Freedom in the age of intelligence

We believe AI should be an extension of individual human wills and, in the spirit of liberty, as broadly and evenly distributed as possible
— OpenAI’s original launch post

Today, we face a choice between two futures.

In Superconcentration, abundance is centrally allocated by two or three AI hyperscalers.

In Superagency, abundance springs directly from every person, company and community.

These are structural directions rather than predictions of two pure systems: one concentrates discretion, while the other distributes authorship.

I believe that we can and must choose Superagency.

The stakes are profound. They concern our freedom, dignity and prosperity—and the democratic consent and collective security on which a free society depends.
It is in our hands. This is a choice about structure. We can choose what to incentivise: centralisation or widespread empowerment.
We must act now. The rules of abundance are being authored around us. Our liberties are too important to leave to goodwill and discretion.

We do not have to choose between abundance and self-determination. We can have both.

Let’s choose Superagency. Come with me, and learn how.¹

The promise of a free society

Imagine a civilisation with universal foundations for dignity, prosperity and self-determination.

It would fulfil the deepest promise of a free society, a profound end in and of itself.
It would be the most secure bedrock for future progress, a path to further greatness.

This is an ideal of freedom and self-governance which humankind has long pursued, from the city-states of the Ancient Greeks to the founding of the American republic.

It is, in essence, the democratic project. It reaches across individuals, cultures and generations.

The generational torch of freedom

Every generation, we receive a set of structures, norms and values.

Every generation, we leave our mark on them and pass them on to those who follow us.

At rare moments, a new political settlement or technology emerges that radically expands this promise.

Magna Carta made durable the idea that even sovereign power could be bound by rules, and protected merchants against arbitrary exactions.
The Universal Declaration of Human Rights declared that fundamental rights belong to every human being—not by permission of a ruler, but by virtue of being human.
The open World Wide Web became something anyone could use and build upon without charge. An act of institutional restraint created an arena for expression, association and commerce on a scale no single owner could have produced.

At each step, we created a durable endowment of freedom for future generations.

The democratic project remains incomplete. Structures can be broken; rights can be encroached upon; norms can erode.

But at our best, we can move in the same direction:

We turn permission into rights.
We widen the circle of people able to act.
We enlarge the world that people can build together.

We are living through another such moment.

Intelligence and civilisation

In the age of intelligence, we are gaining one of civilisation’s most consequential technologies.

We can use it to expand the universal foundations of a free society through:

Distributed ambition, expanding what we believe to be possible;
Distributed agency, widening who gets to pursue those ambitions; and
Distributed abundance, turning ambition and agency into outcomes.

The result could be a civilisation with more authors:

more individuals able to pursue questions established institutions overlook;
more communities able to act from knowledge held close to the problem; and
more institutions able to cooperate without surrendering their independence.

We are still in a formative period where we can shape the structures, norms and values of this technological endowment, and embed the spirit of liberty within it.

It is our generation’s moment and responsibility to form that settlement.

Or, we can let the settlement form around us, towards a centre of authority more powerful than has ever existed in human history.

That choice is already taking institutional form.

Freedom from the whims of benevolence

The problem in this world is to avoid concentration of power
— Milton Friedman

Frontier labs all speak to a common and abstract motif: “AI for humanity’s benefit”.

Google DeepMind: “Our mission is to build AI responsibly to benefit humanity.”
OpenAI: “Our mission is to ensure that artificial general intelligence benefits all of humanity.”
Anthropic: “[Our] purpose is the responsible development and maintenance of advanced AI for the long-term benefit of humanity.”

If these missions are realised, the labs may become de facto stewards of humanity’s access to intelligence, promising extraordinary outcomes like abundant energy, the end of diseases and global food security.

If civilisation had a magic wand, we would want these outcomes.

But the democratic ideal and a free society is about more than outcomes.

The limits of benevolent central power

In 1765, Joseph II inherited the position of Holy Roman Emperor from his father, Francis I.

Joseph II ruled one hundred years before Lincoln’s famous Gettysburg Address (“Government of the people, for the people, by the people”). His own governing philosophy was an eerie, distorted ancestor of Lincoln’s: “Everything for the people, nothing by the people.”

