The Broken Promise of Democratized AI

By C. Rich

The decision to keep the most capable artificial-intelligence models behind closed doors is often presented as prudence: a necessary sacrifice of openness in the name of national security, cybersecurity, and public safety. But from another perspective, it is a profound betrayal. The public supplied the raw material, the cultural archive, the scientific literature, the code, the language, the art, the labor, and the social data from which these systems were built. Then, once the resulting models became valuable enough to reshape research, commerce, education, and power itself, access was narrowed to a small circle of corporations, governments, favored customers, and security-cleared institutions of a democratic society.

The ordinary person is told: you helped create this, you may live under its influence, but you may not fully use it. That is not a neutral safety decision. It is a political choice about who gets to possess productive intelligence. AI labs did not build these systems in isolation. They trained on the accumulated output of civilization: public websites, libraries, research papers, open-source repositories, forums, journalism, books, educational materials, and the innumerable traces people leave when they participate in modern society. Researchers wrote papers under public grants. Programmers contributed code freely. Artists and writers created work that was scraped, copied, analyzed, and absorbed. Teachers, students, hobbyists, engineers, patients, and citizens generated the language through which these models learned to communicate with humanity.

Even where the legal boundaries are disputed, the moral reality is clear. The public’s intellectual life became an input. The public’s future should therefore be an output. Instead, the public is increasingly asked to accept a hierarchy of intelligence. A restricted class receives systems that may conduct advanced research, automate complex analysis, write production code, reason across large bodies of knowledge, and perhaps carry out powerful cyber-defense work. Everyone else receives a filtered, rate-limited, narrower version, useful enough to create dependence, but not necessarily capable enough to create equal opportunity. This divides society not merely by wealth, but by access to cognition itself. The standard justification is national security. A highly capable model, we are told, could find vulnerabilities, generate exploits, assist malicious actors, or scale cyberattacks. These are real dangers. It would be foolish to pretend that powerful AI cannot be misused. But the existence of danger does not automatically justify concentrating a technology in the hands of a few institutions.

Nearly every important human tool has a dual use. Chemistry can cure disease or create poison. Biology can produce vaccines or weapons. Computers can run hospitals or penetrate networks. The internet itself opened extraordinary new possibilities for education, trade, activism, and scientific collaboration, while also enabling fraud, surveillance, harassment, and crime. A free society does not usually respond to dual-use capability by declaring that only elite institutions may possess it. Consider guns. Firearms can be used for murder, intimidation, theft, and political violence. Yet in much of the United States, the response has not been a blanket prohibition based on the possibility of misuse. Instead, society debates licensing, background checks, safe storage, restrictions for particular contexts, and accountability for wrongdoing. One may agree or disagree with that framework, but its underlying principle is important: the possibility that some people will misuse a tool does not by itself erase the public’s claim to access.

The same principle should at least be seriously considered for advanced AI. If the argument is that a system is too dangerous for the public because it can enable bad acts, then it must explain why private corporations, state agencies, defense contractors, and select commercial partners should be trusted with that power. Institutions are not morally pure. Governments commit abuses. Corporations pursue profit. Security agencies make mistakes. Privileged access can itself create danger when it is monopolized, poorly governed, leaked, or deployed without democratic oversight. A system capable of changing the balance of cyber power is not made harmless merely because it sits in a corporate cloud rather than on a citizen’s computer. The deepest problem is not safety research. Safety research is necessary. Careful testing, red-teaming, staged release, and restrictions against clear criminal misuse can all be legitimate. The problem is allowing the same companies that profit from AI to become the sole judges of what the public is permitted to know, build, and use.

When a laboratory says, “This model is too dangerous to release,” several questions follow. Who evaluated the danger? What benchmark was used? What evidence shows that broad access would create unacceptable harm? What less-restrictive alternative was considered? How long will the restriction last? Who audits the company’s claims? Who ensures that “security” is not becoming a convenient cover for market protection? Without answers, security can become a blank check. A company that alone controls a frontier model gains a commercial advantage over smaller firms, independent researchers, universities, startups, and individuals. It can charge for access, choose preferred partners, restrict certain competitors, and influence which kinds of research become feasible. It can also lobby governments to define its own technical lead as a national-security asset, transforming private dominance into public policy.

The result is an alarming feedback loop. The company trains on public knowledge, declares the product too important to release, receives privileged government relationships because of its importance, and becomes still more difficult to challenge. What began as a technological achievement ends as a semi-private infrastructure of authority. This is not the democratization of intelligence. It is the enclosure of it. Open models do not eliminate risk. No serious advocate of openness should claim otherwise. A capable open model can be modified, fine-tuned, and used in ways that closed providers cannot centrally monitor. That is precisely why the debate matters. But openness also distributes defensive power. When models are available to universities, independent security researchers, small businesses, nonprofit groups, local governments, journalists, and technically capable individuals, more people can examine their strengths and failures. More people can build tools to identify vulnerabilities, audit systems, improve privacy, develop local applications, and create safeguards that do not depend on permission from a handful of vendors.

Security through obscurity has limits. Closed systems can prevent casual access, but they do not guarantee containment. Model weights leak. techniques diffuses. Employees leave. Competing nations train their own systems. Smaller models improve. And knowledge spreads through research, replication, and engineering ingenuity. A strategy built entirely around suppressing public access may delay diffusion, but it cannot permanently stop it. The practical question is therefore not whether advanced AI will exist outside the control of a few companies. It will. The question is whether democratic societies will prepare for that reality by cultivating broad technical competence and resilient public institutions, or whether they will leave everyone dependent on a narrow group of corporate gatekeepers until the technology is already everywhere. Open development offers an alternative: build widely accessible models, pair them with targeted safeguards, teach responsible use, fund public-interest research, strengthen cybersecurity infrastructure, prosecute actual criminal conduct, and improve defenses faster than offensive misuse can spread. That approach treats citizens as participants in a shared technological future, not as risks to be managed from above.

There is a difference between temporary restraint and permanent exclusion. It may be reasonable to delay release while a model is evaluated, while a vulnerability is patched, or while a dangerous capability is better understood. But a temporary pause becomes something else when no meaningful accountability exists and when access steadily flows upward, to corporations, governments, and select institutions, rather than outward to the people whose world supplied the model’s training material. The public should not have to accept a future in which the most powerful tools for reasoning, discovery, creation, and problem-solving are treated like private estates. If AI can accelerate medical research, then patients, doctors, researchers, and public universities deserve meaningful access. If it can improve engineering, then independent builders and small manufacturers deserve access. If it can enhance education, then students and teachers deserve more than a deliberately weakened product. If it can help defend networks, then civil society must not be left defenseless while only major institutions receive the best tools.

The burden of proof should not fall entirely on the public to explain why it deserves access to the intelligence built from its own collective work. The burden should fall on those who would withhold it. A democratic society can recognize real dangers without surrendering its technological future to private gatekeepers. It can demand transparency, independent evaluation, time-limited restrictions, public-interest access, strong accountability for misuse, and a serious commitment to open research. It can reject the false choice between reckless release and permanent corporate control. The issue is not whether AI should be governed. It must be. The issue is whether it will be governed with the public or against it.