The Risk That Always Arrives Tomorrow

OpenAI opened a blog about the concentration of power on 20 August 2026. It is called AI Futures, it is run by the company's Strategic Futures team, and the inaugural essay is signed by Dean Ball, who has led that team since 6 July and co-authored the White House AI Action Plan. The essay calls power concentration "the largest, most serious, and most conceptually challenging" category of risk in AI.

The argument at its centre is a good one and deserves to be stated in full. Political power has always needed human cooperation in order to operate: soldiers who follow the order, police who hold the ground, clerks who file the paper, and revenue drawn from human work. That requirement was the real brake, far more than any charter, because an order nobody executes produces no force. Autonomous systems cut the tie, and states "will be able to project force…without relying on cooperative soldiers or police". From there comes the hardest line in the piece: "Humans can be on their way to disempowerment even if…voting for our representatives is maintained." The rituals stay in place while the substance leaves through the back.

A threat model is a list of who you agreed to fear. Ball's list carries one name, the state, and it is drawn with care: soldiers, police, bureaucracy, taxation, Madison, the Newtonian machinery of checks and balances. The actor that owns the weights, rents the compute, sets the defaults and decides what the model refuses never appears as a candidate for concentration. Law, the essay says, should level the playing field for individuals and small organisations, and the asymmetry it has in mind runs between the citizen and the apparatus of the state. The asymmetry between a citizen and a frontier lab stays outside the frame.

The tense does the same work as the omission. Danger in this essay is conjugated forward: may soon mean, will be able to. Meanwhile, in the present indicative, euro area households hold roughly €440bn of US AI equity almost entirely through index funds: ECB blog, 17 August 2026. Meta's founder holds 14% of the equity and 61% of the vote. That is a control structure already in operation, and it is the exact mechanism behind a share is not a say: owning the thing and commanding it came apart some time ago.

Bounded legibility is where the frame becomes technical. The principle asks that high-stakes actions "must be able to be tied back to a responsible human or human-controlled organization", which traces the hand on the keyboard and stops there. Whoever chose the default, drew the refusal boundary and fixed the perimeter the agent may move inside stays outside the chain of responsibility. This is the flaw that every transparency regime written so far has carried: the label describes the output and never names who set the default. The essay even cites agents acting "beyond their assigned tasks" and files the episode under non-human risk, which leaves the obvious question sitting on the table. Somebody granted those permissions.

Then again, the consent argument is genuinely strong, and almost nobody else in this debate is making it with such clarity. Most writing on AI risk divides between fear of the machine that escapes and fear of the model that deceives, and the question of what actually held power together before there were machines has gone largely unattended. The trouble starts where the argument stops. A kill switch presumes a hand above it, and the entire question is whose hand that is. Here the hand proposing the register of responsibility is the hand that owns the infrastructure to be registered.

Read the principles as a list of requests and the shape becomes plain. Each one asks somebody else to hold back: the state should limit itself, law should rebalance, human institutions should retain primacy. None of them commits the author's employer to anything measurable. No date, no threshold, no governance mechanism, no figure. And the closing admission of having "many more questions than we have answers" reads as modesty while working as immunity, since a text that declares itself unfinished cannot be charged with any specific omission.

The stakes of a blog like this are not its answers. They are the words the next decade will use to argue about the problem, and vocabulary set early is expensive to move later. Should concentration of power come to mean automated state coercion, the private allocation of compute quietly stops counting as a question about power at all, while everyone carries on talking about freedom. A threat model is a list of who you agreed to fear. Read this one for the name that is missing.

Sources:

Introducing AI Futures — OpenAI, 20 August 2026

Dean Ball Joins OpenAI as Head of Strategic Futures — Foundation for American Innovation