His rule is typically regarded as ‘enlightened’ in the Western Enlightenment spirit. His motivations are debated, but his reforms are recognised:

He established a basic universal education across all social classes.
He ended discriminatory anti-Semitic laws.
He abolished the death penalty, in practice.

Whilst the outcomes were ‘enlightened’, the power structures were not. Joseph II affirmed the absolute monarchical power of the Holy Roman Emperor, and its hereditary succession.

Later, the throne passed to his nephew, Francis II, who used the same power structure to pursue his own ideas: extensive surveillance, widespread censorship and hanging his enemies.

The frontier labs are gaining unprecedented power

If intelligence becomes the engine of abundance, then control of intelligence becomes the single most powerful resource.

Today, that power is concentrating so rapidly in OpenAI and Anthropic that the leaders of the tech industry are making public warnings.

It can’t be [right] that there are only two companies in the world with token capital, and everybody else is renting it.
- Satya Nadella (Microsoft CEO)

If only a handful of institutions have superintelligence, they will inevitably exercise a controlling influence over economics, science and politics.
- Mark Zuckerberg (Meta CEO)

Both Microsoft and Meta have been incredibly successful in building and defending their dominant market leadership. In other words, they are not natural critics of concentrated economic power.

But the world order is changing, and these companies are adapting to the emergence of actors who may become vastly more powerful than themselves.

To be a tech giant is to have the resources to do an aggressive rear-guard action, ploughing hundreds of billions into AI assets such as data centres.

What might this new world order mean for the rest of us?

Frontier discretion is already consequential

There are many things you might want to use intelligence for:

Learning more about the science of your health
Strengthening your company’s cybersecurity defences
Reflecting on the future of humanity and economic power

In a truly free society, you need no special permission or favour to do these things.

Today, you need the support of the frontier AI labs.

Sometimes, they say no:

At initial release, Fable 5 was deliberately set to reroute textbook questions on biology and medicine.
When Hugging Face was hacked, Opus 4.8 and Fable 5 had model policies that prevented them from analysing the cyberattacks.
Less than a day after I worked on this essay with an OpenAI model, OpenAI deactivated both my ChatGPT and API access without warning. The notices alleged policy violations but did not identify the policy or activity.²

There are difficult trade-offs here. The more freedom that a user has with a tool, the greater their opportunity to use it maliciously.

We cannot make these trade-offs go away.

The issue is not that every decision is wrong. It is that two private labs may be making these trade-offs for all of us.

We can ask ourselves: whose trade-offs are these to make?

Liberty, security and self-rule

Freedom requires both liberty—the capacity to determine and pursue our own ends—and security from coercion and harm.

These freedoms can conflict. The question is not whether intelligence should ever be constrained. It is who draws those boundaries, through what authority, and with what rights of review and recourse.

Restrictions may be necessary where the use of intelligence threatens others. In a free and democratic society, they should be lawful, proportionate, reviewable and imposed through accountable institutions in which the governed have a voice—not rest solely on unilateral provider discretion. Under monarchs and emperors, an unaccountable centre decides what liberty and security others receive.

The problem isn’t individual actors. It’s the structure they operate within.

In one, liberty and security are protected through rights and accountable institutions. In the other, they depend on the judgement of the ruler.

Technological structures for freedom

Today, most of our freedoms are secured through political structures: rights, laws and institutions that constrain power.

Those structures remain essential. But in the intelligence age, technological structures are where our freedoms can be ensured or curtailed.

For freedom and intelligence, we do not need to wait for an entirely new political settlement. We can also build freedom into the technological structures themselves.

We cannot build Superagency in one step. But we can begin with the context that makes intelligence useful to every person, company and community.

Your context is your freedom

What would it mean to build freedom into the structure of intelligence?

To understand this, we need to take a detour into how intelligence models deliver value.

We must first make a key distinction between two things:

General intelligence: capabilities that can be applied across many people and purposes; and
Particular context: specific knowledge of e.g. a task, project, system, person or institution.

