Almost two years have passed since OpenAI released GPT-3.5 to great fanfare. Bill Gates, co-founder of Microsoft, compared the technology’s arrival to his first encounter with the graphical user interface—a breakthrough that reshaped personal computing—in the 1980s. Others predicted that generative artificial intelligence (AI) would rapidly transform economies around the world, leaving many millions unemployed. Yet despite the hype and the worries, ai’s impact has been muted thus far. According to America’s Census Bureau, only 6% of businesses use AI to produce goods and services. Output and labour-productivity growth, meanwhile, remain far below the soaring heights of the computer age in the 1990s.
Why has AI so far failed to live up to its promise? Lessons from the computer age can shed light on the question. As with AI today, the early years of the computer age were marked by predictions of economic transformation. In 1965 Herbert Simon, a giant of computer science, declared that “machines will be capable within 20 years of doing any work that a man can do.” Two decades after Simon’s prediction, the promised productivity revolution remained elusive. In 1987 Robert Solow, a Nobel laureate, famously quipped that “you can see the computer age everywhere but in the productivity statistics.” Only in the late 1990s did the economic transformation at last materialise, leading Solow to acknowledge—three decades after the initial exuberance—that computers had begun to reshape the economy.
Three main factors contributed to the eventual arrival of a computer-age productivity boom: companies ramped up investment in information technology, computer and software prices fell rapidly (see chart), and bosses found new ways to integrate the tech into their operations. Are these factors in evidence today?
View Full Image
Begin with IT investment. Starting in 1995, firms ramped up spending on computer hardware, network infrastructure and software. Between 1995 and 2000, their investment in information-processing equipment and software rose by an average of 20% a year in real terms. Research by Kevin Stiroh of the Federal Reserve Bank of New York has found that firms were investing nearly $400bn in such technologies by 1999, accounting for over 30% of all non-residential fixed investment.
By contrast, recent capital expenditure has been underwhelming. Over the past two years, business investment in information-processing equipment and software has grown by around 4% a year. AI investment may be more focused on intangible assets, such as algorithms and data, which are more difficult to measure than physical capital. Payments to startups for custom tools may show up as operating expenses in the statistics, for example. Even so, you would expect at least a rise in software investment. Instead, spending on both pre-packaged commercial software—such as Microsoft 365—and custom-built systems, including AI tools tailored to specific workflows, is surprisingly low. Growth in software investment over the past year was about three times lower than in the late 1990s in real terms, and remains well below the long-term average.
The second half of the 1990s also witnessed a dramatic fall in the quality-adjusted price of computer hardware and software. From 1995 to 2000 prices for information-processing equipment and software dropped by a third, producing cheaper and better computers. The AI era has yet to see a corresponding decrease in prices: over the past five years, those for software and information-processing equipment have barely budged. Indeed, in the most recent quarter, the price index for these goods rose at an annualised rate of 4%. Even as the underlying technology is becoming cheaper, middlemen who repackage AI tools are increasingly adding margins and driving up prices.
What about the final ingredient in the economic revolution of the 1990s? For a technology to provide productivity gains, companies must retool operations and business models to integrate it. Consider the example of Walmart. In the 1990s the retailer boosted productivity by embedding a new software system—Retail Link—into its operations, granting suppliers real-time access to sales and inventory data. AI adoption today remains largely confined to narrow applications within existing operations, such as a financial-services firm using an AI app for fraud detection. Most firms do not have the data infrastructure required to train custom firm-specific models. To unlock AI’s full potential, more fundamental changes will be required.
Given these constraints, it might be prudent to recall the words of Rudi Dornbusch, an economist who spent his career at the Massachusetts Institute of Technology: that in economics things happen slower than you thought they would and then faster than you thought they could. AI may eventually produce extraordinary productivity growth, but at present it appears to be some distance from the take-off experienced in the 1990s.
Perhaps a more fitting comparison is to the 1970s—a period when technological promise mingled with disappointing productivity growth. The memory chip and silicon microprocessor, which powered the personal computer, were introduced around 1970. Yet 20 years later, less than 10% of the world’s businesses were using computers. As the world moved into the information age with the arrival of email, mobile phones and the internet, productivity growth remained stubbornly low. From 1975 to 1994 labour productivity in America averaged a lacklustre 1.7%. Then things finally got going. The AI revolution seems to be following a similar path.
© 2025, The Economist Newspaper Limited. All rights reserved. From The Economist, published under licence. The original content can be found on www.economist.com
