Did you see Microsoft’s announcement about the integration of the GPT large language model into Office 365 today? It was stunning. When Satya Nadella and his team spoke about their new co-pilot they didn’t speak of it as a product launch, but as the future of work. Based on what they showed (corporate affinity for hyperbole notwithstanding) I’m honestly inclined to believe them. Flicking through the tools they announced, I saw that so much of what I do for a living (fingers crossed not all of it) will be accelerated or replaced.
Large language models (LLMs) are a type of artificial intelligence that can process and understand natural language. Trained on massive amounts of text data to learn and recognise patterns in language use, these models can generate text, summarise, translate, and answer questions.
LLMs have been growing rapidly in size and complexity over the last few years, reaching billions or even trillions of parameters, but since the launch of OpenAI’s ChatGPT, they have exploded through the zeitgeist, culminating in what must be one of the busiest weeks for AI in recent memory.
Even though I’ve been incredibly excited for the future as I watch all these announcements come through, I can’t quite shake the pit in my stomach that tells me the world is about to move quite rapidly from beneath my feet. That’s scary! So, for this post I’m going to take a deep breath and look at what economic principles can tell us about what’s coming.
If this is a hammer…
Before we know who is going to be out of a job, we need to know what the tools can do. While I am by no means an expert in this field, I see the following as the key areas of disruption:
Drafting text, documents and presentations (LLMs); and generating images (DALL-E)
Writing code
Analysing and synthesising data and documents (LLMs).
These are some very broad fields! drafting text, documents and presentations, writing code, and analysing data probably describes what 95% of every white-collar worker does. Those dot points probably capture a large amount of creative work too.
The way I have personally used LLMs to date is through the drafting of entire documents ex machina, before commencing editing. That’s actually how I’ve written this blog, using Microsoft Edge’s co-pilot tool. LLMs can help automate or augment this process by generating high-quality text that is coherent, relevant and accurate, including for reports, policies, proposals, contracts, emails and more.
With the integration into Microsoft’s ubiquitous Office and systems, documents and data stored in word docs, powerpoints and excel spreadsheets will be readable by the LLMs. This means that the models can analyse and even synthesise content based on the information you have saved, as well as what’s available on the web. They can draw findings and, with the right prompting, even make recommendations.
The kind of tasks LLMs will augment occur in basically every sector of the economy. For example:
Lawyers can use LLMs to draft legal documents such as contracts or briefs
Journalists can use LLMs to write articles or headlines based on data or sources
Financial analysts can use LLMs to generate models of data and analyse them
Marketers can use LLMs to create catchy slogans or copy for ads
Researchers can use LLMs to write abstracts or papers based on their findings
Managers can use LLMs to write memos or emails based on their goals
Software developers can use LLMs to write more software and catch bugs sooner.
… Am I the nail?
Economists describe something that comes along that can increase the level of output with the same or fewer inputs as a 'positive productivity shock'.12 The empirical literature on the impact of such shocks is mixed, but there are some first-principals frameworks that can help us understand. In particular, to understand the impacts of such a shock on an industry or sector, there are a few key characteristics of that industry and the market that we need to consider before we can begin to prognosticate.
Traditionally, one such characteristic is whether the sector produces tradeable (can be traded across borders, i.e. cars, iron ore) or non-tradeable goods or services (can’t really be traded over borders, i.e. childcare, haircuts).
For example, if Australia found a way to make cars very cheaply, it could sell more cars globally and increase its market share which could lead to more jobs in the car industry. However, if we invented a robot that could do half the work of a childcare worker, then there would be fewer jobs in childcare because people would not need as many workers to take care of their children.3
I work in a non-tradeable sector, usually writing reports of various kinds for Government. These reports have diminishing marginal returns, such that I suspect enhanced productivity in my sector will lead to a slimming. Funnily enough, the first-order impact of this productivity change would show as a cut to GDP as Government spent less.
What about the rest of you?
I assure, you it’s not just me! Most of the jobs I described above will experience a similar impact - are businesses going to demand an increase in marketing output to offset the decreased labour requirements? How about financial analysts? Journalists? My guess on these is probably not.4
Interestingly enough even those producing tradeable goods, such as software goods and services, may also be in for a rude awakening. The tradability of the goods may not actually matter much when every country is experiencing the same productivity shock. It may just be that, unlike government reports, there may be more demand for additional or new goods that require software skills.
It’s all about the pie.
Up until now, I have been complaining to you that my services will no longer be needed in the economy of the future. You may feel the same. But the important part is that we don’t equivocate on this point: getting more stuff for less effort is a good thing. If the output of the economy is the size of the pie, then the cost of your labour tells you the share of that pie that you’re entitled to. Improvements to productivity mean that there’ll be more pie to go around, but in this case it means that less of it will go to people doing jobs like mine.
Where this can become a political and economic issue in the medium term is where it leads to structural unemployment, like where factories closed, and the workers could not be employed elsewhere.5
In the long run, we should consider these sorts of changes in terms of the labour that is freed up for other endeavours. The transition away from farming did not lead to a life of lounging around under palm trees, it led to people finding new ways to provide value for one another by offering new goods and services, or by increasing the volume of other goods and services that people could afford. For example, how many massage therapists or researchers were there before farm jobs shrunk? Probably not many. But now there's room in the economy for these types of endeavours. In the future, this might mean more research and development, or it might mean that we get to enjoy more services like healthcare, therapy, or childcare. At least I certainly hope so!
A negative productivity shock is whenever the effects of caffeine start to fade away in the late afternoon.
For the purposes of this blog, I am using productivity as a byword for labour productivity. Just know that a change that required fewer non-labour inputs (such as energy, raw materials) would likely result in a different impact. As AI seems to mostly address services, I will focus on labour productivity.
These examples make assumptions about the demand response of the market to price shocks. With different assumptions about the demand curves and level demanded, these outcomes could be reversed.
Lawyers would be an interesting case, as I suspect the cost of litigation presents a significant barrier to a lot of court and other legal actions. Additionally, where occupational licencing exists as an artificial constraint on the supply of legal or other services, enhanced productivity may have a smaller effect on the level of labour demanded.
It’s important to stress that if a worker had their job automated and went on unemployment benefits for the rest of their life, the economy would still be no worse off than if the job hadn’t been automated in the first place.