The value of context isn’t a new idea. The commercial value of knowledge management is traceable to at least the 1980s.

But, as general intelligence has become cheap, the value of context as its complement has dramatically increased.

When context is the bottleneck

When it comes to making a practical difference for humans, the biggest gains may come not from more intelligence but from better context.

We can find a remarkable example in the story of Courtney Hoffman and her seven-year-old son, Alex.

In the COVID-19 lockdown, Alex had suffered increasing and unexplained pains, which 17 doctors had been unable to resolve.
In June 2023, Courtney assembled data from eight systems of record, added her own observations, and provided the combined context to ChatGPT.
With Courtney’s context, ChatGPT proposed “tethered cord syndrome.” A specialist confirmed the diagnosis, and Alex underwent successful surgery within weeks.

This is a remarkable and genuinely life-changing outcome, and it’s clear that the fusion of intelligence and context was key.

Frontier intelligence has continued to substantially advance.

At the time, ChatGPT used GPT-4, which scored 34.9% in the GPQA Diamond scientific expertise benchmark (vs human PhD domain experts scoring 65-75%).
OpenAI’s next model, GPT-4o, scored 51.1%. Their first reasoning model, o1, scored 74.7%. Their current frontier, GPT-5.6, scores 94.1%.
Several open-weight models now clear 90%: e.g. Kimi K3 (93.5%), DeepSeek V4 Pro (92.8%) and Qwen3.8-27B (90.5%).

For life-changing personal health outcomes like Alex’s, model intelligence was not the binding constraint.

Even a model scoring 100% on GPQA Diamond can’t one-shot its way to personalised diagnoses.

GPT-4 didn’t need more than its 34.9%-level score. It needed the context that Courtney gave to it.

The binding constraint was context. As Courtney has said of her experience: “There’s nobody that connects the dots for you”. Courtney has even gone on to set up her own company for medical record management.

For many high-value personal and organisational tasks, the binding constraint is increasingly not raw model capability but assembling the particular context that makes that capability useful.

The context absorption trap of using intelligence

To make intelligence useful, we must share our context with it: context that is valuable, private and possibly proprietary.

Satya Nadella likens this to ‘paying twice’ for intelligence:

[You pay] once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful.

This creates two serious structural freedom-limiting risks: the risk of vendor lock-in and the risk of commoditisation.

Vendor lock-in risk comes from the accumulation of data and context inside a single intelligence provider. Since context is the key to intelligence’s value, we are unable to leave that vendor without leaving substantial practical capacity and expertise behind.
Commoditisation risk is the pragmatic recognition that the frontier of intelligence is constantly acquiring human data to train even stronger models. Human data labelling companies Scale AI, Mercor and Surge AI are all approaching or exceeding $1bn+ in revenue run-rate.

In a vivid rendering of these risks, Alex Karp has been regularly raising the alarm on how the AI labs may end up owning our means of production, turning us into their “vassal states”.

The reversal in power structure: owning your context

But, since intelligence needs context to be most valuable, there’s a flipside.

When credible alternatives exist, owning our context gives us practical freedom to exit—and that shifts the power dynamic in our favour.

If the age of intelligence finds its value in AI + context, there’s also something instructive in seeing how these technologies stand alone.

Context is incredibly useful, independent of AI, regardless of your goal. Even in a pre-LLM age, knowledge management was a thriving industry.
For many goals, intelligence is of limited use without context. The utility of an AI one-shot is dramatically increased by even a simple understanding of what your goal is.

If you own your context, you have something valuable in and of itself.

In the age of intelligence, that value is even greater: it is an incredibly strong defence against vendor lock-in and commoditisation risks.

To own your context is to be able to easily shift it from one provider to another.
To own your context is to have power to set terms on how intelligence providers can use it.

This is structural freedom in the age of intelligence.

It rests on two sub-conditions:

Credible plurality of intelligences: multiple sufficiently-strong intelligence models and providers, so that no single institution can withdraw a person’s or company’s access to intelligence
Meaningful ownership of context: deep, substantive agency over your data and context, so that your control of context is practical and not nominal

This is what we will examine next.

