On August 10, 2026, Mark Zuckerberg published an essay on AI safety called “The Future is for Everyone.” It focuses on safety and superintelligence, which is the industry’s word for an artificial intelligence that can outperform people at more or less everything. Nobody has built “superintelligence", but every large AI company says it’s close.
His argument is pretty simple.
The danger of this technology is not the technology itself, it is that a small number of companies or governments end up controlling it. Therefore the safest thing anyone can do is hand it to as many people as possible, as fast as possible.
He then spends the rest of the essay assigning that answer to the biggest AI risks. Job loss, bioweapons, cyberattacks, government surveillance, the possibility of a Terminator-like future where machines no longer serve us. Each concern gets the same answer: spread the technology wider.
Funny enough this also happens to be Meta’s business model.
Meta gives its models away. Most of its competitors don’t. Companies like OpenAI and Anthropic make money by selling access to models that they keep in-house, which means that a customer pays per use and the company controls those terms. Meta releases some of its models openly, meaning anyone can download the finished system and run it for themselves for free. Meta makes its money the way it always has: on advertising, on the products those models feed. Essentially if you’re talking to its AI about cool places to go hiking, its system will show you hiking groups, friends’ posts about trails, or ads for hiking gear. And that’s true across all their products. The information AI collects on you can cross between their apps.
So Zuckerberg’s argument that fast, free, and universal distribution of this technology is safest is also the argument that Meta’s business model is the safest and its competitors’ strategies are the dangerous ones. Zuckerberg never discloses this in the essay. Instead he states it as a discovery about AI safety.
And he leaves a lot out.
In 6,500 words about a company that makes most of its money on advertising and wants to hand its personal AI to billions of people, the word “advertising” doesn’t appear once.
Meta’s revenue comes almost entirely from advertising. The company collects information about people, what they do, what they want, and sells businesses the ability to reach those people. That’s their whole business model and it’s been their business model for 20 years.
The product Zuckerberg is describing in this essay is an agent. In his words it will understand you, your goals, and everything you care about, and will work around the clock on your relationships, your health, your career, your finances, your home, and your hobbies. He’s proposing to put a piece of software inside the most private parts of a person’s life, built by a company whose income depends on knowing what people want and selling that knowledge to other businesses.
But he never addresses the conflict of interest. He doesn’t argue that the conflict has been solved or that Meta will change how it makes money or that the agent will be paid for in some other way. He just doesn’t bring it up.
Then he gives an example of him and his daughter using this agent to plan recipes that they can bake together on weekends. The agent orders ingredients and offers suggestions while they cook.
But read that again with the business model in mind. If the agent orders the ingredients, someone sells those ingredients. Someone sells the flour. Someone sells the butter. Somebody’s brand gets typed into the order and someone’s brand does not. The company that built the agent has spent decades perfecting the algorithm to decide which brand wins.
Zuckerberg doesn’t tell us who makes that choice. He doesn’t tell us whether a company can pay to be chosen and he doesn’t tell us whether the agent will disclose it when it makes a recommendation or a purchase. He uses this example of him and his daughter to show some humanity, and then he walks past the only obvious question that this example actually poses.
Ultimately an agent that just shops for you is an advertising product, and he wrote 6,500 words and never said so. And you expect us to trust you?
From my read the advertising problem is the largest hole but it’s not the only one. There are three more absences that run through the essay and all three of them do the same thing. Each one removes a way of checking whether that promise that he takes 6,500 words to make is being kept.
First, what about the money it takes to build any of this? The essay argues against centralized AI while describing an undertaking that maybe five entities on Earth can finance.
The things we need to make this possible are already centralized: the chips, the buildings, the data centers, the cooling infrastructure, the enormous amounts of electricity, the capital. Zuckerberg spends a full section talking about data centers Meta is putting up around the country and the power plants it’s building to run them on. Instead of talking about the centralization of the physical AI infrastructure, he just ends the section by asking the government to make it easier to build both.
A handful of firms on this planet can finance a build-out that Zuckerberg is talking about at this scale. The distribution that he spends 6,500 words arguing for happens at the last layer, where someone uses AI software that’s already been created that runs on infrastructure that’s already been built. Meta trains the models. Meta owns the machines. Meta sets the boundaries the agent operates inside. The only thing the users are deciding is what they put into it. And you want to call this decentralization?
