In “The productivity paradox”, I summarised one of the key arguments in The Wealth of Humans: that labour abundance is preventing us from taking full advantage of the digital revolution. Let’s talk a little more about what that means, drawing on the experience of the industrial revolution. Economic historians have not settled every question concerning why the industrial revolution happened when and where it did. We do have a reasonably clear, broad-brush picture of the contributing circumstances.
For the revolution to begin, there needed to be a supportive supply side of innovative potential in place. In Europe, that potential developed with the scientific revolution and what Joel Mokyr calls the “Industrial Enlightenment”: the spreading notion that natural phenomena are governed by rules that can be understood by humans: that can be discovered through experimentation and then exploited. There also needed to be a population of merchants, traders, artisans and engineers with the skills, equipment, and income that would allow them to tinker with machine designs. In northwest Europe this population grew as maritime trade and expanding global commerce supported urbanisation and occupational specialisation within Britain and the Low Countries. Other things no doubt mattered as well, though it is hard to know how decisive they might have been. Were strong property rights relevant? Probably, but it’s hard to know how much so (and they could, in some situations, become an impediment).
At any rate, by the 17th century there was, in northwest Europe, a large, urban, literate, commercially-oriented population with basic technical knowledge. There was also, in parts of the region, an important material abundance, which further boosted innovative potential. The availability of water power was crucial to early mechanisation of textile production. The steam revolution depended on cheap, abundant coal (deposits of which were, in Britain, located in advantageously close proximity to iron ore).
But innovative potential alone was not enough. The demand for inventions also mattered. It was important that firms face pressure to economise on scarce, expensive labour. Robert Allen argues that in Britain, wages were high relative to the cost of other factors of production in comparison with relative labour costs elsewhere in the world. This did not mean that British workers were particularly well off; they weren’t. But in Britain, in comparison with France or China, workers were relatively costly while coal was dirt cheap. The earliest steam engines saved labour but were hugely inefficient consumers of coal. In Paris or China, it would not have paid to have developed the steam engine (and indeed, it did not pay to borrow the technology from Britain for quite a long time, until a century of tweaks to the technology made it far more efficient). In Britain, however, even the earliest, least impressive steam engines found a market.
That was crucial. The existence of a market for steam engines gave developers the opportunity to experiment with tweaks and improvements to the technology, and a competitive incentive to deploy them. Little by little, British engineers developed lighter, more powerful and less coal-hungry engines, which could be used economically in more contexts. As the engines got better, the market for them grew, and the number of people experimenting with the technology increased. Eventually, steam was used in the mechanisation of all sorts of activity, from manufacturing to transport. And it was effective enough to be profitable in places where labour wasn’t especially expensive while coal was. At that point, the rest of northwest Europe quickly began using the most sophisticated steam technology available.
What can we learn from this regarding our own circumstances? Well, consider artificial intelligence. It is certainly easy to imagine a future in which high-quality artificial intelligence is, like the steam engine, a truly general purpose technology, which finds transformative uses in lots of contexts and lots of industries. But how might we get from point A to point B?
Artificial intelligence technology has improved dramatically over the past decade or so. In the 2000s, many experts reckoned that truly driverless vehicles were decades — perhaps a half a century — away. Now, Waymo expects to make truly autonomous vehicles available to select users via a ride-hailing app within a few months. Getting there has not come cheaply. Google has reportedly spent over $1 billion getting its driverless car project to this point. “To this point” is an impressive place, but it remains one in which anyone hoping to deploy autonomous-vehicle technology in a context other than those Google has focused on has a lot more work to do. Joe Ceo, operating a mid-sized manufacturing company which employs a number of forklift and truck drivers, can’t just ring up Waymo and ask for plug-and-play driverless versions of the vehicles he’s currently using.
Rather, to make use of driverless vehicles requires an investment in the equipment itself. It requires investment in the time needed to adjust the software to suit particular circumstances. It means gathering the massive amounts of data needed to train the algorithms. It means spending the money to develop (or maintain access to) the expertise needed to keep the vehicles running. It also requires substantial investments in intangible capital, so that a firm, having made the leap to driverless technology, maximises its potential.
When will it be economical to spend that money? Taking the plunge will look less daunting when the potential market is large, so that the cost of acquiring and adapting the new technology can be spread over many more sales or use cases. It will be more attractive when data demands are low or available data is abundant: because cars will travel routes that are short and simple, or which have already been mapped in extreme detail, for example. Crucially, the technology will be attractive in places where the factor being saved (the labour of the human driver) is expensive. But how many places meet these conditions? In particular: where will the cost of the new vehicle, plus all the engineering, be less than the cost of existing vehicles with human drivers?
This is the crux of the matter. It’s amazing to see videos of Waymo vehicles pulling up to the curb, grabbing some passengers, then motoring on their way. But what’s it gonna cost? Is it cheaper than a human-driven Uber ride? No? Then we may have a problem.
