Artificial intelligence will be a massive net good for society if the government stays out of it. Put the government in the middle, partnered with the largest AI companies, and the same technology becomes a tool for controlling what people are allowed to ask and for freezing the market in favor of the firms already on top.
That is the claim. Here is the proof.
When no one is in charge of a tool, people try things. The attempts that work get copied. The attempts that fail get dropped. Nobody has to approve the next version. That is how Linux overtook operating systems that came from ministries and monopoly vendors. It is how ordinary markets improve a product: users leave what does not work, and a rival is free to ship something better. Economists have described this for two centuries. A later popular label, borrowed from physics, calls it emergent order — structure that appears because each person acts on what he sees, not because a central office drew the plan. Open-source AI belongs to that class. Not because the model is honest. A model repeats and recombines what it was trained on, plus whatever the person in front of it asks. It belongs to that class because a bad version can be abandoned and a better version can be released without a permit.
Regulated AI belongs to the other class. Corporatism is government power plus corporate money. The largest labs want rules they can afford and that a smaller shop cannot. Politicians want a lever over the answers the public receives. Each side has what the other wants. The deal is safety language in public, and a moat in practice.
The Claim of Safety Is Just a Moat to Protect Corporate Juggernauts
A startup with a better model is a threat to a company that already has the contracts, the lobbyists, and the brand. The giant cannot always win that fight on speed. It can win it by helping write a rule that only a giant can satisfy: mandatory audits, pre-release approval, incident reporting to an agency, liability that a firm with a legal department can carry and a ten-person lab cannot. The rule is announced as protection for the public. Its effect is to keep a faster competitor from taking customers.
This is an old arrangement. Railroads, airlines, broadcast licenses, and food rules have all been used the same way. The incumbent helps design the compliance cost. The newcomer looks at the cost and stays out, or sells itself to the incumbent. Politicians receive something in return: the power to decide which systems are “safe” enough to exist, and a partner that will adjust outputs when an administration, an advertiser, or an agency asks. The public is told the alternative is chaos. The alternative is competition.
The record already matches the pattern. Closed models are services. The company keeps the weights and serves answers over a connection it controls. In June 2026 the federal government ordered Anthropic to cut off its most advanced models worldwide. The model did not have to be false. The pipe could simply be closed, because the government and the company were the only two parties who had to agree. Meta spent years releasing weights, then in 2026 began offering the next generation as a private preview to selected partners. The European Union’s AI Act divides models into those with heavy duties and those partly excused, and the fight over who qualifies is a fight over who must report to Brussels. OpenAI took Pentagon work after a rival refused uses that included mass domestic surveillance. The reluctant company was labeled a supply-chain problem. The willing company signed.
A House Judiciary Committee report in 2024 described federal efforts to control artificial intelligence in order to suppress speech. That is the political half of the trade. The commercial half is the compliance wall. Together they are corporatism: the state supplies the coercion and the contracts, and the firm supplies the inventory, the talking points, and the money.
What the Public Actually Loses
A regulated model can be changed from headquarters. If an agency, a political office, or a major customer decides that a true answer is inconvenient, the company can refuse the question, soften the answer, or retrain the system. Every user of that system receives the edited version. There is no second copy in the user’s possession to compare it with. That is the control politicians are buying. They do not need to outlaw a sentence in a statute if the only serious answer machine will not say it.
An open-source model does not give them that switch. Anyone can run it, change it, or offer a competing version. If one release refuses a legitimate question, another release can answer it. Users move. The company that will not serve them loses the business to the company that will. That is the mechanism the large labs and their political partners are trying to close, and “safety” is the name on the door.
None of this requires the belief that an open model tells the truth by being open. It does not. It produces what its training and the user’s request produce. The advantage is that error stays local. A wrong answer can be challenged by another model on another machine. A ministry cannot fix the public record by calling one vendor.
Misuse is real. Criminals will use capable software, open or closed. Researchers have already removed safety limits from released models quickly and cheaply. That is an argument for enforcing the law against fraud, hacking, and violence. It is not an argument for requiring every new model to get permission from the firms that already dominate the market. Those firms have the same technical ability. What they have on top of it is a single place where an edit, or a shutdown, reaches everyone at once.
The window-breakers who call themselves anarchists are not the counterexample. They want a committee and a punishment for dissent. That is a state with worse manners. It is the same appetite for a center, and for the power to make everyone else live inside it.
The Choice
Large AI companies move slowly once they are protected. They despise competition because competition takes revenue. Politicians despise an answer machine they cannot steer, because an answer machine they cannot steer is a limit on their power. The safety regime is how the two make the exchange: the public is restricted, the upstart is priced out, and the giant keeps the market.
Leave the tools with the people who use them. Let a worse product lose users, and let a better one ship without a waiver from the companies it threatens. That is how this technology stays a net good. Put the government in the middle, and it becomes a very expensive way to repeat the official line while the official vendors collect the fee.