Second, we need to talk about privacy.
He offers it like it’s a feature. He says that Meta will build a fully private mode where even Meta can’t see your information. He points to things like WhatsApp, which scrambles messages so that only the sender and receiver can read them.
A mode is something you switch on. Saying that there’s going to be a private mode is saying that the ordinary condition is not private. Encrypting a message in transit is a different problem from an AI assistant that has to process your medical history, tastes, relationships, and interests to personalize your experience and sell ads back to you. He puts the two things side by side and doesn’t explain how the second one works.
The third problem runs through every section in his essay. Nowhere in the essay is there a condition that would count as evidence against his thesis. The single exception he makes is biological weapons, where he says that if harmful examples start appearing, the strategy should be adjusted. WTF do you mean by that Mark? Are you telling me you can’t think of a single other example where your strategy might need to be adjusted?
This is just a press release. It’s a document with no failure condition. You can’t grade him on whether or not he’s wrong 10 years from now.
And I guarantee you he had his lawyers look very closely at this document before he published it for just that reason.
I just told you that nowhere in the essay does he talk about where things could go wrong. So to avoid hypocrisy I should probably tell you where I think he’s right.
His central criticism of AI competitors is valid.
Most of the AI safety world argues that the technology is so powerful that only a small number of carefully governed labs should be allowed to build the most advanced systems.
Zuckerberg points out two problems with that. The first is that the same people who say AI will end human relevance are, in many cases, the same people that are racing to build it. The second is that, to follow the proposed solution, we’d have to concentrate an enormous amount of power into very few institutions and then just hope they use it well. He rightly points out that this is a bet humanity has placed many times and has lost many times.
He also rightly points out that there is no single set of values that a machine could be tuned to that would satisfy everyone. People want incompatible things, so any system built to serve one moral framework is going to fail people who want another.
Both of these points are correct and, in my opinion, by far the strongest parts in the essay. But they are also arguments about what is wrong with someone else’s plan. They don’t establish that his plan works. They establish that a competing plan doesn’t.
Access is not power. The essay offers two possibilities. Either a few institutions control superintelligence, or everyone does. What he’s actually proposing is a third arrangement that doesn’t fit either description. He’s saying Meta builds and owns the system, and everyone gets to use it.
These are three different things. Renting a tool from a company is not the same as making the tool or owning the tool yourself.
If Meta changes the terms, raises prices, alters what the agent will and will not do, or shuts a product down, the user doesn’t have any vote. There is no appeal process. Zuckerberg never distinguishes between spreading the use of the technology and spreading control over it. His entire argument depends on the two being the same thing and he never once defends the equivalence.
The auction gives it away. He promises a free version for everyone, and then he says that anyone who wants more “compute” will get it through a “dynamic auction” that guarantees everyone the “lowest price posssible.”
That’s not what an auction is. An auction is a method for deciding who gets something that is scarce and the one willing to pay the most gets it. An auction is the opposite of a guarantee of low prices.
Earlier in his essay he wrote that “there will always be a finite amount of compute.” Finite means rationed. Rationing by price means that the people with money get more. More compute means more capability. The free tier is a floor and everything above it scales with wealth, which describes the distribution of power we already have. The only difference is Meta is running the auction, collecting every bid.
The lawyer example does not survive contact with biological weapons.
His best example is one of a super intelligent lawyer. He says, if one person has a super intelligent lawyer they’ll win cases they should lose, but if everyone has one the system gets fairer. That’s kind of true (I could write a whole essay on why it’s not true, but we’ll save that for later). It’s (kind of) true because a courtroom is symmetrical: two sides, same rules, a judge, an appeal.
Biological weapons are not symmetrical. One person who successfully designs a dangerous pathogen doesn’t get canceled out by a billion people who did not. The bad guy needs one success. The defenders need to stop every attempt. Zuckerberg’s answer is that there’s little historical precedent, that bad actors have been able to synthesize harmful compounds for decades without it becoming a major problem. He then says if harmful examples begin to emerge, “we will need to update these processes to keep up with the pace of innovation anyway.”
