Reason Set Free

Open or closed, American or Chinese, the models will compete. Humanity's accumulated reasoning need not become collateral in that contest.

The race is shifting from capability to adoption

Washington has begun to worry that it may be winning the wrong artificial-intelligence race.

For years the most visible competition in AI was a race toward the frontier: who could train the biggest model, marshal the most compute, and post the best benchmark scores. On those terms the United States held formidable advantages. Its technology companies could spend sums on data centers and chips that few rivals could match.

China has increasingly chosen a different game.

Many of its strongest AI companies release models whose weights anyone can download, modify, and run independently. DeepSeek, Alibaba's Qwen, Z.ai, and Moonshot AI have built a growing ecosystem of capable, inexpensive open-weight models. Writing for Brookings in the summer of 2026, the analyst Kyle Chan put their appeal bluntly: what you get from a Chinese model is something like ninety percent of an American model's capability at a fraction of the cost. The figure is his rough characterization rather than a stable benchmark, and it will move in both directions as models are released. The strategic point survives the number. The race, Chan argues, is becoming a contest over access and adoption—over which models the rest of the world actually runs.

That leaves American AI companies with an uncomfortable problem. The economics justifying hundreds of billions of dollars of frontier investment depend partly on capturing the value those systems produce, through subscriptions and API fees. But the more expensive and closed a system becomes, the more attractive a nearly-as-good model becomes if a developer can download it, customize it, and run it locally. Chinese companies can trade direct rent for diffusion. If developers from Nairobi to São Paulo to Tokyo build on Chinese models, the model that ranks second on a benchmark can finish first in influence.

This is often described as a battle between open and closed AI. The deeper contest is over who will inherit the reasoning these systems help produce. Models now sit inside scientific work, government analysis, medicine, programming, and ordinary judgment; as they do, a commercial struggle over who supplies intelligence can quietly become a struggle over humanity's accumulated reasoning. That reasoning need not become collateral in the AI race.

The distinction that matters is easy to state and easy to miss. Open weights make intelligence capacity portable: anyone can download the machine. They do not make the reasoning portable—the claims an organization came to accept, the evidence behind them, the disagreements it worked through, the predictions it registered, and the ones that failed. That material still lives inside conversations, inside one vendor's memory systems, and inside the habits of one family of models. Adoption therefore carries a hidden clause: choosing a model can mean handing over the memory of why you believe what you believe.

It need not. If claims, grounds, disagreements, predictions, and revisions can travel separately from the model that helped produce them, then adopting a model stops implying capture by that model's epistemic infrastructure, and lock-in stops being a weapon in the diffusion contest.

Models and reasons are different layers

The case for openness in the digital world was made powerfully before today's AI systems existed. In The Open Revolution, the economist and open-knowledge advocate Rufus Pollock starts from a fact so familiar that its implications are easy to miss. Information is unlike a physical object: give someone your bicycle and you no longer have it; give someone an idea and you still do. Once information is digital, the asymmetry becomes extreme, because reproduction costs almost nothing. Pollock concludes that the rules of physical property fit informational goods badly, and that genuine openness means more than permission to look. People must be able to use a thing, build on it, and pass it on.

AI complicates this beautiful argument.

A model's weights are information, and they copy without diminishing the original. But they are also something novel: information that embodies productive capability. A research paper about computer security gives its reader knowledge. A sufficiently capable model may give that reader a machine that performs thousands of novel security tasks without further instruction. This is why the open-versus-closed debate cannot be settled by the intuition that information wants to be free. Open weights can lower prices, let hospitals and ministries run models on their own hardware, expose systems to scientific scrutiny, and help developers adapt them to settings their creators never imagined. They can also distribute capabilities that are difficult to recall once released.

The dilemma becomes less intractable once we stop treating "AI" as a single thing. The compute used to train a model is one object. The model is another. Access to the model is a third. The reasoning produced through its use is a fourth. The authority to turn that reasoning into consequential action is a fifth. These five need not share a governance regime. A laboratory can stay private, a model can stay proprietary, a genuinely dangerous capability can stay restricted, and an organization can keep tight control over who is authorized to act. And the fourth object—the reasoning produced through legitimate collective inquiry—can still become a common inheritance.

Figure 1 · Five separable objectsFive objects that a single word, "AI," is currently used to name. There are no arrows here: the claim is that they are separable, not that one causes another.

