Why the freedom that matters is control over what AI remembers
Imagine an assistant that has read ten years of your email. It knows your drafts, your doubts, your health questions, the arguments you decided not to send. It may be the most useful tool you have ever had. Yet much of the memory accumulated around that relationship could remain under another company's control. You may be able to export a conversation, but could you reproduce the understanding it acquired elsewhere?
This is an arrangement we are drifting into. A handful of corporations develop the leading models, operate enormous computing facilities, and provide the platforms through which AI reaches us. Centralization has real advantages: pooled hardware, concentrated research talent, and the scale required to train powerful systems. The problem arises when those advantages become entangled with an economic model that accumulates our information and makes leaving costly.
The answer is not for everyone to build a model. It is to separate the model from what persists around it.
Own the layer that lasts
Models change. What accumulates is everything around them: memory, permissions, identity, files, projects, relationships. That persistent layer is where dependence takes hold.
If we controlled it, a model could become a replaceable component. We could move between providers, or use a smaller model on our own machines, without starting from nothing. A provider that changed its policies or prices would risk losing our business, not taking our history with it.
That requires more than a download button. Memory needs formats another system can interpret; permissions need to be specific and revocable; data should remain where it belongs whenever possible; and independent AI systems need common ways to communicate. The practical test is whether we can leave a provider and still use what we have built.
Europe's GDPR recognizes a qualified right to data portability, but that right does not automatically cover every inferred profile or representation produced by an AI system. The legal principle points toward freedom of exit; the technical reality is still incomplete.1
Surveillance has two faces
AI can read communications, financial records, location histories, and personal relationships at a scale no human analyst could match. Governments can use those capabilities to identify dissent or automate judgments about people. Corporations can turn the same capacities toward behavioral prediction, profiling, and commercial influence.
These are not isolated threats. Information collected for business purposes can become available to public authorities through legal demands or commercial channels. The combination may expand surveillance beyond what either institution could readily assemble alone.
The most reliable protection is often to avoid gathering the information centrally in the first place. When sensitive material remains under the user's control, and other systems receive only narrow and temporary access, there is less to seize, sell, or leak. Encryption, data minimization, and independent software become instruments of civil liberty, not merely technical preferences.
The freedom to think privately
We have long been able to test an idea, doubt a belief, question authority, and change our minds without necessarily leaving a transcript. AI places an intermediary inside this private intellectual space.
When someone uses an assistant to examine a political conviction, work through a personal difficulty, question a religious tradition, or develop an unpopular argument, the conversation may become a detailed record of their thinking. Such records should not automatically become assets for advertising, profiling, or indiscriminate state scrutiny.
Local control does not make anyone immune. Devices can be compromised; courts can issue lawful orders. Yet systems that collect less intimate information in central repositories create fewer opportunities for abuse. We already regard private access to books, correspondence, and conversation as essential to a free society. Private access to intelligence belongs alongside them.
The freedom to think should not depend on permission from the institution providing the machinery.
A commons with the right to refuse
Human knowledge was never built through private ownership alone. Language, mathematics, scientific methods, and artistic traditions grew through centuries of exchange. AI draws upon this inheritance, even as much of the resulting commercial value concentrates in systems that outsiders cannot inspect.
An open intelligence commons might allow independent systems to share tested methods, useful tools, and verified findings without compelling anyone to surrender underlying data. A research laboratory could publish a technique while protecting confidential records. An institution could exchange a result without exposing every document used to produce it.
The hardest test concerns knowledge held collectively. Consider a forest community whose understanding of medicinal plants developed over generations. The information is not necessarily the property of any one member; its history may include extraction without meaningful consent. For such a community, openness without authority over sharing can become another means of exploitation.
A legitimate commons must permit people and communities to decide what circulates and what stays private. Creative Commons offers a precedent for specifying permissions around creative work, although conventional licenses alone cannot settle every question raised by AI, collective knowledge, or consent.
The freedom to share and the freedom to refuse belong together.
Use the smallest sufficient intelligence
AI is not weightless. It depends on chips, electricity, cooling systems, water, manufacturing, and networks. The International Energy Agency projects global data-center electricity use rising from approximately 485 terawatt-hours in 2025 to 950 terawatt-hours by 2030, with AI-focused facilities growing especially quickly.2
Decentralization does not automatically solve this problem. A heavily utilized data center may process some workloads more efficiently than a collection of underused personal devices. Manufacturing millions of additional machines has its own environmental cost.
The useful principle is sufficiency: use the smallest model that can perform the task reliably, on suitable existing hardware where possible, and reuse results instead of needlessly repeating computation. A private document search may belong on a laptop; demanding scientific work may be better served by a shared facility powered by lower-carbon electricity.
Efficiency also raises the question of who bears the burden. A data center may use electricity efficiently while concentrating power, water, and infrastructure demands in one community. Local processing may distribute control while shifting some manufacturing and energy burdens elsewhere. Environmental responsibility therefore requires comparing lifecycle costs and asking who benefits, who pays, and who has a say in the decision.
The aim is neither to reject scale nor to celebrate local devices automatically. It is to stop assuming every task deserves the largest available machine.
Who pays for an open future?
This may be the hardest economic question. Many digital platforms gain power by making exit difficult. An open intelligence economy would try to make leaving easy. That removes a familiar source of competitive advantage, even though the underlying software still needs funding, maintenance, security, and support.
There are precedents for another model. Email and the web developed around open protocols; businesses competed in hosting, security, applications, and services rather than owning the protocols themselves. U.S. number-portability rules let customers retain telephone numbers when changing carriers, despite substantial industry resistance.3
Open intelligence could follow a similar path: customers paying for reliability, specialized capabilities, evaluation, integrations, privacy, and support rather than captivity. Consortia, cooperatives, and commercial contributors could help finance shared infrastructure. Portability rights could make it harder to manufacture switching costs artificially.
None of this is guaranteed. Open interfaces are sometimes captured by large distributors; the costs of public infrastructure can be pushed onto volunteers; proprietary convenience can beat a theoretically freer alternative. The central business experiment is whether a company can grow because it makes its customers more capable of leaving, rather than despite that freedom.
What decentralization cannot fix
Distributed systems can surveil. Open networks can concentrate power. Bad actors can exploit independent software, and responsibility becomes harder to assign when one agent relies on another and the result causes harm. Even email, built on open standards, has become highly concentrated among large service providers in many markets.
Freedom therefore needs real security, limited permissions, independently inspectable evidence, and ways to contest errors. Open code does not replace legal rights or accountable institutions. Nor does accountability have to require recording every action a person takes.
The objective is not the absence of governance. It is authority that is limited, answerable, and open to challenge—without allowing either government or corporate power to make surveillance the price of participation.
A question of organization
Much of the discussion of AI's future imagines a race toward one extraordinarily capable machine. The more consequential question may be how many intelligences are organized: who owns them, who governs their relationships, what information they may retain, and whether people can leave.
These choices are being made now, often by default. Their effects will be felt not only in markets and electricity grids, but in the private intellectual lives of ordinary people.
The question is not simply who will build the most powerful intelligence. It is whether those who use it can preserve their memories, question authority, cooperate freely, and leave without surrendering what they have created.
A more intelligent world need not be a less free one. Its freedom will depend on what we choose to build.
Sources and further reading
1. European Data Protection Board, Guidelines on the Right to Data Portability; UK Information Commissioner's Office, Right to data portability.
2. International Energy Agency, Key Questions on Energy and AI, executive summary.
3. Federal Communications Commission, number portability rules; contemporary reporting on industry opposition and the 2003 ruling.