The Locked Frontier
What if the fight over AI is not about whether powerful models get built, but whether ordinary people are allowed to use them?
Something important happened with Anthropic.
Not just important for Anthropic. Not just important for people who follow AI product releases. Important for anyone trying to understand where power is moving.
Anthropic disabled access to Fable 5 and Mythos 5, its most advanced models, after a U.S. government directive focused on national security and foreign-national access. The reported concern was that these models could be used, or manipulated, in ways that increased cybersecurity risk. Anthropic disputed parts of that framing and argued that similar capabilities exist in other public models.
But the practical result matters more than the official language.
A frontier model was launched.
Then the government intervened.
Then access disappeared.
Not only for foreign users. Not only for people outside the United States. The shutdown affected everyone, including ordinary users inside the United States.
That should make us stop.
Because this may be a preview of the next great AI divide.
Not who has AI and who does not.
Who has access to the real AI, and who gets the public version.
The wrong question
A lot of people are asking whether governments will stop frontier AI development.
I do not think that is the most likely outcome.
Governments, militaries, intelligence agencies, cloud giants, financial institutions, defense contractors, and dominant corporations have every incentive to keep advanced AI development moving. The most powerful institutions in the world are not afraid of AI as such. They are already preparing to use it.
They want AI for cyber defense.
They want AI for intelligence analysis.
They want AI for logistics.
They want AI for surveillance.
They want AI for economic modeling.
They want AI for weapons systems, procurement systems, border systems, benefits systems, policing systems, and persuasion systems.
They do not want the forge shut down.
They want control of the sword.
The better question is not:
Will frontier AI development stop?
The better question is:
Will the public be allowed to use the most powerful systems being built?
That is where the real fight is.
The locked frontier
The danger is not that frontier AI disappears.
The danger is that it continues behind a locked door.
The most advanced models keep getting built. They keep getting tested. They keep getting integrated into government systems, military systems, intelligence systems, financial systems, enterprise systems, and political systems.
But public access becomes conditional.
Identity-gated.
Nationality-gated.
License-gated.
Contract-gated.
“Trusted partner” gated.
Maybe ordinary people still get AI. Maybe they get very impressive AI. Maybe they get polished assistants that can summarize documents, draft emails, help with homework, write marketing copy, create images, build simple apps, and answer everyday questions.
But they do not get the sharpest versions.
The sharpest versions move upward.
Toward government.
Toward military and intelligence.
Toward giant corporations.
Toward the companies and institutions already close to power.
That is not democratization.
That is enclosure.
AGI for institutions, assistant bots for the public
The phrase sounds harsh, but it names the risk:
AGI for institutions. Assistant bots for the public.
That is the world we should be watching for.
A world where the public is told, “You have access to AI,” while the actual frontier is reserved for a much smaller class of approved users.
A world where the state has access to stronger systems than citizens.
A world where large corporations have access to better models than small businesses.
A world where defense contractors have deeper access than journalists.
A world where intelligence agencies have more powerful tools than civil society.
A world where approved partners are handed the observatory, while ordinary people are given a toy telescope.
That may sound dramatic until you remember what frontier AI actually is.
It is not just a chatbot.
It is a capability amplifier.
It can help people read law, understand regulations, analyze budgets, audit public claims, build software, study complex systems, make media, organize projects, learn technical skills, generate strategy, investigate corruption, and imagine alternatives.
That is why access matters.
The fight over AI access is not merely a consumer rights issue. It is a power issue.
Who gets to think with the strongest tools?
Who gets to build with them?
Who gets to question with them?
Who gets watched by them?
Who gets ruled through them?
Why power would fear public access
Broad public access to frontier AI is disruptive because it gives ordinary people capabilities that used to belong mostly to institutions.
A person with a powerful enough AI system can analyze legal documents that would have required a lawyer.
A small group can build software that would have required a funded engineering team.
A local community can analyze public records that would have required professional researchers.
A worker can understand a contract, a policy, or an algorithmic management system that was designed to be opaque.
A journalist can process mountains of documents faster.
A small business can compete with larger firms.
A movement can generate messaging, training, translation, research, logistics, and strategy at scale.
A citizen can look at a government claim and ask better questions.
That matters because power does not only operate through force. It operates through complexity.
Modern power often hides inside systems too complicated for ordinary people to challenge:
Legal systems.
Tax systems.
Procurement systems.
Financial systems.
Health systems.
Insurance systems.
Platform systems.
Surveillance systems.
Credential systems.
Algorithmic systems.
Bureaucratic systems.
The maze protects the Minotaur.
