Yan Deli, a senior expert at Tencent and a researcher with Tencent Research Institute, argues that the AI era is defined by paradoxes. In his article, he identifies five: the forecasting paradox, the job quantification paradox, the productivity paradox, the data value paradox, and the industrial revolution paradox.
The forecasting paradox
The article says people have consistently failed to predict AI’s trajectory with much accuracy, whether they are Turing Award winners, Nobel laureates, founders or executives. Some calls have been far too aggressive, while others have been too cautious.
Yan points to earlier examples from AI pioneers including Marvin Minsky, Allen Newell and Herbert Simon. Minsky said in 1970, “In from three to eight years we will have a machine with the general intelligence of an average human being.” The article notes that researchers are still working toward that goal.
It also revisits a 2016 forecast from Geoffrey Hinton, often described as the “godfather of AI,” who said training radiologists should stop immediately because deep learning would outperform them within five years. Yan writes that the prediction did not materialize and that, over the past decade, both the number of radiologists in the United States and their income have risen sharply.
On more recent forecasts, the article cites Demis Hassabis, who said in 2025 that AI may help cure all diseases “over the next decade or so.” It also cites Amodei, described in the article as a highly successful AI entrepreneur, who said in 2025 that AI could double human lifespan within five to 10 years. Yan says these newer claims still need time to be tested.
The same divide appears in debates over artificial general intelligence, or AGI. According to the article, entrepreneurs tend to be much more optimistic than academics. Their estimates generally fall into four buckets: AGI has already arrived, it is one or two years away, it is three to five years away, or it is still five to 10 years away, if not longer. Some of those views have already been disproved, while others remain open.
Academics, by contrast, are described as broadly more pessimistic. Some even reject AGI as a valid destination. Yan cites Iris van Rooij, who said in 2024: “Creating AGI with human-level cognitive abilities is impossible.”
The article says people like to make predictions for many reasons, including promotion, self-motivation, research agendas and simple argument. But the future does not follow a fixed script. In Yan’s telling, it is shaped by collective choices, discontinuities, coincidence and contradiction.
The job quantification paradox
AI’s impact on employment has become a heavily studied topic, and recent work has leaned hard on numerical estimates. Yan lists reports from the Organisation for Economic Co-operation and Development (OECD), the International Monetary Fund (IMF), the World Economic Forum, the United Nations Conference on Trade and Development, the International Labour Organization, the World Bank, Goldman Sachs, McKinsey and the Pew Research Center.
Looked at one by one, the reports carry weight. Put together, the article says, they point in very different directions. The range of estimates runs from roughly 0.4% to 67%, making them difficult to compare in any meaningful way.
Yan argues that any attempt to quantify AI’s impact on employment depends on a credible grasp of where the technology is heading. If leading technologists themselves cannot reliably forecast AI development, economists are left building models that assume either a static technology base or a predefined pace of progress. The article says neither assumption matches reality.
It adds that AI is only one force among many that shape labor markets. Business cycles, industrial conditions, technological development, demographics, worker preferences, employment policy, globalization and unexpected shocks all interact with one another. In that setting, isolating AI cleanly enough to measure its effect becomes nearly impossible. That, the article says, is the paradox.
The productivity paradox
Yan describes AI as a new general-purpose technology with broad applicability, the capacity for ongoing improvement and the ability to generate follow-on innovation. He writes that the term artificial intelligence has been around for 70 years, that the machine learning revolution has been underway for 14 years, and that the current wave has moved quickly from large language models and multimodal systems to world models, agents and physical AI.
Even so, productivity growth has not clearly accelerated. The article cites Rogers (2024) and says economies may even be facing a productivity crisis. Since the release of ChatGPT, growth in labor productivity per hour in the European Union has hovered around 0%, with growth of 0.1% in the first quarter of this year. Across the 14 quarters from the fourth quarter of 2022 to the first quarter of 2026, only three quarters exceeded the long-term average growth rate of 1.0% recorded since 1999.
The U.S. numbers have been stronger. From the fourth quarter of 2022 to the second quarter of 2026, labor productivity in the nonfarm business sector rose by an average annual rate of 2.2%, according to the U.S. Bureau of Labor Statistics. Yan says that made the United States stand out among Western economies, though the pace still only matched the long-run average since 1948.
The article describes this coexistence of rapid technical progress and underwhelming productivity gains as the productivity paradox. It recalls Robert Solow’s well-known 1987 line: “You can see the computer age everywhere but in the productivity statistics.”
