Bill Gates spent a few days recently with his two granddaughters, aged three and one. What stayed with him afterwards was a question that really vexed him. What is education actually going to be for, by the time those two are old enough to need it?
Gates published an essay in late August called “A turbulent AI era and critical choices to make.” He then sat down with CNN to talk it through.
CNN framed the essay as an argument that the AI industry is not admitting the real risk of the technology, and Gates is clear that managing it should be the world’s top priority. The line getting quoted everywhere is that AI will either be the greatest equaliser ever invented or the worst source of injustice.
He names three risks:
The first is people with ill intent. Fraud, cyber-attacks, and at the far end of the scale, bioterrorism. His position is that any model capable of designing novel molecules should be monitored, and should not be copyable into an environment where nobody is watching. Whether that is technically achievable is a fair question. He seems to think it is the price of entry.
The second is the risks to the job market. He is careful to say this has not happened yet. His argument is that it will be significantly affected, because the AI models have become reliable enough to handle well-defined work, and reliability was always the missing piece.
The third is what he calls the psychosocial effect, and this is where his granddaughters come in. His worry is a child growing up with all their friends being AIs, never pushed into the awkward business of dealing with real people. This goes well beyond the usual complaint about screen time. The concern is that a generation quietly opts out of the friction that makes us competent with each other.
Then the part that surprised me. Gates says that over the past year these systems have become far more capable, faster than even he expected, and that the criteria used to review them and the actions taken to minimise the harms are, in his words, really completely missing. Remember, this is the man who has spent over four decades inside tech industry saying the safety scaffolding, most of us assume exists, does not.
On jobs he takes the standard rebuttal head on, and the interviewer puts it to him well. Every technological shift kills jobs. The electric light bulb put lamplighters out of work and we adjusted. Gates answers on pace and breadth. Previous transitions played out over decades, sometimes several generations, and they made society richer, and richer societies invent new work for people to do. This time the change happens across most sectors, rapidly. White collar first and then blue collar as humanoid robots arrive, and the thing replacing the worker is better at defined tasks and works twenty four hours a day.
He knows how this sounds. He says the bar for anyone coming along and claiming this time is different is very high, and then clears it by staking his reputation on the claim anyway.
What actually drove him to publish, by his own account, was the silence.
He expected that once we crossed the threshold of danger, society would start arguing about the risks AI brings, properly. Instead we have people inside the AI industry disagreeing with each other in public while everyone else watches.
He wants academics involved, more corporate voices, politicians across parties listening, because the decisions coming are democratic trade-offs and they need to be made democratically if they are going to hold. He is just as impatient with the opposite reflex, the instinct to simply block things. Stopping a data centre being built does not slow anything down, he says. The data centre shows up somewhere else.
His own proposals will annoy plenty of people. National bodies coordinating across government departments, and an international one modelled on nuclear weapons inspection and aviation safety. A category of human reserved jobs, work we collectively decide people should do even when a machine could do it. And a tax on AI tokens and on robots, partly to slow the incentive to automate and partly to fund the adjustment. The robot tax is not new for him. He floated it back in 2017, the European Parliament rejected the idea that same year, and it has not become policy anywhere that matters since. You can pick holes in all three, and I would. A token tax in one country moves the compute somewhere else, which is exactly the objection he makes about data centres. Human reserved jobs sound comforting until you try to draw the boundary. An international inspection regime needs the cooperation of countries currently racing each other.
Which brings me to an historical similarity, it reminded me of, when I watched it.
In November 1910, six men travelled in some secrecy to a private club on Jekyll Island, off the coast of Georgia. Senator Nelson Aldrich, who chaired the Senate Finance Committee and whose daughter had married John D. Rockefeller Jr. An economist from the Treasury. Aldrich’s private secretary. Henry Davison, a partner at J.P. Morgan. Frank Vanderlip, president of National City Bank, the Rockefeller-aligned institution of its day. And Paul Warburg of Kuhn, Loeb and Company. They were there because the Panic of 1907 had exposed how fragile American banking was, and Congress had set up a commission to work out what to do about it. What came off that island was the Aldrich Plan, and the Aldrich Plan is the direct ancestor of the Federal Reserve.
You might want to read that guest list again.
Half that room were bankers, and they were the half holding the technical pen. Not lobbyists making representations from the outside. The people themselves, drafting the machinery of the body that would go on to regulate their own industry.
It is hard not to wonder whether we are watching the same thing again. The people who understand these systems well enough to write rules for them work for the companies that build them. When a government needs technical input on what a model can actually do, there is only really one place to get it. Look at how the European Union produced its code of practice for general purpose AI, finalised in July 2025. Thirteen independent chairs, over a thousand stakeholders feeding in, and dedicated workshops for the model providers themselves on the reasonable grounds that they were the ones it would apply to. Sensible process. Also, structurally, a room in which the regulated help draft the regulation.
However, the Jekyll Island story has an ending that usually gets left out, and it cuts the other way.
The Aldrich Plan failed. Congress would not pass it. Democrats looked at the governance arrangements, saw that the bigger banks would get more votes in choosing directors, and threw it out. What passed in December 1913 was the Glass-Owen bill, which kept a good deal of the technical machinery from the island but rebuilt the control structure so that Wall Street could not simply run the thing. The guest list shaped the draft. What settled the outcome was what happened afterwards, when people who had never been invited read it carefully enough to argue with it.
There is one thing 1910 had that we do not, though, and it is the reason all of this is harder now. It had the Panic of 1907. A crisis vivid enough that everybody agreed something had to be done, which is what generated the political energy to do it. Nuclear got its treaties and its inspection regime for the same reason. Gates makes that comparison himself, weapons or energy, and says this is a thousand times bigger. The trouble is that AI has no Panic of 1907, yet. One can argue the harm arrives gradually, mostly disguised as convenience, which makes it far harder to organise around.
What changed the Federal Reserve was that enough people outside the room understood the draft well enough to rewrite it before it became law. Gates is asking for those people to show up early this time. The trouble with a vacuum is that it never stays empty for long, and the ones who fill it are rarely the ones waiting to be asked.
Sources: “A turbulent AI era and critical choices to make”, Bill Gates, gatesnotes.com, 26 August 2026. Read the essay
CNN interview: Watch the interview
