2023-24 Winter Link Review
Another season, another jumble of information I think you should ingest. This season’s theme is discordant. I sense they’re all related, but I haven’t been able to craft a coherent theory with them. I want to share knowledge that undergirds my attitude toward AI’s impact on regulation, labor markets, and culture.
Commercial AI* will be a radical increase in labor supply. Tech leaders like Sam Altman eschew the idea that AI will meaningfully damage labor markets, characterizing the change as taking over tasks instead of jobs, allowing the employed to focus on more abstract problems and let the AIs worry about execution, and allowing tunnel-visioned AI proponents to ignore the fact that most jobs are a set of tasks, most people look for work similar in industries similar to their previous one, and most people don’t perform well in the abstract problem space.
*AI in its current state isn’t a huge bane to labor markets. Much of it has the utility of magic tricks or is chewing away at Google’s search market. Klarna’s customer service bot is making waves in the call center industry, collapsing Teleperformance’s share price by almost 30%. My one-liner for when AI is a threat to labor markets is “when the AIs get hands”. Not in the literal sense, but rather the idea that hands let humans interface with the world. Most white-collar labor (a.k.a. high margin captured by workers) happens on computers. When AIs can interface with the digital world to create Adobe Illustrator paths of potential collateral, deliver Excel workbooks with proper analysis, or find matches and schedule visits for home buyers, to name a few, then the labor markets will begin to see ripples. This is an integration problem actively being solved, re: Devin.
The China Shock: Learning from Labor Market Adjustment to Large Changes in Trade
David H. Autor, David Dorn & Gordon H. Hanson
Autor, Dorn, and Hanson may be the most salient piece of writing when it comes to competition with asymmetric input costs. Competition – between China and the US in this paper’s case – has enduring consequences for the labor force exposed to it. The previous consensus amongst trade economists was
1. Trade had not in recent decades been a major contributor to declining manufacturing employment or rising wage inequality in developed countries;
2. Workers employed in regions specializing in import-competing sectors could readily reallocate to other regions if displaced by trade; and
3. Due to the 'law of one price' for skill, any labor market impacts of trade would be felt by low-skill workers generally, not by trade-exposed workers specifically.
China’s ascension in the global market—read as “introduction of a substantially reduced unit cost of production”—broke these assumptions spectacularly. The cheap labor-supply shock combined with free trade eviscerated American manufacturing, as shown in Figure 4. The more a sector’s labor was composed of production workers, the more it shifted to imports from China.
How this appears to have happened is quite interesting; it revolves around China’s astronomical savings rate. Americans spent, the Chinese saved, and Chinese banks/government loaned those savings out, billions of which ended up as American debt. How this happened was convoluted: higher commercial interest rates due to government banks prioritizing state-run businesses, the one-child policy incentivizing savings for families with boys so they can purchase land, and plain old exchange rate manipulation all appear culpable for the trade disparity.
I strongly recommend you read the paper, it’s comprehensive and legible to lay readers with the exception of the theory section. That said, here are my bullet points:
Specialized and protected industries suffered the most when exposed to free trade with China. Not only would those industries shrivel, but the local industries supporting them would as well. These could be anything from the service industry to upstream suppliers of end-product manufacturers. Competition exposure (measured as $1000 increase in competitive imports per exposed-industry worker) created a 1:1 decrease in a region’s general employment and labor market participation. These regions would experience an employment reduction of 2.4M workers between 1999 and 2011, and there’s little evidence the reduction was offset by employment gains made in the region’s other industries.
AI relationship: Not only will people lose in labor markets that AI is introduced to, industries supporting them will suffer from the decrease in available capital.
If a trade shock is accompanied by a trade deficit, the labor reallocation from exposed tradables into non-exposed tradables is drastically incumbered, but still predicted.
AI relationship: Unless tech players like Sam Altman spend this new surplus insteading of loaning or investing (which typically goes to the healthy or nascent industries anyways), there’s going to be a major deficit between domestic economic classes that will be regionally visible.
Commuter-zones with high shares of exposed-industries saw general wage decline, especially amongst non-industry workers. The fallout of comptetition is modeled in income reductions per $1000 of exposure in a trade-exposed CZ. Unemployment would reduce annual earnings per working-age adult by $312 per year and local wage competition takes an extra $213. These CZ’s were experiencing $525 salary losses for every $1000 dollars of exposure.
AI relationship: A region’s exposed industries don’t shrink quietly. They typically damaging the local economy as it goes down.
Wealth transfer payments were grossly insufficient to offset a CZ’s income loss. CZs at the 75th percentile of exposure lost, compared to the 25th, about $549 per year while transfer income amounted to $58 (all per capita and per $1000 of import exposure). The cost of competition is burdened locally, not nationally.
AI relationship: American taxes and revenues are extremely high as they are. To completely offset regional losses due to competition with AI, government spending would have to be 10x what TAA (Trade Adjustment Assistance) had for softening Chinese competition, and that assumes income loss is linearly related to competition exposure (it isn’t, re: CZ cross-industry impact).
Workers of exposed firms don’t switch to unexposed firms, typically bouncing between employers in an exposed industry. Over a decade and a half they slowly gravitate towards less exposed work, but the cohort’s 1991 firm exposure continued to correlate strongly to their 2007 firm’s exposure. Those workers tended to claim disability and unemployment at far higher rates than counterparts in unexposed firms.