Yan groups the main explanations into three categories: mistaken expectations, measurement error and time lags. He cites Brynjolfsson (2017), who argued that lag effects are the most convincing explanation and summarized the pattern as a J-curve. A general-purpose technology, the argument goes, needs rounds of secondary innovation, complementary innovation and organizational change before it leaves a clear mark on productivity.
The article gives historical examples. Steam engines, generators and computers only began to lift productivity significantly 118 years, 91 years and 49 years after their invention, and 54 years, 40 years and 21 years after commercialization, respectively. By that logic, Yan says, AI may still need time before its productivity effects become clear.
The data value paradox
Data, in the article’s phrasing, is the “food” of AI. Systems need huge volumes of it, and they need high-quality input as well. Otherwise, as the familiar phrase goes, garbage in, garbage out.
Yan writes that data sets the upper bound on AI capability and is often called “the new oil,” citing Clive Humby in 2006, as well as “the world’s most valuable resource,” citing The Economist in 2017. He also notes a 2004 Chinese policy document that described data as being as important as energy and material resources.
Yet the article says data’s use value does not translate cleanly into transaction value or monetary value in the way ordinary goods do. It quotes Li Guojie, who said in 2025: “Data can only determine its value in use.” It also cites an OECD review from 2024 of data policy documents from 46 countries. In that review, the policy contexts in which “data” appeared most frequently were innovation, trust, society, market openness, utilization, employment and access, rather than trade.
The article also quotes Chen Changsheng, who said in 2023 that data exchanges were appearing everywhere but that the “trading” inside them was not developing very well. He warned against a policy orientation in which only data that has gone through a market transaction is treated as usable data.
On corporate financial statements, Yan says data has low value density and is hard to monetize, leaving it with a very small share on balance sheets. Citing the latest figures from Shanghai Advanced Institute of Finance at Shanghai Jiao Tong University, he says 136 listed companies in China disclosed items related to bringing data resources onto the balance sheet, with a combined amount of 3.786 billion yuan. Based on those figures, the article says that represented just 2.5% of all A-share listed companies, while the amount itself equaled only 0.3% of China’s core AI industry size.
The three major telecom operators dominated those disclosures. In 2025, their combined amount reached 2.1 billion yuan, or 55.46% of the total amount booked by listed companies. But those data resources accounted for only about 0.06% of their total assets.
That gap between strategic importance and financial representation is what the article labels the data value paradox.
The industrial revolution paradox
The final paradox concerns the repeated claim that each major wave of technology will trigger a “Fourth Industrial Revolution.” Yan says this pattern has lasted for at least half a century. Earlier examples include microelectronics in 1984, computers in 1988, nanotechnology in 1994, the internet in 2000, alternative energy in 2010 and cyber-physical systems in 2014.
Over the past decade, the label has been attached to big data, artificial intelligence, the internet of things, the industrial internet, blockchain, quantum computing and intelligent manufacturing. The article says media narratives have almost been living inside a permanent Fourth Industrial Revolution, even as the technology cast as its driver keeps changing. Yan’s rough count puts the number of candidate technologies at more than 20.
The article notes that some observers call the current shift the third industrial revolution, while others call it the fourth. It also says “industrial revolution” and “industry revolution” are often collapsed into the same English phrase, “Industrial Revolution,” so it does not insist on a strict distinction.
On AI specifically, Yan says many people now treat it as the defining force of the Fourth Industrial Revolution. He cites Demis Hassabis again, this time for a 2026 comment that “the scale and speed of AGI may be 10 times that of the Industrial Revolution.” He also notes that some Chinese entrepreneurs describe AI as the final technological revolution in human society.
Still, the article stresses two points. First, it says an industrial revolution and an economic crisis cannot happen at the same time. Second, industrial revolutions are usually named in hindsight, not recognized cleanly by the people living through them. The First Industrial Revolution, dated in the article from the 1760s to the 1840s, only became widely known under that label 40 years after it ended, helped by Arnold Toynbee. The Second Industrial Revolution, dated from the 1870s to 1914, was only described with that term by economists 40 years after it ended, and its academic definition was standardized 55 years later in David Landes’s 1969 book The Unbound Prometheus.
As for the Third Industrial Revolution, the article says there is still no consensus. Jeremy Rifkin argued in 2011 that it centered on the fusion of the internet and renewable energy. The Economist argued in 2012 that its core lay in the digitization of manufacturing. Erik Brynjolfsson and Andrew McAfee argued in 2011 that the third industrial revolution was driven by computers and networks.
Yan ends by saying AI has already been used to explain the past and is now being asked to define the future as well. Whether it can prove itself on those terms is something only reality can answer.