words, data can be classified as "noble" or "vile." Data determines the upper limit of AI capabilities and is often referred to as "the new oil" (Clive Humby, 2006) and "the most valuable resource in the world" (The Economist, 2017), "equally important as energy and material resources" (2004 government document). The utility value of data is immense, yet it does not possess transactional or monetary value like ordinary commodities. As Li Guojie (2025) stated: "Data can only determine its value when used." The OECD (2024) reviewed data policy documents from 46 countries and found that the contexts in which "data" appears in policies are ranked from high to low as innovation, trust, society, market openness, utilization, employment, and access, with little focus on "transactions." Chen Changsheng (2023) pointed out: "Data exchanges are flourishing, but the 'transactions' within these exchanges are not developing well." "Amid the surge in data market construction in various regions, we should be wary of the behavior orientation that 'only data that has been traded is usable data.'" The value density of data is low, making it difficult to monetize, and its proportion in balance sheets is negligible, almost insignificant. According to the latest data from Shanghai Jiao Tong University's Shanghai Advanced Institute of Finance, 136 listed companies in China have disclosed matters related to data resources on their balance sheets, with a total amount of 3.786 billion yuan. Based on this, it is estimated that the former accounts for only 2.5% of the number of A-shares listed companies, while the latter only represents 0.3% of the scale of China's core AI industry. The three major telecom operators dominate the scale of balance sheet entries, with a total of 2.1 billion yuan in 2025, accounting for 55.46% of the total amount of listed companies' balance sheet entries. However, these data resources account for only about 0.06% of the total assets of the companies, which is trivial. As shown in the table below.

Table: Amount and Proportion of Data Resources Entered by the Three Major Telecom Operators (As of December 31, 2025)
Data Source: Compiled from consolidated balance sheets in financial reports.

Industrial Revolution Paradox

Every time a technological wave arises, some people compare it to the steam engine, electricity, or computers, excitedly proclaiming: "It will trigger the Fourth Industrial Revolution!" This phenomenon has persisted for at least half a century, with early examples including microelectronics (1984), computers (1988), nanotechnology (1994), the internet (2000), alternative energy (2010), and cyber-physical systems (2014). In the past decade, it has intensified with big data (2016), artificial intelligence (2016), the Internet of Things (2016), industrial internet (2017), blockchain (2017), quantum computing (2018), and smart manufacturing (2021), all of which have been assigned the mission of the Fourth Industrial Revolution. As shown in the figure below.

Figure: Distribution of Articles Titled "Fourth Industrial Revolution" on CNKI
Note: The selection criteria include articles with titles containing "Third Industrial Revolution," "Third Industrial Transformation," "Fourth Industrial Revolution," and "Fourth Industrial Transformation."
Regarding the current revolution, some believe it is the third, while others believe it is the fourth; moreover, the terms industrial revolution and industrial transformation are the same phrase "Industrial Revolution" in English. Therefore, this article does not make a strict distinction. We are almost constantly experiencing the Fourth Industrial Revolution in the media. The technologies that trigger this revolution are always "the new leaves urging the old leaves, the flowing waves pushing the waves behind," with preliminary statistics showing at least 20 different technologies. Each time a new technology emerges and is given significant meaning in the context of the Fourth Industrial Revolution, it seems to imply that the previous ones were all wrong. Currently, people generally believe that AI represents the Fourth Industrial Revolution. Demis Hassabis (2026) even pointed out: "The scale and speed of AGI may be ten times that of the industrial revolution." Some entrepreneurs in China view AI as the last technological revolution of human society. AI has proven its past, but how does AI itself prove? We cannot argue, but we emphasize two points. First, industrial revolutions and economic crises cannot occur simultaneously. Second, historical industrial revolutions are not recognized in advance but are narratives constructed afterward; those involved do not realize they are in an industrial revolution. The term "Industrial Revolution" only became known to the public 40 years after the first industrial revolution (1760s-1840s), promoted by Arnold Toynbee. It was 40 years after the second industrial revolution (1870s-1914) that economists began using the term "Second Industrial Revolution," and 55 years later, David Landes standardized its academic definition in "Unbound Prometheus" (1969). There is no unified understanding of the third industrial revolution; Jeremy Rifkin (2011) believes its theme is the integration of the internet and renewable energy; The Economist (2012) considers it the digitalization of manufacturing; Erik Brynjolfsson and Andrew McAfee (2011) argue that the third industrial revolution is driven by computers and networks. Let us look forward to this time being different.