AI and Jurisdictional Choice
Does AI do for cognitive labour what containerisation did for manufacturing?
Until 2022, if a company wanted to leave Delaware, they’d have to get unanimous approval from their shareholders, but following a new ruling, only a simple majority was required. Since then, Tesla and SpaceX reincorporated in Texas, Dropbox and Pershing Square Capital are reincorporating in Nevada, and Meta is also reported to be considering a move to Texas. For Delaware, this is bad news: the state gets about a third of its revenue from business franchise taxes, not to mention the secondary benefits of being the place where everyone files, and disproportionately solves their corporate disputes.
Tax, too, follows this pattern: companies have relatively high levels of jurisdictional choice about where to file. For a few years, there were loopholes in Ireland, the Netherlands and some Caribbean islands, though some of these are being closed now through international treaties.
Wherever firms have high levels of jurisdictional choice, there is a race to the bottom among countries and states competing for their business. Companies have relatively more power.
But companies are somewhat constrained by having headquarters, management teams, shareholders, and they need to do business in major jurisdictions. They want to be on the New York Stock Exchange, even if the NYSE itself has the choice to be on a datacentre in New Jersey. In all these senses, they are strung to location.
How does AI change this?
Over time, AI agents will become an increasingly large share of economic output. While we have previously expressed some scepticism, perhaps there will even be firms entirely composed of AIs in some industries. These will have much higher jurisdictional choice about where they operate—a greater fraction of ‘labour’ can hop between datacentres, rather than being stuck in one place because of human practicalities and preferences.
AI for science is illustrative.
When AI systems can develop hypotheses, design experiments, and interpret experimental data at the level of the very best humans, then making scientific progress is no longer bottlenecked by the throughput of the most talented scientists at elite universities. The scientific process can be ‘deskilled’. Humans will still need to implement these experiments, as robotics aren’t good enough to fully automate the process yet.
We could quite quickly develop new tools to support the research assistants to work better. Carl Shulman has suggested, for example, that augmented reality could abstract the requirements for process knowledge:
“[Y]ou could have a worker previously without training and expertise in the area who has a smartphone on a headset, and we have billions of smartphones which have eyes and ears and methods for communication for an AI to be talking to a human and directing them in their physical motions with skill as a a guide and coach that is beyond any human. They could be a lot better at telepresence and remote work and they can provide VR and augmented reality guidance to help people get better at doing the physical motions that they're providing in the construction.”
However, even once the bottleneck of cognitive labour for science is untethered, there could be other forces keeping it tied to where it already happens: process knowledge will still be in current institutions initially, academic institutions have access to specialised equipment, and it is inconvenient to build a new lab elsewhere. So I don’t intend to make a narrow prediction about what science will look like in the near future, but rather a gesture towards the general trend: where AI systems abstract the cognitive labour from some process, or become an increasing share of output, companies will gain greater jurisdictional choice.
This is especially important from a European perspective. There’s going to be a lot of economic growth from AI — but the majority of the growth effects from general-purpose technologies come from the new products and services made possible, rather than adding it into existing processes. One has to ask, why should we expect this new growth to happen in Europe? When companies are going to have greater jurisdictional choice, depend less on specialised cognitive labour; decisions about where to operate will be driven comparatively by the amount of inference compute available in a market, a pro-innovation regulatory approach, and a low cost of electricity. It seems straightforward to imagine European countries finding it more difficult to compete—on regulation, energy, and abundance of inference compute. That said, these things are fixable.