China is reportedly considering whether its most advanced AI models should remain freely accessible overseas.
That matters because the United States has already shown what government involvement in frontier AI access can look like.
In the last few weeks, we have seen:
- already-released AI models become unavailable because of government policy;
- the broad rollout of a frontier model delayed while a company worked with government officials;
- and now, reports that China is discussing possible restrictions on overseas access to its own advanced models.
We spend a lot of time asking which AI model is best.
But another question may be becoming just as important:
Will you still be able to use it tomorrow?
China is now discussing the access question
On July 7, Reuters reported that Chinese authorities had met with major technology companies to discuss possible restrictions on overseas access to China's most advanced AI models.
Alibaba, ByteDance, and Z.ai were among the companies reportedly involved in the discussions.
The talks were said to include not only closed models, but potentially more open forms of model distribution as well.
Nothing has been decided.
China has not blocked foreign access to its leading AI models.
The restrictions may never happen.
But the fact that the question is now being discussed is significant.
For the last few years, it was possible to imagine AI becoming something close to a global software market.
American companies would compete with Chinese companies.
Open-weight models would cross borders.
Developers would choose models based mainly on capability, price, latency, context length, and product quality.
That assumption is becoming harder to make with confidence.
Because this is not happening in isolation.
The United States has already shown another version of the same problem.
A model can still exist and become unavailable
On June 12, the US government applied export restrictions to Anthropic's Claude Fable 5 and Claude Mythos 5.
According to Anthropic, the directive required the company to prevent foreign nationals from accessing those models.
Anthropic said it could not reliably determine nationality in real time.
So it suspended access for everyone.
The model did not disappear.
The company did not shut down.
The infrastructure did not fail.
Policy changed, and access disappeared.
The restrictions were later lifted, and Anthropic announced that Fable 5 would return globally.
So this particular interruption was temporary.
But the larger lesson remains.
A frontier AI model can exist, work, and remain technically operational while policy determines whether users are allowed to reach it.
That is no longer a theoretical risk.
OpenAI showed a different version of the same shift
OpenAI's GPT-5.6 rollout followed a different path.
When OpenAI announced the model family on June 26, it did not immediately make the models broadly available.
OpenAI said it had shown the models and its release plans to the US government.
At the government's request, the company began with a limited preview for a small group of trusted partners.
OpenAI also said it did not want this kind of government-access process to become the long-term default.
On July 8, Reuters reported that the broader GPT-5.6 rollout was moving forward after additional testing and discussions between OpenAI and US officials.
The important point is not simply that GPT-5.6 is becoming more widely available.
It is that the timing and scope of a frontier-model rollout became something discussed with government before broad public access.
This does not mean every AI release is now directly controlled by the state.
But national-security policy is moving much closer to the point where advanced AI reaches ordinary users and developers.
What happens if AI stops being one global market?
It is too early to say that AI will split into national blocs.
That outcome is not inevitable.
I hope it does not happen.
But it is no longer difficult to imagine.
One possible future could include:
- a US-centered AI ecosystem;
- a Chinese AI ecosystem;
- separate European regulatory requirements;
- different access conditions depending on country, nationality, industry, or organization.
In that world, asking "What is the best model?" is not enough.
You also have to ask:
- Can I access it from where I live?
- Can my organization use it?
- Will the API remain available in my region?
- Can I export my data if access changes?
- Can I move my work to another model?
- What still works if the provider, policy, or network changes?
Today, choosing a provider because it is cheaper or more capable may be completely rational.
The problem begins when years of memory, work, configuration, and dependency are built around the assumption that the same access conditions will continue indefinitely.
Provider lock-in may become a bigger problem than we expected
Developers already understand provider lock-in.
We know that an API can change.
Pricing can change.
Models can be retired.
Terms can change.
A company can shut down a service.
But frontier AI introduces another layer.
The provider may not be the only actor that determines access.
Government policy may influence:
- who can use a model;
- where it can be used;
- when it can be released;
- whether overseas access remains available;
- and which organizations receive early access.
That changes the continuity problem.
It is no longer only:
What happens if my provider changes its product?
It may also become:
What happens if access to the provider itself changes because of policy?
AI resilience may matter as much as AI performance
None of this is an argument for abandoning cloud AI.
Frontier cloud models are likely to remain more capable than local models for many tasks.
They may offer better reasoning, better coding assistance, larger context windows, stronger multimodal capabilities, and access to infrastructure that most individuals cannot run locally.
Use them.
Use the best model available for the work in front of you.
But using a powerful service is different from designing your entire digital life so that nothing survives when access to that service changes.
I think AI resilience may become as important as AI performance.
That means designing systems where:
- models can be replaced;
- providers can be changed;
- data can be exported;
- memory does not exist only inside one company's service;
- a local model can provide minimum continuity when cloud access is unavailable;
- the user's accumulated state survives even when the reasoning engine underneath it changes.
A local model does not have to beat the best cloud model to be useful in this design.
It can simply keep the system alive.
It can preserve access to the user's own state.
It can provide a fallback path until another source of capability is available.
Why I am building doll
This is one of the reasons I am building doll.
doll is not an attempt to build a foundation model that competes with OpenAI, Anthropic, or Chinese AI labs.
It is an attempt to keep memory, identity, settings, data, work state, and continuity on the user's side while allowing the underlying AI model to change over time.
The design principle is simple:
local continuity first, cloud capability when available.
Cloud models can be used for additional performance.
Local models can provide a continuity path.
The model can change without forcing the user's entire AI environment to disappear with it.
I do not know whether AI will actually divide into national blocs.
I do not know whether future access restrictions will become broader or more permanent.
I hope they do not.
But after the events of the last few weeks, the possibility is harder to dismiss.
The most important AI may not be the smartest one.
It may be the one you can still use tomorrow.
Sources
Anthropic — Statement on the US government directive affecting Fable 5 and Mythos 5
Reuters — OpenAI set to launch most capable GPT model after delayed rollout
Reuters — Beijing is looking at curbing overseas access to China's top AI models, sources say
Originally published in a longer form on the doll website.
Top comments (0)