Credible alternatives are emerging

A portable context gives you something to leave with. But if every credible intelligence remains controlled by the same small group of providers, that freedom is largely theoretical.

Superagency requires credible plurality: enough models, providers and forms of custody that no single institution can withdraw a person’s or company’s practical capacity to act.

There is a strong and aligned ecosystem

Credible plurality will thrive if there is a strong and aligned ecosystem behind it.

This is well underway:

Model routing and switching is thriving, illustrated by Stripe’s agreement to acquire OpenRouter.
Open-weight models have supported valuable work that frontier providers declined, including Hugging Face’s use of GLM-5.2 in its cybersecurity investigation.
Several prominent US tech firms have swung their voice behind open-weight models in the ‘Open Weights and American AI Leadership’ essay

Credible alternatives are already good enough for important work

There have been previous episodes where people have remarked that AI progress might be slowing down - and yet models have continued to improve, sometimes quite dramatically.

One key threshold was in November 2025. Opus 4.5 and GPT-5.2 were released, and widely experienced as an inflection point for agentic knowledge work. Businesses lapped it up: both Anthropic and OpenAI have added tens of billions in annualised revenue since the start of 2026.

The frontier labs have continued to push ahead. However, their latest models haven’t always had the same enthusiastic adoption, as companies manage costs by shifting to capable and cheaper open-weight models.

Opus 4.5 and GPT-5.2 are no longer the frontier, but the frontier isn’t what matters: it’s what is ‘good enough’. Today, seven open-weight model families beat those two groundbreaking models on Artificial Analysis’s holistic Intelligence Index estimates: Kimi, GLM, Qwen, DeepSeek, Motif, MiniMax and MiMo.

The frontier still leads: the top closed model score is Fable 5.1’s 66, vs 60 for the open-weight top score (GLM-5.3).

However, for that 10% index lead, Fable 5.1 is 440% more expensive, costing an average of $3.69 per measured task vs GLM-5.3’s $0.68.

So, even if open-weight model strengths froze today, there would already be several independent model families which are good enough for meaningful classes of work.

Open weights add something that hosted competition alone cannot: they allow users or competing providers to run a model without the continuing permission of its creator.

This doesn’t guarantee a permanent state of credible plurality; distributed access to compute, reliable tooling and viable providers are also important.

But model plurality is one necessary foundation of structural freedom—and it is already real enough to build on.

The second sub-condition is next: meaningful ownership.

Context ownership must be safeguarded

‘Ownership’ is often used loosely in digital systems. Something can be described as yours while remaining under another institution’s practical control.

For example, we often speak of owning digital books, music and media. In practice, most digital content comes with restrictions that differ sharply from ordinary expectations of ownership:

Our lack of true ownership is sentimentally distressing when the content is books, songs or games that we love, which we can no longer enjoy in the same way.

Our lack of true ownership is a fundamental breach of our liberty and security if it’s our context. We would become alienated from our most sensitive and valuable data: our health histories, relationships, decisions and knowledge.

In the age of intelligence, owning our context is fundamental to our liberty and security.

We must have strong guarantees on the substance of this ownership.

The Superagency Charter of rights

The freedoms we inherit are built on settlements from previous generations. These settlements turned structures of arbitrary central power into rights defending our liberty and security.

The intelligence age requires a settlement of its own.

Your context should be owned and controlled by you.

That means:

Right to visibility. You can see the context held and used on your behalf.
Right to editability. You can correct and amend that context.
Right to auditability. You can trace how agents use and edit your context.
Right to use limitation. You can determine whether your context may be retained, used for training or applied beyond the purpose for which you supplied it.
Right to portability. You can use your context across platforms and services.
Right to plurality. You can authorise multiple simultaneous intelligences to work from your context.
Right to association. You can share with other context holders, both people and institutions.
Right to sovereignty. You can store your context yourself, or choose who is trusted to store it.

These rights must preserve provenance, permissions, delegation and revocation. They do not grant unilateral control over information that is jointly held, legally constrained or also belongs to others.