If we suppose that in at least some important contexts it makes economic sense to use driverless vehicles in the wild, then that is a very big deal. Because the more that driverless technology is used in real-world environments, in markets in which there are real competitive pressures, the more opportunity there will be to experiment, the more data the vehicles will gather, the better and cheaper they’ll get, and the more new market opportunities will consequently be opened to providers of the technology. Eventually Joe Ceo will find it worth his while to junk the vehicles he’s got and bring in autonomous ones, which will by then have gotten much cheaper and clever enough to get working immediately without a lot of bespoke customisation of the technology.
That’s the magic point at which the world is utterly changed. But whether and when we get there depends upon the initial competitiveness of the expensive, high-maintenance, limited-use versions of the technology we’ve got now. And that, critically, depends on things like the availability of cheap labour.
We are accustomed to thinking about the supply-side constraints on technological progress. Like: are we coming up with good ideas? Are there enough STEM workers? Is the regulatory environment inhibiting innovation? But the demand for technological progress also matters. Are input prices such that firms stand to make money after going through the long, costly, risky process of developing working prototypes? If wages are low, the answer will often be no. Why would anyone want a robot therapist, when robots are not likely to be as good as available humans? They wouldn’t, given the price of available humans. Triple that price, however, and robots, despite their flaws, begin to look attractive. And once they are actually being used within a market, it becomes much easier to make them better and cheaper still. (Humans might recoil at this possibility and question whether we ought to want a society in which machines displace well-trained, passionate humans from such tasks. The Luddites felt similarly.)
Now, the availability of cheap labour might not forever prevent us from building a world of driverless cars. As it happens, we live in an economy in which artificial intelligence is being pursued aggressively by massive technology companies, which are swimming in money and are relatively shielded from competitive pressures. That is why Google can pour billions into long-term moonshot projects like driverless vehicles. It also means that Google can continue to develop and improve the technology even if the short-term economics are not particularly favourable to it. And maybe, over time, that yields driverless products with attractive cost profiles. There are also other paths through which autonomous vehicles could arrive. In particular, both insurance companies and drivers of personal vehicles have good reasons to want cars to come equipped with ever more driver-assisting technology. But if we’re gaming out when firms will be using driverless vehicles in productivity-enhancing ways, the price of labour matters a lot.
And driverless vehicles are just one case among many. In any circumstance in which there is the promise of automation (using AI or anything else) it’s worth asking: are labour-market conditions such that the risks and expenses associated with developing a workable AI solution for that particular context are worth the trouble? What this generally means is: do labour costs matter enough to the business, and are the potential labour-cost savings big enough, that it makes sense to rethink our business model and give tons of money to Google (or whomever) in order to take a flyer on this new, immature technology? We need there to be one or a few examples in which the answer is yes; otherwise the opportunities and the market pressures that lead to the improvement and perfection of that technology don’t arise.
One last thing to consider is what might happen as new labour-saving technologies find their way into parts of the economy. Suppose, for example, that trucking is eliminated by driverless vehicles. What happens to the truckers? Some might drop out of the labour force. But most will try to find work somewhere else. And in doing so, they will place downward pressure on the wages of similarly skilled workers. That, in turn, will weaken the case for automation in other parts of the economy.
Now, AI might get so good, so fast, and humans might be so miserably unable to adapt, that low wages end up not mattering very much, and mass automation happens anyway. That was the fate of the horse. As the internal combustion engine began displacing horses, horse prices tumbled. The collapse in the price of horses slowed the pace of mechanisation in agriculture, by making it less attractive to farmers to buy a new tractor, but only by a little. Mechanisation continued, and horse prices kept falling until, in most cases, it was no longer worth it to farmers, or anyone else, to keep them. (This was bad news for horses.) On the other hand, it might be the case that the adoption of new automating technologies is limited and halting, because wages are low and advances in one part of the economy reduce the advantage to saving on labour costs in other parts of the economy.
But what is important to remember is that the supply of innovative potential is not the only thing that matters. If we’re asking where the labour-saving technologies are, we need to ask where there is intense pressure to save on labour costs.
If cheap labour is a constraint on innovation and productivity growth, then what do we do? A substantially higher minimum wage would address the problem, but would create others, like high unemployment. A generous universal basic income would increase worker bargaining power and might set a floor on job pay and quality, but it would cost a fortune and is probably better suited to a world in which a great wave of automation has already taken place. More powerful unions might do the trick, by negotiating higher pay and reduced hours (and by generally annoying employers). Governments might also try engineering labour scarcity by prioritising full employment, by investing heavily in infrastructure, by lavishly supporting adult education, and by increasing employment in fields like teaching.
Ultimately, we need to find ways to raise labour compensation and to allow workers to reduce their working time. People spend a lot of time worrying about what we’ll do with ourselves when the machines take all our work. We may have it backward. The problem might instead be that if we can’t find things to do with ourselves, the machines won’t take our jobs in the first place.