Inside Story

Machine questions

What does history tell us about automation’s impact on jobs and inequality?

Julian Thomas Books 3 October 2023 1626 words

Conflicts of our time: US Federal Trade Commission head Lina Khan testifying in Congress about her anti-monopoly actions against Google and other technology companies. Michael Brochstein/ZUMA Press Wire

When it appeared twenty-five years ago, Google’s search engine wasn’t the first tool for searching the nascent World Wide Web. But it was simple to use, remarkably fast and cleverly designed to help users find the best sites. Google has gone on, of course, to become many things: a verb we use in everyday language; a profitable advertising business; Maps, YouTube, Android, autonomous vehicles, and DeepMind. Now a global platform with billions of users, it has profoundly changed how we look for information, how we pay for it and what we do with it.

The way we talk about Google has also changed, reflecting a wider reassessment of the costs and benefits of our connected lives. In its earlier days, Google Search was enthusiastically embraced as an ingenious tool that democratised knowledge and saved human labour. Today, Google’s many services are more popular than ever, though Google Search is the subject of a major antitrust case in the United States, and governments around the world want to regulate digital services and AI.

In Power and Progress, Daron Acemoglu and Simon Johnson take the project of critical reappraisal further. Their survey of the thousand-year entanglement of technology and power is a tour de force, sketching technology’s political economy across a broad historical canvas. They chart the causes and symptoms of our contemporary digital malaise, drawing on a growing volume of journalism and scholarship, political economy’s long tradition of analysing “the machine question,” and the work of extraordinary earlier American technologists, notably the cyberneticist Norbert Wiener, the network visionary J.C.R. Licklider, and the engineer Douglas Engelbart.

If, as Acemoglu and Johnson argue, our digital economy is characterised by mass surveillance, increasing inequality and destructive floods of misinformation, then the signal moments from the past will inevitably look different. From this angle, the great significance of Google Search was its integration with online advertising, opening up the path to Facebook and a panoply of greater evils.

The strengths of Power and Progress lie in the connections it makes between the deficiencies of current technology and the longer story of innovation and economic inequality. History offers many opportunities to debunk our nineteenth-century optimism in technology as a solution, and to puncture our overconfidence in the judgement of technology leaders.

A particular target is the idea that successful innovations produce economy-wide benefits by making workers more productive, leading to increased wages and higher living standards generally. The theory fails to capture a good deal of historical experience. The impact of new agricultural technologies during the Middle Ages provides a telling example. Between 1000 and 1300, a series of innovations in water mills, windmills, ploughs and fertiliser roughly doubled yields in England per hectare. But rather than leading to higher incomes for most people, living standards appear to have declined, with increases in taxation and working hours, widespread malnutrition, a series of famines and then the Black Death. Average life expectancy may have declined to just twenty-five years at birth.

The cities grew, but most of the surplus generated by improved agriculture was captured by the church and its extensive hierarchy. A religious building boom proceeded on spectacular lines. Vast amounts were spent on hugely expensive cathedrals and tax-exempt monasteries: the same places, as Acemoglu and Johnson note, that tourists now cherish for their devotion to learning and production of fine beer. The fact that better technology didn’t lead to higher wages reflects the institutional context: a coercive labour market combined with control of the mills enabled landowners to increase working hours, leaving labourers with less time to raise their own crops, and therefore reduced incomes.

If medieval cathedrals give rise to scepticism about the benefits of tech, it follows that we should think more carefully about the kinds of technologies we want. Without that attention, what the authors call “so-so automation” proliferates, reducing employment while creating no great benefit to consumers. The self-checkout systems in our supermarkets today are a case in point: these machines simply shift the work of scanning items from cashiers to customers. Fewer cashiers are employed, but without any productivity gain. The machines frequently fail, requiring frequent human intervention. Food doesn’t get any cheaper.

The issue then is not how or whether any given technology generates economic growth, but which conditions make possible innovations that create shared prosperity. The recent past provides examples of societies managing large-scale technological change reasonably well. The postwar period of sustained high growth and “good jobs” (for some but not all) had three important features: the powers of employers were sometimes matched by unions; the new industrial technologies of mass production automated tasks in ways that also created jobs; and progressive taxation enabled governments to build social security, education and health systems that improved overall living standards.