AI relationship: People don’t change their core competencies easily. If they’re trained in skills that are assumed by AIs, they’re likely to continue pursuing jobs with those skills even if they can’t secure work with them. Instead, they found themselves on government payrolls.
The economic blowback of Chinese trade wasn’t proportional amongst the laborforce. Exposed workers in the upper tercile, while experiencing some loss relative to unexposed counterparts, recovered well and typically relocated outside of manufacturing. Those in the lower tercile lost far more than unexposed ones and tended to continue in manufacturing.
AI relationship: Accountants, lawyers, engineers, and similar roles at manufacturers could weather the China shock because those skills more easily translate to other industries than line workers’ . When the demand for heads in accounting collapses in lockstep with the demand in many other cognitively-demanding roles, it isn’t clear what avenues will be wide enough to fit all of these people.
The iron law of oligarchy
Wikipedia
It’s a Wikipedia page, so it’s a summary in and of itself. I’ll still provide a few words and relate the idea to AI.
In many ways, this is an application of the principal-agent problem to sociology. All organizations, whether liberal democracy or Soviet communism, will drift towards oligarchy. The scale and complexity of growing organizations demand consolidation of decision-making (power), a professional class emerges, and humans fundamentally serve themselves, so the organization's decision-making tends to be towards itself and disproportionately towards its leaders.
There’s a lot of talk about AI regulation, and it’s a salient topic that may determine the future, so I will comment here.
Regulating tools is different from regulating outcomes. Killing a human with pre-meditated intent, whether with a rock or rifle, is illegal and punished equally. Even impersonation can be criminal in some cases. Regulating a technology – who can create, distribute, wield, and how – is its own world. Most of the things we don’t want AI to do are already illegal, but lots of the push for regulation is of that vein (EU AI Act Title 2, Article 5, Biden Executive Order). Much of the case for this kind of regulation is reliant on the notion that AI is like nukes and ought to be treated as such.
In a world where which AIs are approved by government red teams, who can use it set by government accreditation levels, or how it can be used policed by three-letter agency policies, the trend will be away from consumer benefit and towards government and incumbent.
Culture over Policy: The birth rate decline
Ruxandra Teslo
A new addition to my reading list, Ruxandra takes a position on biology and culture versus material circumstance. The social world is often much harder to sway than the material one, but sometimes the material world isn’t enough of a factor in decisions to change things. The epitome of this, and the issue she is the steady descent of global birth rates. Many policies across many countries, from maternity leave to tax breaks, have been applied to attenuate to this problem, but effects, if any, disappear with time. One of my favorites isn’t a policy, but a commercial to encourage Danes to take a vacation and have children: Do It For Denmark!
The central thesis of Ruxandra’s piece is that trying to social engineer the subtleties of human life doesn’t work. She compares it to the efficacy of central planning in economics over free markets. Her homeland of Romania is a centerpiece in how force was able to achieve the desired metric, only to be circumvented (often grotesquely) and backfire, eventually reverting to the “cultural mean” and a damaged social environment once the state lacked the capacity to enforce such policy. I would compare it to businesses that over-fixate on KPIs, only to realize it was at the expense of other variables that are revealed to be valuable when the house of cards begins to crumble.
She analyzes the trends and postulate the factors of fertility decline that I will summarize below:
The trend seems very much related to religiosity, not industrialization. This is found in 18th century France, before their industrialization and revolution, and strongly correlated with the transition of local loyalty to the French state from the Catholic Church. This was modeled by looking at the genuflection of some clergy to the King in 1791 and looking at the start of birth rate declines in their diosces.
The historical decrease wasn’t dominated by fewer couples having children, but rather average family size reduced. This can also be found using the French data: the likelihood of a couple having more than 6 kids increases by 10 points if you increasing the number of “refractory clergy” (clergy that rejected the King’s demands for oaths) to 100% of the diosces.
In the US, the decrease in the size of families was mitigated by an increase in the number of women having children. Childlessness at 44 hovered between 20-25% until the generation born around WW1 when it decreased to almost 10% for the generation born during the Great Depression. Women born after WWII have childlessness rates between 15 and 20%, and this is forecasted to increase significantly.
Parents of recent generations say they want their children to prioritize financial and intellectual achievement (“Extremely important”: 88%, 88%) more than marital or parental status (“Extremely important”: 20%, 21%). This contrasts with their own experience who rate time spent with children as significantly more meaningful than paid work.
Rusandra deftly describes much of the attitude of our generation, a characterization that I think is missing parts but is certainly not wrong.
Where is AI in all of this?
The fourth point – younger generations and their parents almost fantacially prioritize financial stability, intellectual achievement, and social status broadly before considering a family – is a recipe for disaster. An abrupt labor shock for this generation would be nothing short of bad. If lots of upper-mid status work gets automated/hyper efficient without a compensatory increase in demand or new demand for equally reputable roles, then most younger folk will be faced with career and financial achievement ceilings. We’re generally unwilling to settle down until they have found someone who also proved themselves in the socioeconomic arena, so I expect my generation’s fertility to be much worse than previous ones.
I haven’t come to a conclusion on what a country full of people with little stake or interest in the future looks like, but I can’t imagine it’s good.