Generative artificial intelligence tools are changing the way we use the internet. Some people are concerned that such systems could replace countless jobs; others expect quick productivity gains. As with past technological innovations, change is likely to be less dramatic than the hype suggests.
From checking for typos and recalling facts through to writing poems and creating art, artificial intelligence (AI) systems can now do tasks it would have been hard to imagine them being capable of even a few years ago.
These new generative AI systems, like ChatGPT, DALL-E2 and OpenArt, have exploded in popularity over recent months. Recent data suggest that ChatGPT has over 100 million users and OpenAI (owner and developer of ChatGPT) receives approximately one billion visitors to its website each month. Analysis from Swiss bank UBS indicates it is the fastest growing consumer app in history.
While technology trends like web3 and the metaverse have attracted plenty of investment and media attention, whether they will have much real-world impact is still unclear. Generative AI seems to be different.
There is broad consensus among technology researchers and commentators that it is a highly significant development – at least as significant as the smartphone; probably as significant as the web; and perhaps as significant as electricity. Some go further still, claiming that it is the precursor to machine super-intelligence, and therefore represents an existential risk to humanity, which must be mitigated urgently.
This might seem surprising, given that AI systems have been part of UK citizens’ daily lives for more than a decade. Whether or not we are aware of it, with their ability to detect patterns in historic data, AI systems have been ranking our social media feeds, determining what digital ads we are shown, and recommending films we might like to watch on streaming services.
So, what is different about generative AI? To paraphrase technology analyst and investor Benedict Evans, its novelty lies in running the pattern-matching machine in reverse. Instead of identifying existing examples that fit a given pattern, it ‘generates’ new examples of the pattern. The output of generative AI systems is anything that can exist as a digital file – text, imagery, audio or video. The result is that an unprecedented super-proliferation of content is underway.
What are large language models?
ChatGPT – one of the generative AI systems that has burst onto the scene – is an example of a large language model (LLM). Here, we focus on the economic implications of such systems.
LLMs produce text outputs in response to natural language instructions from a user (known as ‘prompts’). Trained on vast corpuses of text from books and the web, LLMs work by making iterative statistical predictions about what word should come next in a sequence. As a result, they can be used to produce articles, essays, computer code, stories, poems, song lyrics, messages, letters, political speeches and eulogies that are more or less indistinguishable from those written by humans.
The best-known LLMs are those developed by Silicon Valley company OpenAI, largely thanks to the popularity of its consumer-facing application ChatGPT. Released in November 2022, it is underpinned by OpenAI’s GPT-3.5 model (premium subscribers can use the even more powerful GPT-4).
Other examples of LLMs include Google’s LaMDA, accessible to users of its ChatGPT-like beta product Bard, Meta’s LLaMA and Anthropic’s Claude. Microsoft, meanwhile, has invested more than $10 billion in OpenAI and integrated GPT-3.5 into its search engine Bing. And well-capitalised start-up firms are in the process of ‘productising’ LLMs for specific use-cases, including copywriting and the drafting of legal contracts.
Can LLMs really replace human workers?
Once they have got over marvelling at its ingenuity, ChatGPT users often find themselves irritated or enraged by its shortcomings. Partly because of its user interface, many instinctively expect it to work like a search engine, retrieving information from a database and presenting results with a high degree of factual accuracy. But this is not what LLMs are designed to do.
Instead, GPT-3.5’s probabilistic approach frequently produces what have been called ‘hallucinations’ – factoids, non-existent URLs and fabricated academic references that are all the more hazardous because they are so plausibly articulated. A more banal frustration is the need to copy-and-paste ChatGPT’s outputs into other software to make use of it. With this fact-checking and administrative burden, ChatGPT users could be forgiven for expressing scepticism at the idea that LLMs are poised to replace 300 million jobs.
But what this neglects to consider is that LLMs can also be used programmatically, via application programming interfaces (known as APIs). As an experiment, Ankur Shah and I built a database of information about UK insurance products. We then wrote a prompt for an insurance product review article and programmed a simple system to populate the prompt with the product data, send it to OpenAI’s API, and push the LLM’s output directly into a web content management system.
We were able to publish hundreds of online review articles like this one in less than an hour, at a cost of around $7. Completing the same project with human freelance writers would have taken several months, and cost closer to $70,000. Further, including real product data in the prompts pre-empted the LLM hallucinating incorrect cover limits or imaginary policy features.
So, if used correctly and with careful prompts, LLMs like ChatGPT could indeed change the way in which certain jobs are done. In light of this practical experience, it is plausible that LLMs will mean disruption for writers – most acutely in fields like content marketing, where subject matter expertise and a distinctive authorial voice are less important than in journalism or fiction.
