New technologies create excitement. The invention of the railway led to the UK’s “Railway Mania” of the 1840s, with investors piling into railway shares. The radio also appealed to the imagination in the 1920s. And more recently, the euphoria over Internet adoption caused the Nasdaq to quintuple between 1995 and 2000. The hysteria stems in part from high expectations about how far and how fast innovations can boost human well-being and productivity, in addition to the “fear of missing out”. But in any case, the initial bubble burst as reality caught up with expectations.
The rise of generative AI – particularly large language models, such as OpenAI’s ChatGPT – has sparked a similar frenzy. Since ChatGPT’s launch last November, glowing reports of the technology’s potential economic impact — which can automate tasks from essay writing to code generation — have been flooding in quickly. Goldman Sachs estimated it could lead to productivity gains that could boost global GDP by 7 percent over a ten-year period.
Estimating the benefits is largely a guessing game. Most experts agree that it will take time to pay off. Indeed, the impact of previous technologies has often been conceptualized using the “J-curve effect”, where productivity may initially fall when applied, before rising sharply. Will generative AI follow a similar path?
Electricity, railroads, and computers all took decades before sparking a productivity boom. In comparison, AI technology is much less capital intensive and does not require large-scale development of new infrastructure. ChatGPT is accessible with a mouse click; more than 100 million people have already done so. Generative AI systems, however, will require massive computing power, which does not come cheap. Companies will have to retrain their workforce and adapt their business models. This will take time, although the user-friendliness will make adoption easier than previous technologies.
Other factors could offset that lead, such as regulation. Given its power, AI gurus have already called for a moratorium on the further development of advanced models. Unlike historical inventions that replaced human brawn, generative AI can perform cognitive tasks such as writing, analysis, and design. It can strengthen these activities together with people, but policies and legislation must also evolve to govern it and deal with the social and workforce consequences. The uncertainty, disruption and compliance will be speed bumps along any J curve.
Generative AI can also directly undermine productivity. For efficiency gains to benefit an economy, the time freed up must be used productively. But the technology can create new problems. It can be used to simulate data, manipulate data, and help students pass off their homework as their own. Dealing with that can be difficult. Ironically, it also allows productivity killers like junk email and online distractions to be more productive. By improving fraud detection, it could eventually help it clean up its own mess.
How highly generative AI can climb the curve depends on its usefulness once – and indeed if – it lifts the dip. It could boost productivity in knowledge-based jobs, speed up physician diagnosis and legal contract drafting, but some service sectors will be less affected. By accelerating the research process itself, it could drive technological progress and iterative productivity gains. Complementary changes are also important. Railroads eventually increased efficiency, but that’s because industry and trade also flourished. If governments apply AI, for example to reduce form filling, it would reduce other drags on productivity.
There is no doubt about the potential of generative AI. Its ability to stimulate cognitive activity, which is more difficult to value, means we may not be able to accurately measure its impact. But as previous technologies demonstrate, productivity gains are not guaranteed until the technology can be deployed effectively. We have to keep our feet on the ground.