For technology to work for everyone, the forces that can temper the powers of corporations — effective regulators, labour and consumer organisations, a robust and independent media — play an essential role. The media are especially important in shaping narratives of innovation and technical possibility. Our most visible technology heroes need not always be move-fast-and-break-things entrepreneurs.

Finally, public policy can help redirect innovation efforts away from a focus on automation, data collection and job displacement towards applications that productively expand human skills. Technologies are often malleable: they can frequently be used for many purposes.

Acemoglu and Johnson would like us to divert all that frothy attention on AI to what they call machine usefulness, focused on improving human productivity, giving people better information on which to base decisions, supporting new kinds of work, and enabling the creation of new platforms for cooperation and coordination: a course they see as far preferable to a universal basic income.

Kenya’s famous M-PESA, introduced in 2007, is one of many examples, offering cheap and convenient banking using basic mobile phones. On a larger scale, the web is also a human-oriented technology because its application of hypertext is ultimately a tool for expanding access to information and knowledge. Acemoglu and Johnson concede that the idea at the heart of Google Search can also be understood in this way: a mechanism that works well for humans because it is constantly reconfiguring itself in response to human queries.

The authors’ ideas for positive policy interventions can usefully be read alongside those of the Australian economists Joshua Gans and Andrew Leigh, whose 2019 book Innovation + Equality remains less used than it should be.

One way to read Power and Progress is as a historically informed guidebook for the conflicts of our time — in the courts, where Lina Khan’s Federal Trade Commission has launched far-reaching cases against Google and Amazon, in the new regulatory systems emerging in the European Union, Canada and elsewhere, and in the wave of industrial actions taken by screen industry writers and auto workers in the United States.

In Australia, we are also at a point where governments will soon make decisions about the kinds of technology we want to support or constrain. We can have no certainty about the outcomes of any of this, but Acemoglu and Johnson argue that such conflicts are both necessary and potentially productive. They diverge here from one of the main currents of liberal technology critique: where writers like Carl Benedikt Frey, whose The Technology Trap (2019) covers some of the same terrain, see redistributive policies as necessary for managing the consequences of automation, Acemoglu and Johnson point to the positive potential of political and industrial conflict for reordering technological agendas. They want to place more emphasis on our capacity to choose the directions technology may take.

The recently concluded Hollywood writers’ strike offers an intriguing example. The key point is that the screen writers didn’t oppose the use of generative AIs such as ChatGTP in screenwriting. Instead they secured an agreement that such AIs can’t be recognised as writers and that a studio may not require the use of an AI. If a studio uses an AI to generate a draft script that it then provides to a writer, the credit or payment to the writer will be the same as if the writer had produced the draft entirely themselves; and a writer may use an AI with the permission of the studio without reducing their credit or payment.

The settlement clearly foreshadows the extensive use of generative AIs in the screen industries while offering a share of the benefits to writers. The critical point, as some reports have noted, may be that the revenue-sharing deal with writers preserves the intellectual property interests of the studios, since works created by an AI may not be copyrightable.

Meanwhile, AI raises other important issues about automation, quite apart from the focus on work. When we are relying on machines to make or inform decisions, we are also moving into the domain of institutions, with the obvious risk that existing technology-specific laws, procedures and controls can be bypassed, intentionally or otherwise. This, after all, was what robodebt did with a very simple automated system. In the absence of wide-ranging institutional adaptation and innovation, more complex modes of automation will pose greater risks.

More generally, the authors’ framing of the “AI illusion” appears to be premature. Power and Progress was clearly substantially completed before the appearance of the most recent versions of ChatGPT. Accustomed as we are to AI’s many failures to match its promises, we should now be considering the surprising capabilities and broad implications of large language models. As Acemoglu and Johnson would insist, if generative AI does turn out to be as powerful as many believe, then it will necessarily be capable of far more than “so-so” automation. •

Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity
By Daron Acemoglu and Simon Johnson | Basic Books | $34.99 | 546 pages