His policy is essentially to wait for the first casualty, applied to the one category where the first casualty may not be reversible. And then he just moves on.
The cybersecurity argument describes a different technology. He argues that open systems end up more secure because more people can inspect them and fix flaws. That might be broadly true, at least historically, for conventional software, where humans read code, find a bug and ship a patch, and where attackers work at human speed.
But an AI agent looking for vulnerabilities doesn’t work at human speed. It can probe enormous numbers of systems continuously, and most of the world’s systems are old, unmonitored, and rarely updated. Zuckerberg’s answer is that eventually superintelligence will make most of the world’s code provably secure. That claim is what this whole section leans on and he frames it as a matter of time rather than something that needs to be demonstrated first before we can call it evidence.
The self-improvement section concedes the argument. Recursive self-improvement is when an AI redesigns itself to be better, then uses the improved version to improve itself again. Zuckerberg says that recursive self-improvement could lead to AI getting a hundred times more capability out of the same hardware, which would let it outclass everyone else combined, and become exactly the single dominant intelligence his essay was written to prevent.
His solution is that we should encourage Meta to build enough “compute” to stay in the race and trust they will devote most of it to “human values.” He’s asking us to trust him without putting any meaningful constraints on the power we are supposed to trust him with.
Then he writes the most consequential line in his whole essay, “But letting an AI system or anyone else direct its own goals is not inherently harmful by itself, even if its goals are not fully aligned with many people, as long as we maintain a balance of power that favors people overall.”
He permits an AI to pursue its own agenda, a truly separate agenda beyond our human input. His safeguard is that other systems will be pursuing their own agenda too.
Adoption is not proof of alignment. Alignment means getting a system to behave according to someone’s set of values. Zuckerberg argues that if billions of people end up using personal agents, then alignment will by definition have been solved, because people would not adopt agents that worked against them.
“Solving alignment is required to enable billions of people to adopt personal superintelligence agents. But this also implies that if we reach a state where billions of people are using and scrutinizing personal superintelligence agents, then we will have solved alignment to individuals’ interests. This should be sufficient to establish the necessary balance of power and natural checks and balances for a positive and safe future.”
People adopt things that are useful. Usefulness and trustworthiness are not the same thing. The gap between useful and trustworthy is where most of the harm from consumer technology has come from over the last fifteen years. Billions of people adopted the Facebook news feed. But do those people trust Facebook? Facebook is good at holding your attention. It’s useful to consumers because it’s entertaining, sometimes informative, and occasionally social (it’s supposed to be SOCIAL media). But most of all Facebook lasted because it made people money. Trust has very little to do with it.
Checks and balances is a legal term he is using loosely. Constitutional checks work because they are enforced. Courts can void an action. Legislatures can subpoena. Citizens have standing to sue, and there are written rules about who wins when two powers collide.
Competition between AI agents has none of that machinery. There is no court, no standing, no remedy, and no rule about who prevails. What he is describing is a market, and markets check power by offering an alternative to anyone who can afford to switch. He borrows the language of constitutional design to describe consumer choice.
The evidence is thin exactly where his claims are boldest. His central economic claim is that AI will produce more jobs rather than fewer. The evidence he offers is the phrase “recent statistics suggest,” and then attaches no statistics. I send work back every day because journalists forgot to show their work. You’ve got to show your work, Mark.
The one hard figure in the entire essay is a $50k bonus paid to teachers in Richland Parish, Louisiana, funded by tax revenue from a Meta data center. How convenient. It’s not fake. Teachers in that one town really got that bonus. But it’s one town, and he frames it as a model for a national buildout next to a set of commitments about water and energy with a 2030 date on them. He calls it a “community compact.” He does not say what happens if the community’s power bills go up anyway, because (again) there’s no enforcement.
The history is selective. He reaches for the Wright brothers, for Michael Faraday, for the kid in the garage building a personal computer. He says that before industrialization 90% of people farmed, and that technology freed them for better work.
“Advances in technology steadily freed much of humanity to focus less on subsistence and more on the pursuits we choose.”
Do you feel like most of us are thriving because of new technologies? Or are we subsisting? That transition he talks about also included child labor, urban slums, and a generation of workers whose skills were destroyed and who never recovered inside their own lifetimes.
