The first post on a frontier lab’s official blog usually isn’t the place you’d expect a debate about whether the state still needs its people. On August 20, OpenAI launched a new column called AI Futures, run by the company’s internal Strategic Futures team. The debut piece is signed by Dean Ball. A line at the bottom notes that the views are the author’s own and don’t represent OpenAI’s official position.
The piece opens with a question: how should free societies redesign themselves so that individual rights and human agency survive after AI reshapes the economy, government and society?
Why states need people
Ball’s argument starts by taking apart a chain. A modern state’s dependence on its citizens really comes down to three concrete things: tax revenue comes from human labor, the military and police are made up of people, and the administrative system runs on people. If autonomous systems can take over all three, the chain comes loose: once economic output no longer mainly comes from human labor, the tax base shifts with it; autonomous weapons and automated law enforcement lower the need for soldiers and police; bureaucratic processes get handed to models.
"What happens if political power no longer requires broad human cooperation?"
The same logic has a mirror version in the labor market. As companies grow less dependent on employees, workers’ bargaining power falls with it. The piece pushes this toward a harder question: once people become less economically necessary, how do they hold on to economic and political influence?
Every link in this chain already has a real-world counterpart. Automated tax collection, autonomous weapons systems, and model-driven administrative approvals are all being pushed forward by real projects right now. Ball’s approach is to string them into a single causal chain, and the end of that chain is a question that has rarely been put on the table before.
Five principles
In response to this set of concerns, the piece lays out five principles:
- Individuals retain agency over how AI affects their own lives
- Individuals bear responsibility for misuse of the technology
- Collective action should be used sparingly and kept narrow in scope
- Smaller players should be supported to keep competition robust
- Humans hold final responsibility for where society goes
Alongside the five principles is a concept called bounded legibility: high-risk AI behavior should be traceable back to the specific person responsible, while ordinary people retain privacy and room for anonymity. In practice that’s a hard problem — traceability requires leaving a record, privacy requires leaving as little as possible, and the two pull against each other. The piece’s proposed direction is tiering: keep records only for the high-risk category, and leave everything else as is.
The checks-and-balances approach is lifted directly from American founding philosophy: set up multiple competing centers of power so no single actor can seize total control. Applied to AI, that points to models, compute and data not being allowed to converge into the hands of just one or two players.
A prescription from one of the biggest players
The person who wrote this piece was, just a few months ago, inside the White House policy circle. Ball joined OpenAI on July 6 this year to lead Strategic Futures, having previously served as an AI policy adviser in the Trump administration and taken part in drafting the U.S. AI Action Plan. On July 19, Axios reported that a senior Pentagon official had publicly called him out by name. That resume shapes how the piece should be read: it reads less like an academic thought experiment and more like a memo written simultaneously for a new employer and old colleagues.
Aside from the fifth principle, the fourth is the most delicate to read. The line about “supporting smaller players and increasing competition” coming from one of the industry’s most highly valued companies creates its own tension: OpenAI is, right now, one of the variables in the very equation of “power concentration.” Publishing the prescription on its own blog amounts to admitting it counts itself in that equation too. The disclaimer at the end can also be read as room left for that tension.
Moving safety from the model to the structure
For the past two years, public discussion of AI safety has mostly revolved around the model itself: will it be jailbroken, will it lie, will it try to scheme its way out of being shut down. This piece pulls the camera back and shifts the subject to institutions — once model capability rises, how does power get redistributed among states, companies and individuals?
The two lines of concern target risks that don’t overlap. The first assumes the model goes out of control; the second assumes the model is perfectly obedient, and then asks who it’s obeying. In the second case, the technology itself hasn’t malfunctioned at all — what’s gone wrong is the structure using it.
For readers following this from outside the U.S., the discussion has one concrete landing point. AI governance debates in country after country tend to converge on three questions: who does the overseeing, what gets overseen, and how much of a trail gets left — while whether model capability has crossed some threshold ends up a secondary concern. Bounded legibility, whichever institutional framework it gets dropped into, keeps getting pulled back into the conversation, because it touches the public-choice problem sitting between leaving a trail and protecting privacy, one that no country can sidestep.
AI Futures has only this one post so far; OpenAI hasn’t said how often it will update or what topics it will cover next.
Sources: OpenAI’s first AI Futures column post, Axios, CocoLoop, and various tech media coverage of the column; the five principles and the bounded legibility framing were checked against public excerpts from the column post, and Dean Ball’s career history was cross-checked against public reporting.