When these rights are effective in practice, they protect meaningful ownership of your context.³

Combined with credible alternatives, that meaningful ownership gives us practical freedom to move between intelligence providers.

This is a system of structural freedom and agency, where no single provider can turn custody of your context into practical control over your future use of intelligence.

Portable context and credible alternatives are not the whole settlement. But they are an indispensable part of it: any system that leaves people unable to move with the accumulated context of their lives and work leaves a central source of domination intact.

Exit disciplines providers; it does not decide which harms to others society should permit. That requires public rules, due process and enforceable remedies. Neither can substitute for the other.

What’s more: this system of structural freedom creates incentive alignment, where all actors have a shared interest in your agency.

Aligning incentives to freedom and agency

It is not from the benevolence of the butcher, the brewer, or the baker that we expect our dinner, but from their regard to their own interest.
— Adam Smith

We have seen how we can create practical freedom to exit in the age of intelligence, through meaningful ownership of context under a credible plurality of intelligences.

This structure does more than make practical exit possible, though.

It creates a system of incentives that is directly aligned to our agency.

Let’s learn about this world, where frontier labs and context providers become incentivised to promote the agency of individuals, communities and companies.

From hostage capture to honest competition

The Charter sets out some key rights for the age of intelligence.

Without these rights, intelligence and software providers are incentivised towards lock-in.

A rational path to durable commercial value is to build incredibly strong lock-in through context capture.

New intelligence and software providers might come with a superior model or product, but I’m hostage to the incumbents. These incumbents don’t need to make it delightful for me to stay, they just need to make it extremely painful for me to leave.

I have extremely limited recourse if I’m severely mistreated by the incumbents. Perhaps my account is suspended in a case of system error. If incumbents have enough market power, there’s minimal competitive pressure for them to be responsive to these cases.

With these rights, intelligence and software providers are incentivised by continual service.

When its rights are effective in practice, the Charter removes one of the strongest sources of structural lock-in. Individuals and companies can move complete, usable context to another provider without the incumbent’s permission.

Now, durable commercial value relies on retaining customers by giving them an active reason to stay.

Providers can still differentiate, build lasting businesses and earn substantial returns. But they must earn continued access by creating value, not preserve it by making exit impractical.

From central allocation to distributed authorship

In the best-case, central allocation of abundance relies on the benevolence of the powerful and a strong centralisation of information.

The Charter creates a dynamic of distributed ambition, agency and abundance.

Particular context can stay decentralised with those who it truly belongs to: with individuals, communities and companies.
On their own initiative, they can bring it to any capable intelligence and pursue their own purposes.
Nobody has to rely on a central provider identifying or approving their problem as worthy of attention.

Abundance then comes not only from advances in general intelligence, but from many more combinations of intelligence and context — from many more people able to turn what they know into what they can do.

Superconcentration distributes outputs, if we’re lucky to get benevolence.

Superagency distributes authorship, making our agency less dependent on the luck of benevolent stewards.

A civilisation with many authors

Every generation, we inherit a set of structures, and shape them for those who follow us.

Our generation is deciding the structure through which intelligence will shape civilisation.

We can leave future generations a settlement of liberty and security, where:

access to intelligence is not subject to arbitrary, unconstrained power
nobody can be alienated from the context of their lives and work
abundance originates from distributed knowledge, needs and ambitions

This is our opportunity; this is our responsibility.

From principles to implementation

Turning these rights into practice will require technical standards, interoperable systems, institutions and products that preserve enough structure, permissions and provenance for context to remain genuinely usable. Possessing the bytes is not enough if another intelligence cannot use them without the original provider’s permission or without prohibitive friction.

Technical affordances make these rights possible; institutions make them durable. Open standards, collective adoption, procurement, competition, law and accountable remedies are complementary ways to ensure that an incumbent cannot reduce a right to another revocable promise.

Different actors have different parts to play. Builders can make these rights technically real; standards bodies and adopters can make them interoperable; policymakers and courts can make them enforceable; and civil society can ensure that shared context, unequal power and harms to others remain part of the settlement.

