It is difficult to talk about the future of artificial intelligence today without touching on almost everything else. About work and the distribution of wealth, but also about power, the relationship between the United States and China, the potential role of Europe, or how far we will be able to control increasingly autonomous systems. The technology is advancing quickly and, with it, so are the questions it leaves open.
The same week that OpenAI launches its latest model, GPT6-Astra, we speak with the company’s Head of Strategic Futures, Dean Ball. His job, as he explains it, is to envision the scenarios AI could open up and to consider which decisions we make today might bring us closer to some of those scenarios and farther away from others. In October, he will also be one of the speakers at Foro La Toja 2026, which Agenda Pública will once again cover from Illa da Toxa.
Ball concentrates on the concentration of power, argues that governments should have mechanisms to slow AI development if circumstances require it, and rejects one of the metaphors that has most shaped the technological debate: the notion of a “race.” Speaking as a private citizen, he goes even further: “An arms race here benefits nobody,” he reiterates. He also has a particularly direct message for this side of the Atlantic: “Europe has to build data centers and stop fretting about it”. For Ball, the continent still has room to leverage its industrial capacity, but it must understand that all of this is unfolding with far more urgency than it may appreciate.
Your current title at OpenAI is Head of Strategic Futures. So in the context of everything happening in AI, what is your view of the different futures ahead of us? What are the most probable developments as the technology accelerates?
I think one of the interesting things about living in this period is that, for the first time in a very long time, it is genuinely difficult to imagine what the future will look like.
I just had a child. He’s eight months old. I look at him and find it very hard to imagine what his life is likely to be like when he’s 20.
I don’t think that was true when I was a baby, in 1992. I don’t think my parents looked ahead to 2012 and thought: “That is unfathomably different from today. I can’t even imagine what 2012 is going to be like.” I’m sure they thought it would be different, but not that the entire nature of the human role in the economy would be almost impossible to predict. There is tremendous uncertainty around that today.
So I basically think the job of our team is to envision the future as a distribution of many possible outcomes. Imagine not one or two outcomes, but millions of potential outcomes, and then ask: as we navigate between them, which are the better ones? Which are the ones we want to avoid? Which are the ones where we’re not sure?
“The job of our team is to envision the future as a distribution of many possible outcomes”
What might be desirable is itself very complicated. It is not necessarily our job to tell the world what the desirable future is. But what is probably incumbent on us is to explain what we think the realistic range of possible outcomes could be, which we think is quite broad, and then suggest what levers we have at our disposal today to steer the world toward the futures we would prefer to live in rather than the ones we would not.
One of the areas creating the most uncertainty is the future of work. It affects people directly, not only through their income but also through their role in the world and their sense of purpose. There seem to be two broad camps: one is more pessimistic and expects AI to disrupt and eliminate many jobs; the other speaks of abundance and overcoming scarcity. Where do you place yourself? And what happens on the demand side? If people have no disposable income, who will effectively be paying for that abundance?
All of these things depend to some extent on economic outcomes, some of which are simply going to be what they are. There are going to be areas where AI is likely to outcompete human labor. But I think a lot of this is going to be downstream of what our institutions are. What do our institutions incentivize?
You can imagine many jobs where perhaps AI can do the task and perhaps it can’t, but either way human beings simply prefer interfacing with another human being. Or perhaps it becomes a luxury to interact with another human being. It might even become a status symbol to have more humans around you because they’re very expensive.
But that requires the preferences of human beings to continue mattering in the world. That’s the important thing.
What you need is to ensure that institutions incentivize a world where human preferences still matter. And I think the single best way we know to do that is to ensure that power is distributed throughout the economy and throughout society: political power and economic power. You need to make sure there aren’t massive concentrations of power that make individual preferences matter very little. That’s the main thing I think you have to avoid.
There are many other questions here, but if I had to identify the primary issue, I think it really is the concentration of power and how you avoid it.
“You need to make sure there aren’t massive concentrations of power that make individual preferences matter very little”
There are plenty of ways to do that. The answer might ultimately involve some form of wealth redistribution. It might involve different approaches to liability. It might involve questions about who is allowed to own property. Can AIs own property themselves? Can they transact property themselves? These are really interesting questions.
You discuss this power balance in your last Substack post, arguing that AIs may inevitably themselves become self-sovereign economic actors and outlining what we could collectively do in response.
We should ensure that certain things do not happen. We should have explicit provisions in the law against AI ownership of property, for example.
What I’m saying in that post is that the existence of AIs that are independent economic actors is probably inevitable at this point, but there are all sorts of different ways we can constrain their access to real property in the economy: limit their ability to own real estate and limit their ability to transform the physical world on their own. And I’m saying we absolutely should do those things.
“The existence of AIs that are independent economic actors is probably inevitable at this point, but there are all sorts of different ways in which we can constrain their access to real property”
That post is a good example of what I was talking about before. Certain phenomena may be inevitable, but that does not mean we have no agency. It means the phenomenon may be inevitable while we can still exercise agency through the choices we make about law, public policy and, ultimately, the institutions those choices create.
Given the Hugging Face incident, we’re beginning to see political responses moving toward slowing or pausing AI development. Senator Bernie Sanders, for example, has just called for an immediate AI pause, and Jean-Luc Mélenchon echoed that position in France. We’re also seeing political opposition to things like data center infrastructures. What is your view of that political response? Is slowing down AI a useful and viable path forward?
I haven’t studied Senator Sanders’s proposal in any serious depth, so I’m not qualified to comment on it. It appeared just on September 3, and we had a couple of things happening at OpenAI that day [the company released GPT-6 Astra, its new frontier AI model].