Call the institution that holds the fourth object a Reason Commons: a governed community, and the infrastructure it runs, that preserves a reasoning record so its participants can inherit, contest, and improve what has already been worked out. The reasoning record is the durable artifact—claims linked to their grounds, the assumptions those links depend on, the objections raised against them, the predictions registered, the outcomes observed, and the revisions that followed. A reasoning process is the activity. The record is what the activity leaves behind.

Consider a claim this essay itself makes: that an organization switching AI providers risks losing the memory of why it believes what it believes. Inside a reasoning record, that claim would not sit alone. It would carry the observations behind it, the assumption it depends on—that consequential reasoning is currently trapped inside vendor systems—and the objection that a future model could reconstruct the reasoning from old documents. It would carry a prediction specific enough to be wrong: that teams which migrate providers spend measurable time re-deriving decisions they had already made. If the assumption failed, every conclusion resting on it would become visible as a candidate for revision, including this essay's.

Documents preserve conclusions, not reasoning

Most of what we call collective knowledge is stored not as reasoning but as documents. A paper reports what its authors concluded. A meeting transcript records what people said. A database holds observations. An AI conversation returns an answer and perhaps an explanation.

What a later reasoner needs is different: what the group was trying to achieve, which facts it accepted, which causal claims rested on which observations, which assumptions were disputed, which alternative explanations survived, which prediction was made before the result was known, what actually happened, which part of the model failed, what changed because of the failure, who supplied the evidence, and what remains uncertain.

Those relationships are rarely stored. They are rebuilt. New employees retrace old decisions. Researchers reconstruct old debates. Organizations feed a model a pile of documents and ask it to infer what happened. An institution that has already paid for a difficult piece of reasoning pays again to recover it.

Could a future model simply reconstruct the argument from the documents? Sometimes, and better each year. But reconstruction cannot recover evidence that was never linked to the claim it supported, a disagreement already flattened into a summary, or a prediction that exists only because someone wrote it down before the outcome was known. A model reading the file afterwards cannot distinguish a forecast from a memory of one. Even where reconstruction works, it is a cost paid repeatedly and a result nobody can audit.

Imagine an organization that relies on one model provider for five years. Thousands of consequential questions are asked and answered. Assumptions surface. Strategies fail. Experiments succeed. Employees leave. Models are deprecated and replaced. If the reasons behind those five years live only in conversations, in that provider's memory systems, and in the tacit behavior of one family of models, then changing providers is not a software migration.

Figure 2 · What lock-in costsEach arrow marks a consequence the paragraphs above have already asserted.

A reasoning record breaks that chain, because the reasoning is the durable object rather than a by-product of using a tool. A claim stays linked to its grounds. An assumption stays linked to the inference it supports. A disagreement stays attached to the exact proposition in dispute instead of dissolving into a meeting summary. Predictions are registered before outcomes are known, and when reality contradicts one, the losing belief is not quietly rewritten: the contradiction and the revision both become part of the inheritance. Participants change while the reasoning continues.

Then a proprietary American model, an open-weight Chinese model, a specialized local model, and a human expert can all enter the same inquiry. They can challenge the same causal claim and inspect the same evidence without any of them owning what accumulates.

The provider becomes transient. The reasons persist.

A commons needs boundaries and rules

The word commons invites the wrong picture: an unowned field to which everyone has unlimited access. Elinor Ostrom spent much of her career showing why that picture is mistaken. In *Governing the Commons* she studied communities that managed fisheries, forests, grazing land, and irrigation systems without relying entirely on either privatization or centralized state control. Her question was institutional rather than moral: how can interdependent people keep obtaining joint benefits when each of them has opportunities to free-ride, shirk, or exploit the rest?

The commons that lasted were governed. They had boundaries. Participants helped make the rules. Behavior was monitored, conflicts could be resolved, sanctions existed, and larger systems nested governance in layers. Ostrom cautioned repeatedly against turning these observations into a universal recipe, since the details had to fit local circumstances. But her empirical conclusion undid the simple choice between private ownership and central control: under the right institutional conditions, groups govern shared resources themselves.

For AI, a commons need not mean that every reason generated anywhere becomes public. A hospital's reasoning commons can end at the hospital. A company can hold confidential strategic reasoning inside one. Scientists can share some reasoning openly and embargo the rest. Security institutions can restrict operational detail tightly. A commons needs one narrower guarantee. Within its boundary, no participant may capture the reasoning so completely that the community can no longer inherit, contest, or extend it.

