According to technology industry leaders and many experts and academics, artificial intelligence (AI) is about to transform the world as we know it with unprecedented improvements in productivity. While some believe that machines will soon do everything humans are capable of, ushering in a new era of limitless prosperity, other predictions are at least a little more grounded. For example, Goldman Sachs predicted that generative AI will boost global GDP by 7% over the next decade, and the McKinsey Global Institute predicts that the annual GDP growth rate could increase by 3 to 4 percentage points between now and 2040. Are they realistic estimates? As I noted in a recent article, the outlook is much more uncertain than most forecasts and conjectures suggest; However, although it is basically impossible to predict with certainty what AI will do in 20 or 30 years, something can be said about the next decade because most of these short-term economic effects must involve existing technologies and their improvements.
It is reasonable to assume that the greatest impact of AI will come from the automation of some tasks and improvements in the productivity of some workers in certain occupations. Economic theory offers some clues to evaluate these aggregate effects: according to Hulten’s theorem (named after the economist Charles Hulten), the aggregate effects of “total factor productivity” (TFP) are simply equivalent to the product of the participation of automated tasks by average cost savings.
Although the average cost savings is difficult to estimate and varies by activity, there are already some studies on the effects of AI on certain tasks; For example, Shakked Noy and Whitney Zhang examined the impact of ChatGPT on simple writing tasks—such as summarizing documents, or preparing grant proposals, or routine marketing materials—and Erik Brynjolfsson, Danielle Li, and Lindsey Raymond evaluated the use of ChatGPT. AI assistants in customer service. Together, these investigations suggest that the generative AI tools currently available save 27% in labor and 14.4% in total. What do we know about the proportion of tasks that will be affected by AI and related technologies? Based on recent study numbers, I estimate it to be around 4.6%, which implies that AI will only increase TFP by 0.66% in 10 years, or 0.06% per year (of course , as AI will also fuel an investment boom, the increase in GDP growth could be a little higher, perhaps in the range of 1% to 1.5%).
These are much smaller numbers than those of Goldman Sachs and McKinsey; To increase them, you have to drive productivity improvements at a micro scale or assume that more tasks in the economy will be affected, but neither of those scenarios seems plausible. Labor savings well above 27% not only fall outside the range offered by existing studies, but do not align with the observed effects of other even more promising technologies. For example, industrial robots transformed some manufacturing sectors and appear to have reduced labor costs by 30%.
Similarly, we are unlikely to see replacement in much more than 4.6% of tasks, because most manual and social tasks are not even remotely within the reach of AI. According to a 2019 survey of US companies, only 1.5% had any investment in AI. Even if those investments have accelerated over the past year and a half, widespread AI adoption is a long, long way off.
Of course, AI could have more significant effects than my analysis allows if it revolutionizes the process of scientific discovery or creates many new tasks and products; The recent discoveries it facilitated on crystal structure and advances in protein folding do indeed suggest these possibilities. But these breakthroughs are unlikely to be a major source of economic growth over the next 10 years. On the other hand, my own estimates might be excessive; The early adoption of generative AI naturally took place in sectors where it performs reasonably well, that is, in tasks where success can be objectively measured. In those cases the models can learn from external information and available historical data.
However, for many of the 4.6% of tasks that could be automated within the next 10 years—evaluating applications, diagnosing health problems, and offering financial advice—there are no such clear indicators of success, and usually involve complex variables that depend on the context. In those cases, learning from external information is much more difficult and generative AI models must rely instead on the behavior of existing workers. In these circumstances there will be less room to significantly improve human work; I estimate then that a quarter of 4.6% of tasks fall into the “difficult learning” category and will experience smaller increases in productivity. Once we apply that adjustment, the 0.66% TFP growth figure drops to approximately 0.53%.
What about the effects on workers, wages and inequality? The good news is that, compared to the effects of previous waves of automation—such as those based on robots or software systems—the effects of AI could be more widely distributed across demographic groups. In that case, it won’t have as big an impact on inequality as previous automation technologies. However, I find no evidence that AI will reduce inequality or drive wage growth. Some groups—especially native-born white women—are significantly more exposed and will be negatively affected, and capital will outperform labor overall.
Economic theory and available data justify a more modest and realistic outlook for AI. There isn’t much support for the argument that we shouldn’t worry about regulation because AI will be the proverbial rising tide that lifts all boats. AI is what economists call a general-purpose technology, we can do a lot of things with it, and there are better things than automating work and driving the profitability of digital advertising… but whether we uncritically embrace techno-optimism or let the industry be technology sets the agenda, we could waste much of its potential.
Follow all the information Economy and Business in Facebook and xor in our weekly newsletter
