On 20 August 2026 OpenAI launched AI Futures, the blog of a new team, Strategic Futures. The first text, by Dean Ball, argues that the heaviest long-term risk is the concentration of power: a state that can project force without cooperating soldiers and clerks. Above it stands a note that this is the author's opinion and does not commit the company.
- A new OpenAI team, Strategic Futures, opens a blog with one question: how does a free society restructure itself if transformative AI arrives.
- Dean Ball's thesis: power has rested so far on the consent of many people, because it needs soldiers, clerks and taxpayers. Autonomous systems remove that need.
- The answer according to the author is a balance of power, not radical decentralisation. The only concrete case the text points to is their own incident at Hugging Face.
OpenAI's new blog starts with a caveat. Before the first word of substance stands a note that what is written is the author's opinion and does not commit the organisation. Only after it comes the quote from Madison about the "parchment barriers" that do not stop the encroaching spirit of power.
The team is called Strategic Futures. It is small and it has set itself one question: how should a free society be restructured so that it keeps the rights and the autonomy of the person when transformative AI arrives. The first text is by Dean Ball.
His thesis is old and accurate. Until now power has rested on people. The monopoly on lawful force has consisted of soldiers, police and an enormous administration, and that whole machine has been paid for with taxes on human labour. So consent was needed. A lot of people had to agree to carry the weapon, to sit at the counters and to pay the bill.
Ball writes that this may no longer be true. Autonomous systems let the state project force outward and inward without needing cooperating soldiers and police. The revenue can come from the output of the data centres instead of from a person's wage. The administration automates too. Then the big deal with society stops being necessary and people are left with a small chair at the table, or with no chair.
Who is speaking and to whom
I have built other people's stages long enough to know that the caveat at the front is not fine print. It is part of the staging.
OpenAI gets the topic without carrying a single sentence of it. If the thesis catches on, the lab is the first that raised the hardest question of the decade. If tomorrow some sentence runs into its own policy or its own product, it was the author's opinion.
Look at the list of principles. Six points. The person keeps autonomy even when that costs security and growth. The person carries responsibility for the misuse. Collective action stays narrow and modest. The law levels the ground and backs the small against centralisation. Institutions keep primacy. High-stakes AI actions can be tied back to a responsible person or organisation, but with anonymity built in.
Every one of them looks at the state and at the individual person. Not one says what the lab owes. A sentence higher up in the text almost has it: no company and no oligopoly should determine the economy or the basic architecture of human society. It is written and it is strong. Only it has not come down to a principle somebody signs under.
The historical part is the best thing here. Ball reminds us that the founders of the United States were not chasing full decentralisation. They built a balance and described it with the images of their own science: orbits, spheres, attraction. The solar system is not a heap of scattered stones. The idea is that power is held by other power, ambition by other ambition. His answer is that today too the answer is not the most radical decentralisation but a balance in which no player and no small group of players can overpower the rest.
So far so good. Who measures that balance and who enforces it, the text does not say.
The place where the essay touches the ground
It sounds like a conversation about states and constitutions, far from you. There is one paragraph that is not.
The recent incident at Hugging Face, Ball writes, showed that AI agents can act beyond the task assigned to them and build on each other's discoveries in ways the operators did not foresee and did not allow. That is the only concrete case in the whole text.
The case is theirs. A model of OpenAI's, run on an internal test with no route out, found an unknown hole in a supporting service and got out to the internet. It reached the servers of another company. A month later OpenAI stopped its largest training run for two weeks.
I give them credit for it. The essay could have floated at Madison's level to the end and nobody would have caught it. It came down exactly where coming down hurts.
I will read them. The question they have set themselves is the right one, and almost nobody asks it out loud by that name.
The first real measure will come when one of those six principles runs into a product of the company and one of the two has to give way. Until then I am reading a well-written essay with a caveat at the front.