Recording reasoning has costs of its own. People argue differently when every assumption is on the record, and a complete provenance trail is also a surveillance tool. A legitimate Reason Commons therefore has to make several things true at once. A conversation is not automatically shared reasoning. Not every utterance is captured, and fewer still are ratified. A bounded commons can hold confidential reasoning without exposing it. Provenance can be preserved without being universally visible. And a model's interpretation of what someone said remains a proposal, distinct from the words themselves. The right to leave some reasoning unrecorded belongs to the design rather than standing against it.

Pollock and Ostrom can look like opponents here. Pollock presses for universal freedom to use, build on, and share information that is not private. Ostrom insists that a commons is not the same thing as open access, and that durable ones are frequently bounded and governed. AI needs both: Pollock's presumption against manufactured informational scarcity, and Ostrom's recognition that collective resources survive because institutions govern participation, responsibility, monitoring, and conflict.

The alternative to enclosure is not anarchy. It is governance.

Competition can destroy the substrate it depends on

In their forthcoming Evolving Prosocial AI, Paul Atkins and David Sloan Wilson read AI development through cultural evolution and multilevel selection. Their concern is not whether competition exists, but where it operates and what it selects for. Corporations rationally compete for market share, nations for strategic advantage, individuals for convenience—and actions that succeed at one level can damage the larger system that made those successes possible. Atkins and Wilson call this an extractive or "cancer" pattern: lower-level units thrive by undermining the conditions the higher level requires.

The U.S.–China contest exposes exactly that trap, and the incentives operate recursively. A company that releases a powerful open model can capture an ecosystem. A company that keeps one closed can capture rents. A nation that restrains a capability fears its rival will release the same capability first. A nation that releases first makes everyone else's restraint strategically pointless. No villain is required: each actor can make a locally rational choice while all of them move toward an equilibrium none would have chosen collectively.

Figure 3 · The social trapFour choices that are rational for whoever makes them, and the collective result none of them wanted.

Atkins and Wilson's proposed response is not a plea for altruism but a design problem: arrange institutions so that cooperative behavior becomes competitive, changing the selection environment itself. They draw explicitly on Ostrom's commons principles to do it, and they distinguish destructive from constructive competition. Competition becomes constructive, they argue, when it takes place inside a cooperative structure whose integrity it does not destroy.

Applied to the open-versus-closed fight, this suggests something other than abolishing the competition. It suggests changing what the competitors are allowed to enclose. Let companies compete to produce better intelligence. Let open and proprietary models compete on price, capability, safety, privacy, latency, specialization, and service. Let countries compete to produce better research ecosystems. But do not make accumulated collective reasoning the prize awarded to whoever wins the model market.

That sounds like wishful thinking until the apparent conflict is stated precisely, because both sides of it are real. Frontier models cost hundreds of billions of dollars and somebody has to fund them; the obvious way to justify that spending is to own the model and everything produced through it. A community that wants to inherit its own reasoning needs that record to stay portable and governable outside any single vendor. Written out, the two requirements look irreconcilable. They look that way only because of a premise neither side has examined: that capturing the value of a model requires owning the reasoning produced through it. It does not. A laboratory can keep its weights, its pricing, and its service proprietary while a separate, governed commons holds the claims, the grounds, and the predictions. Models can compete without forcing communities to surrender the reasoning they helped produce.

Figure 4 · The apparent conflictSolid boxes are statements the argument asserts. Both dashed boxes are things it goes on to deny: the conflict is only apparent, and it rests entirely on the assumption beneath it.

AI makes validation the scarce resource

Pollock is right that information is non-rival: an idea cannot be overgrazed like a pasture. But AI introduces a different commons problem, because it can generate ideas faster than anyone can determine whether they are true. A million agents can produce a million plausible causal explanations, objections, policy proposals, mathematical conjectures, and research hypotheses at negligible marginal cost. Experiments still cost money, measurements still take time, experts remain scarce, and reality does not accelerate because token generation does.

The scarce resource is no longer the production of reasoning. It is attention and validation capacity. Every agent has an incentive to contribute one more potentially useful thought at almost no private cost, and every contribution imposes a possible inspection cost on everyone else. An ungoverned reasoning commons therefore suffers not from depletion but from pollution.

The pasture survives. It disappears beneath plausible prose.

Figure 5 · PollutionTwo arrows converge because neither cause is sufficient by itself: volume becomes pollution only against a capacity to check it that does not grow.

This is why a Reason Commons cannot be Wikipedia with more AI. It needs an epistemic constitution, and four rules do most of the work.

1Observation stays distinguishable from inference. A model's proposal is not accepted reasoning, and ten models repeating the same inference do not produce ten independent pieces of evidence; their errors are correlated through shared training, shared prompts, and shared context.

