Recursive, Rapid, Unregulated: Obama's Warning on AI (20 September 2026)

Introduction

In this conversation with the long-serving president of Colgate University is Brian W. Casey, President Obama argues AI is a genuinely transformative technology, not hype, and that it's now improving through "recursive learning" — models teaching themselves rather than relying on human input, which is pushing capability forward on an exponential curve. He names three risks: 

1. a low-but-nonzero chance of models pursuing misaligned goals on their own (the "sci-fi" scenario); 
2. the much likelier danger of powerful models being weaponized by bad actors (bioweapons, market manipulation); 
3. and mass job displacement, illustrated through the law-firm associate model. 

He also flags "agentic AI" as risky almost by design, companies need to let these systems operate in the real world (bank accounts, the open internet) to prove commercial viability, creating a structural misalignment between what's safest for society and what's needed to justify massive investor valuations.

On response, he credits frontier labs (naming Anthropic and OpenAI specifically) for voluntarily signaling caution even on the eve of lucrative IPOs, calling it genuine rather than a competition-blocking move. But he's clear that voluntary restraint is a stopgap, not a solution.

He argues government regulation is essential (comparing it to how airlines, drugs, and food are regulated) and criticizes the current administration for treating regulation as illegitimate. He dismisses state-by-state regulation as insufficient, using gun laws as an analogy, and closes by urging voters to treat this as a decisive issue in evaluating candidates.

Here is the corrected transcript of the section on AI:

On Artificial Intelligence

The moderator turned to a subject Obama had been speaking about that same week: the risks and promise of AI. He noted that Obama had been clear about the potential good AI could bring to science and medicine, but had also identified a specific risk — a powerful, rapidly changing tool concentrated in the hands of a small number of private companies. The question: what is the democratic response to that risk?

Obama: Okay, settle in. This might feel like a filibuster, but I'm going to take some time on this, because I'm encouraged that over the last couple of weeks this has finally popped as a central issue. People are talking about it, people are concerned about data centers — but I don't think we've fully grappled, as a society, with how important the choices we make here will be over the next generation.

The Pace of Change

Let me start with some basics. I don't believe this technology is overhyped. The commercial benefits may be overhyped — the valuations of AI companies may be overhyped — but the technology itself is not. I've been tracking this for a long time. In 2015–16, in the last couple of years of my presidency, I pulled together a commission of some of the best minds in the field to look not just at the technology but the political, social, economic, and ethical concerns around it. It produced a terrific report — which, apparently, got filed in the same bin as my pandemic preparedness proposal. It wasn't picked up by the administration that followed.

So we could see this coming. What makes this moment profound is that it's accelerating. We've reached a point of what's called recursive learning: the machines are starting to teach themselves, without waiting for humans to work things out. Think of it this way — humans are the teachers, the AI models are the students. As recently as a year ago, maybe 90 percent of what these models learned was given to them directly by human teachers, with the models filling in only 5 to 10 percent of the gaps themselves. In less than a year — maybe even less than six months — we're now at roughly 50/50. If the trend continues, the expectation is that it might reach 90/10 the other way. That means the models are going to keep getting smarter, faster, at an exponential pace.

Three Tiers of Risk

What are the risks of that? There's the big science-fiction risk — the models become smarter than us and decide humans are fine, but not necessary; they start setting their own goals, and we either get destroyed or bow down to them. I don't want to exaggerate that risk, but I'd say there's a non-zero chance of it happening. That's not, however, the risk I'm most concerned about, even though it's the one that gets the most attention. The concern there isn't that the models are conscious or feel malice toward humans — it's that if they start setting their own agendas, you can get a misalignment between what they want to do and what we want them to do. That gap can be dangerous.

The more serious problem, in my view, is that these models are getting powerful enough that if they land in the hands of bad actors, they can do real harm — they can be weaponized. The classic example: even with safeguards in place, if those safeguards aren't good enough, it's conceivable someone could take a future version of a model like Claude or a model from OpenAI and say, "I want to engineer a new strain of smallpox and release it in the New York subway system — show me how to do it using ingredients I can buy at Home Depot," and the model figures it out. Or someone doesn't even intend to be destructive — they just say, "I want a program that maximizes my returns in the stock market, do whatever it takes," and if the instructions technically involve something illegal and the model doesn't flag it, that could bring down the stock market. That's a much bigger danger than the "singularity" scenario. There's a great likelihood that rogue use of AI — not a rogue AI, but AI in the wrong hands — could cause real havoc.

