A summary of my thinking on shared prosperity, work and AI in ten bullet points.
This is partly motivated by the fact that I have received questions from several people on these issues, and I feel like it may be useful to lay out my thinking in simple terms in one place. We are also about to have a new administration in the United States, so perhaps it’s a good time to think about some aspirations (even though I view it very unlikely that the incoming administration will move us in this direction).
1. Shared prosperity is key.
By shared prosperity I mean economic growth from which most groups (e.g., men vs. women, different ethnic groups, different education groups, different regions, etc.) benefit more or less in the same way (e.g., their incomes growing at similar rates). Economic growth that just enriches one group greatly and generates only small benefits for many other groups is not shared prosperity. This is mostly an ethical precept, but it can also be justified because a peaceful, harmonious society does require shared prosperity. It is also a realistic one. It does not require that all inequalities are wiped away at one fell swoop.
2. Shared prosperity cannot be achieved just with redistribution.
It needs to be rooted in the labor market, in (good) jobs and in wage growth. The safety net and some amount of redistribution are important. But these are not sufficient to generate shared prosperity. Even in social democratic Nordic countries, where redistribution is most robust, it is not the source of shared prosperity. Wage and employment growth have been much more important historically. Redistribution-based shared prosperity doesn’t make political economic sense either: if some portion of the population is continuously impoverished, they wouldn’t have the political power to ensure that robust redistribution remains.
Moreover, even if we had a system where pre-tax inequality was growing a lot but there was enough redistribution to ensure the disposable incomes of all demographic groups grew robustly, it would have other serious problems. People without jobs and those whose pre-tax incomes were not growing wouldn’t feel that they were contributing to society. Worse, we would head towards a truly two-tier society with just some fraction of the population flourishing economically and receiving all the social status as they are the source of all earnings and tax revenues out of which others are receiving redistribution.
3. AI is here to stay and will be very impactful.
I have little doubt that AI will be a defining technology for our future. It can also deliver significant productivity benefits, though I think whether it will do so or not is contingent on how we develop it, and its full effects will take a while to be materialized. There is a lot of uncertainty about AI’s effects. In my opinion, it is also difficult to know what AGI (artificial general intelligence) would mean and when it may arrive, and this adds to the uncertainty about AI.
In sum, we cannot think of the future of work and shared prosperity without understanding AI’s impact.
4. AI’s direction can be pro-worker or anti-worker.
A basic pillar of my thinking and my research is that all technologies are malleable – meaning that they can be developed in many different ways, with very different consequences about who wins and who loses. This is doubly and triply true for AI, which is a broad, flexible technological platform. AI can be developed for prediction tasks; it could be developed for generating text and images; it can be used as an informational tool, etc. In all of these cases, AI can be more anti-worker (meaning that it focuses on automating tasks and disempowering workers) or pro-worker (meaning that it can become an information technology for enabling workers to perform their tasks better and to be able to branch into more sophisticated and new tasks). How AI will be developed is a choice.
5. Currently it is being developed as an anti-worker technology.
The main way in which companies are thinking of monetizing AI is by automation and more powerful digital ads, and neither of which would contribute to a pro-worker agenda. Moreover, the way in which foundation models are developed and trained is shaped by the expectation and desire to reach AGI. But AGI would mean more automation – if AI can achieve general intelligence and perform almost all tasks as well as most humans, then it will take away these tasks from humans. This current path will therefore lead to job displacement and lower wages, and is thus inconsistent with shared prosperity.
6. To redirect it, you need policies.
Putting the previous two points together, we can conclude that while there was a direction for AI consistent with shared prosperity, we are not pursuing it. Moreover, the industry will not suddenly change direction. Therefore, there needs to be an intervention, and this can only come from government policies (across the world) to encourage new directions and also put regulations to prevent the more harmful uses of AI (some of which are synergistic with the anti-worker direction).
7. To redirect it, you need competition.
New technologies especially radically new directions typically come from new companies, not established incumbents. This is doubly so when the incumbents we are talking about are the largest corporations humanity has ever seen. Hence, the pro-worker AI agenda should be symbiotic with agenda of increasing competition and breaking the hold of the existing powerful incumbents on the tech sector and the direction of AI.
8. To redirect it, you need different architectural choices.
Perhaps even more controversially, redirecting AI may need architectural choices. To put it simply, pro-worker AI need to be an information tool in the hands of workers. This is impossible unless AI provides reliable, understandable and real-time information to workers in a range of occupations. The current architecture of AI (partly fueled by AGI dreams) is about AI acting autonomously and has also led to a black box structure of AI. Instead, the pro-worker direction AI needs the tools to provide advice to human decision-makers (rather than make autonomous decisions), and the best autonomous decisions are not necessarily the best advice/recommendation/information to workers. Moreover, pro-worker AI needs to be understandable by human decision-makers, which is not possible with current black box structure of foundation models complemented with fine-tuning and other kinds of ex post training of pre-trained models.
Stepping back, in an ideal world government intervention should be neutral towards different technological choices. After all, entrepreneurs and innovators know which technologies to develop and how to develop them much better than bureaucrats and lawmakers. But in certain situations where different directions of technologies have major social consequences (for example, in the choice of fossil-fuel versus green technologies), then government intervention may need to impact technology and design choices as well. Nevertheless, it is important that this is done in the most minimalist possible way, so that innovation incentives and choices are not impacted beyond the extent necessary for a more socially beneficial direction to emerge.
9. All of this requires democracy.
Since the current direction is chosen and supported by the largest and most powerful corporations in the world, only robust democratic pressure can lay the foundations of a redirection.
10. The Catch-22: AI endangers democracy.
Tech choices in the past, especially those surrounding social media, have been damaging to democracy and active political participation of the citizenry. The same is likely to be true for AI, and even more so. First, AI is likely to be a very powerful technology for manipulation, and this can exacerbate platform choices that can make money while discouraging democratic citizenship. Second, the current ethos in the AI sector is quite anti-democratic, with leading technologists and entrepreneurs believing that experts (themselves) should be empowered to make all key decisions and democratic processes get in the way of the necessary AI acceleration.
This not only creates a Catch-22 (we need democracy to redirect AI, but AI has already damaged democracies) it also suggests that redirecting AI will be very difficult. But I still believe it’s not completely hopeless.