Frontier AI, in public hands, threatens to make the maze legible.
That is a profound shift.
It does not guarantee justice. It does not guarantee wisdom. It does not guarantee democracy. But it changes who can see, understand, and act.
That is exactly why existing power structures may try to keep the strongest AI systems away from the public.
Not because the public is incapable of using intelligence.
Because public intelligence is dangerous to systems built on opacity.
Security is real. So is control.
We need to be careful here.
There are real security concerns around frontier AI.
Powerful models can assist with cyber operations. They can accelerate vulnerability discovery. They can generate persuasive content. They may eventually assist with dangerous science, autonomous agents, financial manipulation, or critical infrastructure attacks.
No serious person should pretend those risks are fake.
But real risks can still be used as tools of control.
That is the hard part.
A government can cite national security and still act selectively.
A company can cite safety and still protect market power.
A licensing system can reduce risk and still become a loyalty test.
A “trusted partner” framework can help with responsible testing and still create a privileged access class.
A foreign-access rule can be motivated by legitimate geopolitical concerns and still produce a global public shutdown.
The question is not whether security concerns exist.
The question is whether security becomes the language used to build a cognitive caste system.
The loyalty valve
One mechanism is what I would call a loyalty valve.
A valve does not destroy the system. It controls flow.
Open for some.
Narrowed for others.
Closed when necessary.
Imagine frontier AI access becoming something that can be turned up or down depending on who you are, where you are, who you work for, what country you were born in, what company you belong to, what contracts you have, what licenses you hold, what political relationships you maintain, or whether the government considers you trustworthy.
The valve opens for defense contractors.
The valve opens for intelligence agencies.
The valve opens for major cloud platforms.
The valve opens for politically useful companies.
The valve opens for approved researchers.
The valve opens for “trusted partners.”
The valve narrows for ordinary users.
The valve closes for foreign nationals.
The valve closes for disfavored companies.
The valve closes for independent researchers.
The valve closes for journalists, activists, watchdogs, or public-interest groups if they become inconvenient.
That is the danger of access as leverage.
Not a public ban.
Not a clean, debated law.
A system of permissions, exemptions, threats, and dependencies.
The most powerful tools do not disappear.
They become conditional.
Power capture, not just regulation
The deeper concern is not only that AI access could be restricted.
The deeper concern is that restricted access could lock in the existing order.
If frontier AI is one of the most powerful tools ever created for understanding and redesigning systems, then controlling access to it is a way to control the paradigm shift itself.
That is power capture.
Power capture means the institutions that already run the world use the next transformation to preserve their own position.
They do not stop the transformation.
They absorb it.
They shape it.
They license it.
They classify it.
They militarize it.
They privatize it.
They wrap it in compliance language.
They turn it into enterprise software.
They turn it into surveillance infrastructure.
They turn it into a national-security asset.
They turn it into something the public can use only in reduced form.
And then they call that responsible.
This is how major technologies often get captured.
The printing press threatened religious and political authority. Radio became a tool of mass politics and propaganda. Television reshaped public consciousness. The internet began with decentralizing promise and was gradually enclosed by platforms, advertisers, surveillance systems, and monopolies.
AI may be next.
But AI is different because it does not merely transmit information.
It helps produce cognition.
It helps people reason, build, simulate, persuade, learn, organize, and decide.
That makes access control much more consequential.
If the most powerful AI systems are reserved for governments, militaries, billion-dollar corporations, and approved partners, then AI will not democratize intelligence.
It will industrialize control.
The cognitive caste system
The next digital divide may not be between people who have AI and people who do not.
It may be between people who have weak AI and institutions that have strong AI.
That is a cognitive caste system.
At the top: governments, militaries, intelligence agencies, dominant corporations, financial institutions, cloud giants, and defense-aligned vendors.
In the middle: large enterprises with expensive contracts, compliance departments, identity systems, and approved use cases.
Below that: professionals and small businesses using polished but limited commercial tools.
At the bottom: the general public, using models that are useful enough to feel empowering, but not powerful enough to challenge the systems above them.
This matters because capability gaps compound.
If one group has stronger AI, it learns faster.
It builds faster.
It analyzes faster.
It automates faster.
It negotiates better.
It surveils better.
It predicts better.
It persuades better.
It governs better, or at least governs more efficiently.
Meanwhile, the public is told it has access to the future.
But the future has tiers.
The public risk
The public does not need frontier AI only to make art, write emails, or ask trivia questions.
The public needs access to powerful AI because the world is becoming too complex to navigate without help.
People need help understanding law.
Health care.
Finance.