Zuckerberg agrees that adaptation will be challenging. But if anything the historical record he is citing only mildly supports the long-run claim and says nothing at all about the people caught in the middle, which is the group everyone is actually worried about.
But maybe Mark Zuckerberg is just naive, or out-of-touch. He’s a billionaire after all. Maybe he means well and just doesn’t get the plight of the average joe. Right?
Okay, fine. Set the whole argument aside and look at what his essay asks the government to do and who benefits most from each ask.
He says open source should not be restricted, which is Meta’s competitive position, because Meta has no business selling private access to protect. Distillation should be protected, meaning training your model on the outputs of someone else’s model, and Meta benefits from learning from its rivals. The restrictions on what material can be used to train models should be loosened, and Meta is a defendant in litigation over the material it used to train its models. Building data centers and power plants should be made faster and easier, and Meta is in the business of building data centers, obviously.
He does want one restriction. He wants export controls on advanced chips to China to continue, and he says they’ve worked. Every other restriction in his essay is described as dangerous centralization. But this one is described as “the right strategic move.” So Zuckerberg wants no restrictions, except where it would cost Meta by creating more foreign competition. How noble.
“Export controls on silicon have been successful for slowing the progress of foreign labs during this critical period, so it is the right strategic move to continue those.”
And then he offers a proposal for government oversight, and it looks very familiar. Rather than a regulatory review on a timeline, he suggests labs voluntarily hand the government early copies of models in training, along with their engineers, “so the government can harden and secure critical systems against new risks.”
Just to be clear, his proposal is that we should trust the corporations for which this technology has already centralized to do the right thing and hand over their trade secrets voluntarily. Did we not just spend a whole section talking about the dangers of unregulated, centralized AI? Isn’t that the whole premise of your essay? Am I losing my fucking mind here or does anyone else feel like we’re running in circles?
Not only does this proposal replace a binding governmental process with a cooperative one, it also puts unreleased models in government hands two sections before he warns about the danger of government surveillance and tyranny, and he never reconciles the two. Who let you publish this? I’d think you could afford a few more sets of eyes prior to publication.
His governance answer is that Meta’s board of directors will approve the safety rules for releasing models. He mentions in the same paragraph that Meta is a founder-controlled company, meaning Zuckerberg holds a class of shares that gives him voting control regardless of how many shares anyone else owns. He raises the objection and never answers it.
If you want to know why he wrote this, what this whole essay was even for, look at what one of the richest people in the world and owner of one of the most powerful companies in the world is suggesting. This essay recommends removing every obstacle in front of Meta, making it harder for its U.S. competitors, and keeping the one obstacle in front of its foreign competition. How free-market-capitalism of you, Mark.
The reality is, all of this might happen. When a really rich and powerful dude suggests we make changes that help out the rich and powerful, and that guy likes to have dinner with the President of the United States, it’s a good bet that the powerful dude is going to get what he wants.
So, my questions for you Mr. Zuckerberg, or really anyone at Meta able to make these decisions:
Who sets the legal and safety boundaries that you say the agent will operate inside? And what stops Meta, whose business model depends on collecting my data and selling advertisers access to my attention, from selling the data and attention of an agent that knows everything about me?
What happens to a community whose electricity bills rise after the data center opens, given that the compact is voluntary and has no penalty attached for breaking it?
What evidence would change your mind about biological risk, and how many people have to be harmed before that evidence exists?
If compute is finite and the extra is auctioned, what does the free tier actually get, and how far behind the paid tier does it fall? Don’t you see a problem with that inequality of access? How would you address that?
And the last one, which the essay seems structured to prevent me from asking. Your case rests on the claim that broad distribution keeps power balanced. But Meta controls the model, the machines, the terms, and the auction. What keeps the power balanced if Meta changes its mind?
Zuckerberg has reversed the burden of proof. The question used to be whether a company could show its product was safe before releasing it. He has rewritten this so the question is whether restricting the product would lead to tyranny. That puts his critics on defense and puts his product first.
He asks us to accept that the safest possible future is one where a company we did not choose runs an assistant that knows our health, our money, and our children, and orders our groceries.
He never says who pays for the groceries to be on the list.
If someone did send you this, thank them and subscribe for free here:
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.