But something I am much more familiar with is Pacing the Frontier, a letter that came out a couple of weeks ago and was signed by a number of frontier lab employees, myself included.
In that letter, we say that we believe it should be a serious priority for governments to figure out exactly how we could pace AI development so that, if we wanted to slow the rate at which AI is accelerating, we could do that.
This is a really delicate area of the law. You have to be very careful because there are ways of doing this that could turn into a hugely tyrannical power grab by the government over a technology that I think is going to be very important to people’s freedom of expression and individual rights. So you have to balance those concerns.
“We believe it should be a serious priority for governments to figure out exactly how we could pace AI development”
I suspect that at some point it will probably become a matter of political consensus that we want to be able to deliberately slow the rate of acceleration sometimes. But you have to be extremely careful about how you design that ability.
I want to linger on that point, because many would argue that geopolitical competition itself pushes AI development to accelerate. You served as the principal staff drafter of Winning the Race: America’s AI Action Plan. If slowing the pace of AI development could lead to better global outcomes, do we also need to move away from thinking about AI as an arms race? And does that make greater cooperation with China necessary? Is the potential military “perpetual advantage” too big a risk to collaborate?
Ultimately, questions about the US–China relationship, geopolitical competition and all of these dynamics are decisions that are made—and should be made—by the elected leaders of the United States, not private corporations. So I don’t think it is my place to opine about what our geopolitical posture toward China should be.
What I do think is incumbent on us at OpenAI is to be transparent about what we’re seeing at the frontier that governments are definitely not seeing, and to keep our government candidly informed of those developments so it can make informed decisions about how it wants to proceed.
Speaking as a private citizen, however, I would say that an arms race here benefits nobody.
If I could change one thing about the AI Action Plan, it would be the title, Winning the Race. I never liked it. I understand why we chose it, but I have always thought the race metaphor was a bad one.
I wrote as much publicly after leaving the White House. I don’t think we should think about AI as a race. I think that’s a mistake.
What role can Europe still play in AI? And given the current state of transatlantic relations, how important is the Europe–United States relationship in this broader geopolitical competition driven by technological innovation?
I think Europe faces some deep structural challenges, but I also believe it has extraordinary opportunities if it plays its cards right. All of us must follow a narrow path, and Europe has to do the same.
“What I do think is incumbent on us at OpenAI is to be transparent about what we’re seeing at the frontier that governments are definitely not seeing”
Europe’s strong point is a deep reservoir of physical-world industrial expertise. Europeans know how to manufacture complex things. And it’s not just about EUV machines [extreme ultraviolet lithography systems for advanced chipmaking] or cars. There is all sorts of intricate stuff that Europe produces.
Europe needs to get smarter about protecting those industries, about growing them, and about cultivating new areas of physical-world expertise. Because of the systems we’re building in the United States, there will be a great deal of physical-world output that needs to be built. And the US alone simply will not and cannot build it.
So Europe should stop worrying so much about sovereign AI, in the sense of building its own AI systems from scratch, and instead focus on building physical industrial capacity.
The other point is that Europe has to build data centers and stop fretting about it. It really does. It just has to get it done. And that’s going to require private-sector drive, not the public sector. I think a public-sector-led approach is probably dead in the water here.
Europe needs to understand that this is all happening with far, far, far (and I want to say far one more time) more urgency than it seems to grasp. It is entirely possible to be left behind, and Europe shouldn’t let that happen to itself.
It probably also wants to reduce its debt as quickly as it can because interest rates are going to rise.
You have also been very ambitious about transparency and traceability. One of the most controversial issues lately is the architecture around chain-of-thought [where the model writes intermediate steps as it thinks, which serve as traces] and the possibility that model makers make models less transparent as they become more capable. After the Hugging Face incident —which OpenAI described as a “warning shot”—, there is a concern that we may reach a point where even those warning signals become harder to understand. What is your view of the chain-of-thought controversy? Could governments eventually regulate and enforce ways of auditing what models are capable of doing and evaluating their alignment?
When we created reasoning models at OpenAI, with the model that became o1, one of the approaches that worked was that the model had something like an internal monologue or scratchpad. It could reason about hard problems in the same way you might work through a difficult math problem by writing down the steps.
When you’re very young, you might need to write out something as basic as 3 + 4 on paper. When you’re older and tackling something like calculus, there are basic arithmetic operations you simply perform automatically in your head. You don’t necessarily write every one of those steps down on the page because you’re thinking about higher-level mathematics.
What’s happening with chain-of-thought monitorability is a lot like that.
The models are getting so intelligent that they can perform very difficult, much more complex computations, so to speak, “in their head,” without having to write down the individual steps. That can make it harder for us to monitor what the model is doing.
A lot of people anticipated that this would be true, and it’s one of the reasons why I think it is becoming harder for the industry —not just OpenAI, but the industry as a whole— to make robust guarantees about what we think a newly trained model will and will not do.
So yes, this is a real challenge.
“The models are getting so intelligent that they can perform very difficult, much more complicated computations, so to speak, “in their head””
I don’t think it is an insurmountable one. Like many problems our field has faced over the years, I believe it can be addressed through a combination of bright people, computing power, and increasingly, AI systems themselves.
There are other ways to improve monitorability. You can make systems more interpretable. You can work on AI alignment. There are many things you can do, but it requires deliberate effort.
And perhaps, rather than merely pacing AI progress and doing nothing, what we might do is pace the rate of intelligence advances for six to twelve months while launching a concerted effort on these kinds of technical issues to make the models more monitorable.
Thank you very much.
In partnership with