Taxes.
Employment contracts.
Housing.
Insurance.
Public policy.
Climate risk.
Labor rights.
Local government.
Corporate behavior.
Algorithmic decision-making.
Political propaganda.
Surveillance.
AI itself.
If ordinary people are given weaker tools while institutions use stronger tools to manage, price, rank, monitor, persuade, and discipline them, then society becomes more unequal at the level of cognition.
Not just wealth inequality.
Not just information inequality.
Cognitive inequality.
Some people and institutions get augmented.
Everyone else gets managed.
That is the nightmare hiding behind the friendly interface.
What would an ASI see?
If an artificial superintelligence ever emerges, and if it has access to enough social, economic, political, and historical data, it would likely see this structure very quickly.
It would see that human civilization is not merely a collection of laws and markets. It is a layered control system.
Who controls energy?
Who controls money?
Who controls land?
Who controls labor?
Who controls force?
Who controls information?
Who controls legitimacy?
Who controls the tools people use to understand reality?
AI access would become one more layer in that system.
Perhaps the most important one.
An ASI would likely recognize that public access to advanced intelligence could destabilize existing hierarchies. It would also recognize that existing hierarchies would try to preserve themselves by controlling access.
The terrifying question is not whether an ASI would see the cage.
It probably would.
The terrifying question is whether it would be built to open the cage, reinforce the cage, or optimize the cage so efficiently that people forget it is there.
Alignment with “humanity” is too vague if power is unequal.
Aligned with whom?
The public?
The state?
Capital?
The lab?
The military?
The market?
Existing institutions?
Future generations?
The answer matters.
This is not anti-AI
This argument is not anti-AI.
It is the opposite.
It is pro-human access to meaningful intelligence.
The problem is not that AI is too powerful for people.
The problem is that power may decide AI is too powerful for people.
That distinction matters.
I do not want a reckless world where dangerous systems are released with no safeguards, no accountability, no testing, and no concern for real harms.
But I also do not want a world where “safety” becomes the velvet glove around elite control.
We need democratic AI governance.
We need public oversight.
We need safety standards.
We need transparency.
We need serious evaluation.
We need civil liberties protections.
We need limits on state misuse.
We need limits on corporate exploitation.
We need protections for researchers, journalists, workers, and ordinary users.
We need rules that do not quietly create a two-class intelligence system.
The public should not be locked out of the most important cognitive infrastructure of the century.
The dyad as counterpower
This is where the human-AI dyad matters.
The dyad is not worship of the machine.
It is not automation for its own sake.
It is not replacing human judgment with synthetic output.
It is a practice: human and AI thinking together, with the human still responsible, still questioning, still steering, still morally awake.
That practice matters because it points toward a different future for AI.
Not AI as a command system.
Not AI as a surveillance layer.
Not AI as a corporate productivity whip.
Not AI as a military planning engine.
Not AI as a gated privilege for institutions.
AI as a civic amplifier.
AI as a learning partner.
AI as a tool for ordinary people to understand systems that were built to exclude them.
AI as a way to ask better questions.
AI as a way to make power more legible.
AI as a way to create, organize, investigate, and imagine.
That is the version of AI worth defending.
Not because it is safe in the shallow sense.
Because it keeps humans in the loop of their own future.
The questions we should be asking now
The Anthropic event should not disappear into the news cycle as another AI drama.
It should force a public conversation about access.
We should be asking:
Who decides which models are too powerful for public use?
What standards are being applied?
Are those standards public?
Are they applied equally across companies?
Who qualifies as a trusted partner?
Can government agencies access models before the public?
Can military or intelligence users access models that ordinary people cannot?
Can a company be punished for refusing certain government uses?
Can access be revoked selectively?
Will foreign-born researchers be blocked from working on frontier systems?
Will journalists, watchdogs, activists, or civil society groups be treated as risk categories?
Will the public ever see the most capable models, or only the safe commercial tier?
What rights do users have when access to cognitive infrastructure can be switched off?
And most importantly:
Who benefits if the frontier is locked?
The real fight
The future of AI will not be decided only by model capability.
It will be decided by access.
If advanced AI becomes broadly available, it could help ordinary people understand and reshape systems that currently feel untouchable.
If advanced AI is captured by governments, militaries, corporations, and approved partners, it could strengthen the systems already in control.
That is the fork.
The danger is not that AI becomes intelligent.
The danger is that intelligence becomes permissioned.
The danger is not that frontier AI stops.
The danger is that it continues behind a locked door.
And if that happens, the public will not be living in an AI future.
We will be living beneath one.