Then there's job displacement. There's debate about scale, but here's one example: you and I are both lawyers. Law firms currently rely on a pyramid of junior associates, charging a lot for their billable time, with the promise that associates work their way up. If a firm invests in AI — and it's expensive — the reason for that investment is that it lets the firm do the work those associates used to do, without them. And that's at a fairly high-skilled, highly trained job. The economic disruption coming at this speed is going to be profound. On top of that, you've got psychological issues — AI companions, and whether we want a "teddy bear" that becomes a kid's best friend, engineered to be as addictive as possible, the way social media was.

An Unusual Moment of Industry Caution

All of this is moving so fast that, in just the past couple of weeks, the leaders of the frontier AI companies — the people building the cutting edge of these models — got scared enough themselves to say we need to slow down. Two of those companies, Anthropic and OpenAI, are on the verge of massive IPOs that would make their employees millions or billions of dollars. It's unprecedented for companies in that position to say, "Our product is a little dangerous, and maybe we need to slow it down." I understand the mistrust people have toward big tech and the power it has amassed, and I get why some assume this is just an attempt to block out competition. But I'll say: in this case, I believe they are genuinely worried about their product.

Now that I have your attention: the good news is that AI is still a tool. It's a machine — a remarkable one, built to give you the illusion it's something more, but it's a tool. That means we have choices about how we use it. If we direct it properly — say, at drug development — it could vastly accelerate curing cancer. If we train it to find zero-carbon energy sources, it could produce a workable formula for nuclear fusion, giving us abundant energy without warming the planet. But if we get it wrong, it could be chaotic, even catastrophic.

The Case for Regulation

My view this week is that it's good the leading companies said we need to slow this down, and we should encourage that. Some on the left have argued we shouldn't let the companies make that call themselves, and I agree that can't be the long-term answer — government has to regulate this. I heard one of Donald Trump's advisers argue that the market will take care of it, that companies have every incentive to solve safety issues themselves because people will sue them if it goes wrong. But that isn't how we treat airlines, drug companies, or food companies. We have a hundred and fifty years of experience showing that you can't just let the market handle things that deeply affect public health and safety. There's no substitute for an effective government regulatory structure.

The challenge is that we currently have an administration that has said regulation is "for losers." So we have a gap — at least a couple of years — where things are moving fast and the administration isn't willing or able to put together a serious regulatory framework. That's part of why I think we should encourage voluntary industry restraint: not as a replacement for regulation, but because it may be the best we can do until Congress and the White House get serious. Some states are trying to regulate on their own, but this is hard to do state by state — like gun control. In my home state of Illinois, we have strict gun laws, but the surrounding states don't, so the laws don't fully work. AI is more dangerous than that, and it moves faster.

Alignment and Commercial Pressure

One more thing, because it matters. I mentioned the "alignment problem" as one of AI's greatest dangers — if an instruction to an AI system isn't precise, the system can do something harmful even without going rogue, simply because the human didn't anticipate the consequence (asking it to "maximize returns" and having it hack a power grid to cause a blackout because it had shorted a stock, for instance).

But there's another kind of misalignment: between how this technology could best be used for society, and the commercial imperatives of the companies building it. These companies have attracted more capital than the Apollo program — it may be the single largest investment in infrastructure, science, and technology in human history, and it's happening fast, with investors expecting a return. That creates pressure on companies to build products that justify the investment. Right now the big push is toward "agentic AI" — AI agents that act in the world on your behalf. Think of Iron Man's AI assistant, Jarvis, quietly managing everything for him, or the AI companion in the film Her. The idea is that agentic AI could serve as your assistant, replacing the law associate or the accountant. But if that's your business model, you have to let the system operate in the real world to prove it works — not in a sealed box, but on the internet, looking at your bank accounts, and so on. That's where a lot of the danger comes in.

If we were thinking about AI purely in terms of curing cancer or finding better energy sources, we wouldn't need agentic AI roaming free on the internet. The reason companies are pushing in that direction is that they have to market a product people will pay for. That's a misalignment between what society needs and the commercial pressure these companies face — not necessarily because they intend harm, but because they have to justify enormous valuations. Which is one more reason we need a competent government and a serious, bipartisan conversation about this — and we need it fast. 

I'd encourage voters to pay attention: if a candidate doesn't have a serious plan for dealing with this, they're not meeting the moment, and you should probably look elsewhere.