Unlock Editor’s Digest for free
FT editor Roula Khalaf picks her favorite stories in this week’s newsletter.
The National Bureau of Economic Research has published a new paper by MIT superstar economist Daron Acemoglu that attempts to make AI dreams of a productivity renaissance, supercharged growth and reduced inequality a reality.
At this point, it seems almost heresy to say that AI won’t revolutionize everything. A year ago, Goldman Sachs economists estimated that artificial intelligence would increase annual global GDP by 7 percent over 10 years — or nearly $7 trillion in dollar terms.
Since then, Goldman’s prediction has become almost sobering, with even the IMF predicting that AI “has the potential to reshape the global economy.” FTAV’s personal favorite is ARK’s prediction that AI will help accelerate global GDP growth to 7 percent per year. 🕺
Professor Acemoglu – a likely future Nobel laureate – takes the other side. Alphaville’s emphasis below:
I guess so [total factor productivity] the effects of AI advances over the next 10 years will be modest – an upper bound that ignores the difference between hard and easy tasks would be about a 0.66% increase in total over 10 years, or about a 0.064% increase in annual TFP growth. When the presence of heavy tasks among those to be exposed to AI is recognized, this upper bound drops to around 0.53%. GDP effects will be slightly larger as automation and complementarity of tasks will also lead to greater investment. But my calculations suggest that GDP growth over the next 10 years should also be moderate, in the range of 0.93% – 1.16% over the 10-year totalassuming investment growth due to AI will be moderate and overall in the 1.4% – 1.56% range if there is a large investment boom.
As Acemoglu says, it’s “modest, but still far from trivial.” But as he notes, we also have to consider the fact that some of the most common use cases for AI are bad – ie deepfakes, etc.
Fighting them can boost growth in the same way that rebuilding a hurricane-ravaged city boosts growth, but it still reduces overall welfare. Alphaville’s emphasis below.
. . . When we include the possibility that new AI-generated tasks may be manipulative, the welfare impact may be even smaller. Based on figures from Bursztyn et al. (2023) concerning the negative effects of AI-powered social media, I present an illustrative calculation of social media spending, digital advertising and IT defensive attacks. These could add up to 2% to GDP, but if we use the figures of Bursztyn et al. (2023), their welfare impact may be −0.72%. This discussion suggests that it is important to consider the potential negative welfare implications of new AI-generated tasks and products.
Acemoglu is also skeptical that artificial intelligence will have a major impact on inequality — neither making it significantly worse nor improving it. Overall, however, his work suggests that “low-educated women may experience a small decline in wages, overall inequality between groups may increase slightly, and the capital-labor income gap is likely to widen further.”
The skepticism is interesting because Acemoglu is one-third of an influential trio of MIT economists who spearhead the university’s Shaping The Future Of Work initiative.
The professor emphasizes that the potential of generative artificial intelligence is great, but only if it is used mainly to provide people with better and more reliable information than chatbots prone to hallucinations and mechanically reconstituted images.
My assessment is that there are indeed much greater gains to be made from generative AI, which is a promising technology, but those gains will remain elusive unless there is a major shift in industry orientation, including perhaps a major shift in the architecture of most common generative AI models, such as are LLMs, with the goal of focusing on reliable information that can increase the marginal productivity of different kinds of workers, rather than favoring the development of general human-like conversational tools. The general nature of the current approach to generative artificial intelligence might be unsuitable for providing such reliable information.
Simply put, it remains an open question whether we need basic models (or the current kind of LLM) that can engage in human conversations and write Shakespearean sonnets if we want reliable information useful to educators, health professionals, electricians. , plumbers and other craftsmen.
Further reading:
— Economics of Manicure (FTAV)
— Investment outlook for the year ahead or ChatGPT? Take the Quiz (FTAV)
— Generative AI will be great for Generative AI Consultants (FTAV)