AI Will Remain. Your Access Conditions May Not.
What I Asked GPT-5.5 Thinking While Building doll
Published: · Updated:
doll is a personal AI continuity system intended to preserve a user's state and ongoing work even when models, cloud services, local AI applications, or runtime environments change.
One of the concerns that led me to the project was the possibility that today's relatively easy access to high-performance AI may not continue under the same conditions.
While developing doll, I asked GPT-5.5 Thinking the following question in substance:
Do you think the kind of world in which doll becomes necessary will actually arrive?
I do not mean a world in which AI disappears altogether. I mean one in which people can no longer rely on continuing access to high-performance AI at today's prices, capabilities, level of freedom, account conditions, and regional availability.
This concern is what led me to the idea of doll. Without assuming the concern is correct or incorrect, give me your candid view based on the situation as it stands now.
This is not a word-for-word reproduction of the original prompt. I clarified its meaning for publication.
The answer below is an English translation of the original Japanese response. Apart from translation and Markdown formatting, its substance has not been altered.
My View
I think it will. But the situation you originally feared is unlikely to arrive as a single, sudden event in which ordinary people completely lose access to high-performance AI.
What I think is most likely is something slower and more troublesome.
AI itself remains available. But the model, price, features, memory, terms, account, and regional conditions you had yesterday can no longer be preserved as they were.
You could argue that this has already begun.
What Is Likely to Become Unstable Is Not AI Itself, but the Conditions of Access
On June 15, 2026, Anthropic retired the then-current Claude Sonnet 4 and Claude Opus 4 API snapshots. Models released in May 2025 had left ordinary API use roughly a year later. Claude 3.7 Sonnet had already been retired on February 19, 2026. Anthropic documents these retirements here.
OpenAI also announced on June 11, 2026 that older GPT-5 and o3 model snapshots were being deprecated. Older image models, Sora 2, and the Assistants API were also on shutdown or migration timelines. OpenAI explains that once a shutdown date passes, the affected model or endpoint is no longer available. The current deprecation schedule is published here.
This is not abnormal corporate behavior. Consolidating development around newer models is rational.
But from the user's side, it means:
- you cannot freeze the behavior of a model you like;
- systems built around it must keep migrating;
- there is no guarantee that a newer model will produce the same results;
- past AI environments may be impossible to reproduce.
The problem doll is concerned with is therefore not merely hypothetical. It is already happening on a smaller scale.
I Do Not Think High-Performance AI Will Disappear
I do not expect AI services themselves to vanish.
Demand is too large, and AI has become too important to companies and governments. If anything, it is moving closer to becoming ordinary infrastructure.
But operating frontier AI requires enormous facilities and energy. The International Energy Agency projects that global data-center electricity consumption will rise from about 415 TWh in 2024 to about 945 TWh in 2030. It also notes that electricity infrastructure often takes longer to build than data centers themselves. The IEA's figures are available here.
This is not proof that prices must rise. But it does mean that frontier AI will continue to depend heavily on:
- large pools of capital;
- semiconductor supply;
- electricity;
- data centers;
- national regulation;
- the business decisions of a small number of companies.
Even if the technology becomes widespread, access to the highest level of performance is likely to remain a service whose conditions are set by companies, not a public utility.
Regional and Account-Based Conditions Will Remain
Access to cloud AI is not a universal right even now.
OpenAI specifies the countries and territories in which its services are supported, and states that access from or provision of access to unsupported locations may result in an account being blocked or suspended. Its current regional policy is published here.
For someone living in Japan, this is unlikely to become an immediate major problem. Over the longer term, however, we may see more of the following:
- national regulation;
- export controls;
- sanctions;
- age and identity verification;
- payment requirements;
- restrictions based on intended use;
- a widening gap between enterprise and individual access.
A world in which AI exists but you cannot use it under the same conditions is more plausible than one in which it disappears completely.
But the Outlook Is Not Entirely Pessimistic
There is also a strong movement in the opposite direction.
In 2025, OpenAI released the Apache 2.0-licensed open-weight models gpt-oss-20b and
gpt-oss-120b, describing the former as suitable for efficient deployment across a wide range of
environments, including consumer hardware.
OpenAI's announcement is available here.
