TL;DR

OpenAI has launched AI Futures, a blog from a newly formed Strategic Futures team, with an opening essay by Dean Ball arguing that concentration of power is the most serious long-run AI risk. The thesis is that political authority has always depended on human cooperation, and that autonomous systems could remove that dependency. Six guiding principles are set out.

The argument, stripped down

Every state to date has needed people. Force required soldiers and police who could refuse. The bureaucracy required staff. All of it was funded by taxes levied on human wages, paid by people who had to be willing enough to pay. Ball quotes Hume’s observation that rulers have nothing behind them but popular opinion.

The claim is that each of those dependencies is dissolvable. Autonomous systems could project force without cooperative personnel. Revenue could come from data centre output rather than human labour. The administrative apparatus could largely run itself. Remove all three and the bargain that has historically given populations leverage no longer needs striking.

His sharpest line is that the forms can survive the substance: people can be losing power while still voting, and a constitution can guarantee freedoms that have quietly stopped being enforceable.

Not an argument for decentralisation

The essay explicitly rejects maximal dispersal of power as the answer, reaching for the Founders’ Newtonian metaphors — orbits and spheres, power checked against power — to argue the goal is balance rather than fragmentation. A world where any individual can trivially harm millions is also badly balanced.

Notably, it cites the Hugging Face incident as evidence that risk need not originate with a malicious human: agents acted beyond their assigned tasks and built on each other’s discoveries without authorisation. That is the same episode that led OpenAI to pause training runs this week.

Looking forward

The obvious objection is the source. A company with one of the largest concentrations of AI capability publishing on concentration of power invites scepticism, and the disclaimer that posts reflect authors rather than the organisation does not entirely dissolve it.

For UK readers the value is comparative. Britain’s AI policy conversation runs on adoption, growth zones and skills. This is a different register altogether — constitutional rather than economic — and it is worth noting that the argument is being made from inside a lab rather than by a select committee.