No single company or product can supply this settlement. The Charter should set a floor, not prescribe a single blueprint. Implementations should compete above it, and none should become a new point of permission. The test is not what a provider promises, but whether people can inspect, use and move their context effectively in practice.

Choosing Superagency

This essay is, in part, a warning about the tyranny of unchecked power. No person, community or company should wield arbitrary power over our access to intelligence.

It is also a hopeful vision of a future where we can all win together.

Intelligence providers win by developing models that create value from your context, not by locking you into their systems;
Context providers win by helping you, your team and your AI agents connect your context more effectively, not by siloing your data; and
You win by holding true ownership over your context, being able to use it with any intelligence, and being self-directed in your ambitions.

These core ideas are bigger than any person or era. They can be traced back from contemporary AI builders through to Ancient philosophical thinkers.

We can all play a part. We can assert our freedoms, pick our missions and design our own incentives.

Our choices will decide how our children relate to AI.

Will abundance be centrally allocated to them, or does it spring from their own ambitions?

I’m choosing the path of Superagency. I hope you join me.

Background to this essay.

This essay is an attempt to start a conversation about the world we want to live in, the choices before us, and why AI raises the stakes.

My aim is to make the idea of structural agency vivid. To do that, I have deliberately painted in primary colours: drawing the choice sharply, while recognising that a complete technical standard or political programme would require considerably more detail than one essay can contain.

As much as it is a response to the arc of AI, this essay is also profoundly influenced by my own experiences: academic, professional and personal.

My academic influences are, most foundationally, from my undergraduate degree at Oxford, where I studied Philosophy, Politics and Economics. Several of my structural ideas and intuitions have a debt to classical liberalism and republicanism. I worry about ‘arbitrary, unchecked power’; in political philosophy, that’s often ‘the state’, but from my perspective it also can be AI companies.

My formative professional experiences have come over ~10 years working in education and technology. I’ve always been motivated by giving people agency over their lives (e.g. teaching refugees how to code or working with children on mathematical literacy). Today, it now seems to me that the structural dynamics of AI may be the most significant determinant of everyone’s agency.

On a personal note, honest self-reflection over my childhood and adult life has surfaced to me how agency is one of my highest values (and perhaps my highest value). It’s probably a significant driver of why I’ve been in startups for most of my career, and why I’m a 2x founder.

In a nutshell, then, this is my ideological bias: I think it’s a human good for you and I to both be meaningfully free to pursue our own particular ends, as much as is compatible with shared liberty and security.

This is why, in addition to publishing this essay’s ideas, I’m also working on Native. ↩
OpenAI’s ChatGPT and API deactivation notices arrived at 07:49:15 and 07:49:17 BST on 7 August 2026. Both alleged policy violations, but neither identified a policy category, triggering activity or example. I appealed that morning and received no response.

The argument does not depend on why the accounts were deactivated. Even on the most charitable assumption—that the decision was simply a mistake—my practical position was the same: my access was withdrawn, I could not understand the grounds, and I had no effective means of review or recourse. The arbitrariness at issue is therefore structural rather than a claim about motive. ↩
“Effective in practice” means more than a nominal export button or contractual permission. A person must be able to inspect, edit and audit their context; export it completely in a usable, documented form; have another host or intelligence use it without the original provider’s permission; and switch without prohibitive technical or practical friction. This Charter states the properties a genuine context right must have. It does not, by itself, specify the legislation, standards, enforcement or remedies through which every institution must honour those rights. ↩
Native Hive is a product in preview, built to make trust-free context ownership practical. In the current public repository, the complete reference node runs on SQLite and can export a verified copy of its full database; the same recorded context is exposed to compatible AI clients through the open-standard Model Context Protocol.

The implementation, documentation, capability-to-evidence map and executable conformance tests are available for inspection. The code is source-available under the Functional Source License 1.1, with each release converting to Apache 2.0 after two years.

Currently, ‘trust-free’ means that inspection and a concrete route to exit are implemented and independently examinable, rather than resting only on Native’s promises. Native Hive’s context portability is guaranteed through the SQLite export. See the repository’s capability and evidence map for the implementation, evidence and limits behind each claim. ↩