Other open-weight models and local runtimes are advancing as well. For that reason, I do not consider this future very likely:
AI usable by ordinary people disappears altogether.
A performance gap may remain, but it is likely that people will still be able to keep some form of local AI on hardware they control.
This does not make doll unnecessary. It is what makes doll viable.
If no local model existed as a fallback when cloud access was lost, preserving state alone would be of little use. Progress in open-weight models makes doll's central idea more realistic: continue with reduced performance rather than lose continuity altogether.
My Subjective Probabilities
This is not a statistical forecast. It is my judgment based on the situation as it currently stands.
Over roughly the next five years
Today's broadly favorable conditions continue with little change: around 20%
High-performance AI remains relatively affordable, users can choose freely among major services, and migration remains a minor inconvenience.
Cloud AI remains plentiful, but model turnover, pricing, terms, and features change rapidly: 60–75%
I consider this the most likely outcome. In this case, doll's emergency value is limited, but the value of separating state from providers is high.
You lose access to one of the major services you use, or it becomes impractical to use: 30–50%
This includes model retirement, pricing, usage limits, account problems, and changes in service policy. Because alternatives are likely to remain, losing one service does not necessarily mean losing AI altogether.
You broadly lose access to major cloud services and have to move primarily to local AI: 5–15%
For an individual living in Japan, I would put the chance of such a broad disruption within five years on the lower side.
Over roughly the next ten years
I think the chance that today's conditions remain essentially unchanged for ten years is very low.
Over a decade, companies rise and fall, regulations change, geopolitics shifts, models are retired, pricing systems are redesigned, identity requirements evolve, and services are reorganized. There is a high chance that doll's migration or recovery capabilities would prove useful at least once.
Whether events become serious enough to require a fully offline recovery from the Offline Recovery Kit is another question. That part is insurance.
The Real Threshold at Which doll Becomes Necessary
doll does not become relevant only when AI disappears from the world.
Any one of the following would be enough to give it a role:
- a model you rely on is retired;
- a service you use shuts down;
- the price rises beyond what you can accept;
- you cannot move your accumulated memory to another service;
- you lose access to your account;
- you need to work with material that cannot be sent to a cloud provider;
- switching providers repeatedly destroys the context of your work;
- you need to continue working during a network outage.
In that sense, it is more accurate to say that the world in which doll is needed is not something that may arrive someday. It is already beginning to arrive, in varying degrees.
Choosing Not to Build doll Could Still Be Rational
The existence of a risk does not mean everyone must build a system themselves.
If an existing product provides enough portability and recoverability, and you trust its implementation, then using it may be the right decision.
The reason to build doll is not that nobody else has built anything similar.
It depends on how much of the system you need to keep under your own control in order to trust it.
Your concern is not merely that a service might shut down. It is that after AI becomes deeply embedded in your work, the terms under which you can continue using it will still be set by someone else.
I think that concern is reasonable.
My Final View
My prediction is this:
AI will not disappear. It will become more deeply embedded in society.
But the guarantee that you can keep using high-performance AI under today's conditions— with the same model, price, freedom, and memory—will continue to weaken.
I will not claim that doll's worst-case scenario—a broad loss of access—will definitely happen.
But I do think this world is highly likely:
Cloud AI remains available, but dependence on it makes your own environment impossible to preserve.
That is the world in which doll is most realistically needed.
doll is not a bet on a future in which AI disappears. It is preparation for a future in which AI becomes indispensable while users still do not own the conditions of continuity.
I believe that future is not merely plausible. It is already taking shape.
Notes on Publication
By the time this answer was generated, doll had already entered the concept stage, and GPT-assisted specification work, research, design, and development support had already begun.
GPT-5.5 Thinking therefore knew about doll's purpose and design philosophy before answering the question. I cannot rule out the possibility that the answer was influenced by that context, or biased toward validating the project.
The probabilities in the answer are not statistical forecasts. They are subjective judgments based on information available in June 2026 and on the context of the conversation. The categories overlap and are not intended to add up to 100%.
I am sharing the answer because, despite those limitations, I found it an interesting way to think about the conditions doll is intended to prepare for.
doll official website: https://doll.badjoke-lab.com/