The same goes for customer services. Chatbots powered by LLMs and trained on domain-specific data are a major upgrade on their predecessors – contrast a GPT4-powered bot like Intercom’s Fin, for example, with Aviva’s clunky online assistant. The same bots can be connected to speech APIs, opening the potential for contact centres to be fully automated.
For organisations able to launch these systems, this would mean material reductions in operating costs. For customers, it would mean an end to queuing for web-chat or telephone support, since AI systems can handle hundreds of interactions simultaneously. But for front-line customer services workers, the prospect will seem rather less utopian.
What about productivity?
These efficiency gains and operating cost-savings ought to have a favourable impact on productivity in some industries. This is especially true in sectors like financial services, telecoms, media and education. But it does not necessarily follow that higher sectoral productivity will lead to productivity improvements across the whole economy.
Indeed, despite widespread adoption of the previous generation of digital technologies, productivity growth has stagnated since the global financial crisis of 2007-09, to the continuing puzzlement of economists.
Figure 1: Output per worker, by country
Source: Feenstra et al (2015), Penn World Table (2021)
It could be that we have been living in an unproductive bubble, or that most organisations have not yet worked out how they can use mobile apps or big data analytics to become significantly more productive.
Whatever the reason, the implication is that we should not take for granted the idea that LLM-enabled cost-savings will produce productivity improvements for the whole economy. Back in 1987, economics Nobel laureate Robert Solow famously quipped that ‘the computer age is everywhere except in productivity statistics’. The same could well be true for AI.
But as Diane Coyle, one of the Economics Observatory’s lead editors, writes, the real value of LLMs and other generative AI systems will not come from enabling a small number of technologically-advanced companies to slash their costs or invent new products. Rather, it will come from changing how things are produced – as assembly lines did in the 1910s, or just-in-time production in the 1980s.
To this end, the most important facet of LLMs may well be their aptitude for writing computer code. The UK faces a chronic shortage of software developers, accentuated by Brexit and the Covid-19 pandemic – in August 2022, there were more than 30,000 vacancies in this field. In this context, there are two roles that LLMs could play.
First, they can increase the productivity of today’s developers, partly closing the skills gap. One lever is GitHub Co-Pilot, an LLM-powered tool sometimes described as ‘autocomplete for code’. This tool already has more than a million users and enables developers to write software up to 55% faster than they could previously.
But this pales in comparison with the second possibility, which is that large numbers of people with little or no coding experience could start using LLMs to build software. Until now, the main constraint on what computer systems can be built has been the availability of workers with skills in Python, PHP, JavaScript and so on; but in future it may well be the capability to imagine what a system might do and specify in natural language how it should function.
If we believe that LLMs will indeed change how software is produced, nurturing that capability through industrial and education policy would be a smart move for policy-makers.
Conclusion
When trying to make sense of the economic implications of new technologies, one of the biggest challenges is the obfuscating effect of hype and ‘criti-hype’. In the case of LLMs, technology executives have incentives to talk up the science fiction-inflected risk of artificial general intelligence obliterating humanity, as it increases the perceived value of actually existing products (like chatbots), which might otherwise seem trivial.
At the same time, in academia, think-tanks and the media, the booming market for commentary and opinion on the social and ethical implications of generative AI seems to give incentives for alarmism.
As is often the case, the reality is more mundane. The economic impact of LLMs will be felt first in content creation and customer services, before software development is changed dramatically – with any luck, to the benefit of wider productivity.
Where can I find out more?
- AI and economic productivity: expect evolution, not revolution: Jeffrey Funk makes the case that it will be years before AI has a material impact on economic productivity.
- Pre-empting a generative AI monopoly: Diane Coyle encourages competition policy-makers to be on the front-foot as the next generation of big tech companies emerges.
- Hard Fork podcast: Tech journalists Kevin Roose (New York Times) and Casey Newton (Platformer) report regularly and entertainingly on the latest developments in generative AI.
- AI and productivity growth: Poll of top American economists’ views on AI’s potential impact on growth, conducted by Chicago Booth’s Clark Center for Global Markets. The Center has put the same question to European experts, as well as asking them about wider social impacts.
- The power and perils of the ‘artificial hand’: considering AI through the ideas of Adam Smith: Speech by Gita Gopinath (First Deputy Managing Director, International Monetary Fund), commemorating the 300th anniversary of Adam Smith’s birth (University of Glasgow, 5 June 2023).
Who are experts on this question?
- Jeffrey Funk
- Lucy Hampton
- Lee Vinsel
- Simon Willison
- Daron Acemoglu
- Diane Coyle