2Disagreement attaches to specific claims. A dispute recorded against the exact proposition in question survives transmission. A dispute absorbed into synthetic consensus does not.

3Consequential predictions are registered before the outcome is known. A prediction written down afterwards is a memory of one, and the record has to say which it is.

4A failed prediction propagates. Whatever depended on it is marked for revision, and a revision that rescues a claim by narrowing its scope owes the community a new prediction.

Ostrom's design principles raise an awkward question here: what does a sanction look like in a commons whose resource is reasoning? Not a fine and not expulsion. Something closer to reliance-weight—a contributor whose predictions keep failing, or whose claims keep arriving without grounds, is relied upon less. The scarce achievement stops being the production of something intelligent and becomes the earning of collective reliance.

Why a durable reasoning record is newly possible

None of this is a new wish. Argument mapping, IBIS, the semantic web, nanopublications, and preregistration registries have each tried to make reasoning structured and durable, and most of them stalled against the same obstacle: somebody had to do the encoding. Turning an argument into linked claims, grounds, and assumptions was skilled, tedious work whose benefit went mostly to other people, later. The structure was valuable to everyone except the person creating it.

That is the condition that has changed. A model can now read ordinary conversation and draft the structure: propose the claims, propose the links to evidence, name the assumption an inference depends on, flag a contribution that may duplicate something already in the record. Drafting structure is not deciding what is true. The draft arrives as a proposal, and a person accepts it, corrects it, or declines it. Only then does it enter the record, carrying its author, its provenance, and whatever disagreement it attracted—and later evidence may still overturn what the community accepted.

Inside a governed reasoning record, a model can then do more than add to the flood. It can detect when two groups are relying on incompatible assumptions. It can trace a claim back to the evidence that supports it. It can surface an old prediction on the day its outcome becomes observable, and show which conclusions depend on a belief that has just failed. It lowers the cost of finding what deserves human review. It does not decide what the community is entitled to trust.

Whether that trade is actually favorable is an empirical question rather than a slogan. Verifying and maintaining an AI-drafted reasoning record has to cost less, over repeated future use, than reconstructing the reasoning later from the raw material, and it has to do so without increasing dangerous omissions. That is a hypothesis with a testable shape. It is the one the Reason Commons project is currently built to test, and it has no human results yet.

This also answers a fair question about who moves first. The actors with the deepest resources—the model vendors—profit from proprietary memory, so waiting for them, or for a standards body, or for a procurement rule, is waiting for an incentive to invert on its own. The alternative is to start small and specific: one bounded commons, one community that already knows it is interdependent, one inquiry whose reasoning is worth inheriting. Reason Commons is that first pilot, and this essay is one of its inquiries.

Reasons should travel without forcing consensus

Most AI policy today is about models: how capable they are, whether their weights should be released, which chips can be exported, what safeguards a provider must install. Those are the machinery of intelligence rather than the institutions of reasoning. After billions of humans and machines have reasoned together for decades, where will the record of what they learned reside? In the databases of a few AI companies? In each organization's incompatible archive of documents and chat logs? Nowhere durable at all, so that every new generation of models re-argues what its predecessors already tested?

Or in separate national infrastructures—American models embedded in American institutional memory, Chinese models embedded in Chinese institutional memory—with no way for a claim made inside one to be examined from the other?

That last possibility is where the essay's opening returns. The diffusion contest is producing two adoption spheres. If the reasoning record travels with the model, those spheres harden into two epistemic infrastructures that cannot inspect each other's grounds, and disagreement between them degrades into assertion. If the record travels separately, two communities that profoundly disagree can still discover the propositions on which they overlap.

Atkins and Wilson favor polycentric AI governance: communities coordinate across levels while keeping local authority, extending Ostrom's approach to nested governance. A Reason Commons should use the same architecture—many locally governed commons connected by shared conventions, rather than one global database of human thought. Some reasoning stays local. Some crosses a boundary. Some becomes globally open. Authority stays distributed.

The interoperability that matters is not that everyone believes the same things. It is that reasons travel without losing their structure. A proposition should be separable from its author. Evidence should keep its provenance. A disagreement should survive transmission. A participant, human or model, should be able to say "I accept this inference but reject the assumption beneath that one" and have the distinction land somewhere durable.

The internet is the obvious analogy, and it is worth qualifying rather than repeating. Companies compete ferociously on top of protocols that none of them owns; the common substrate does not eliminate capitalism but makes a different kind of capitalism possible. AI could support the same shape. But TCP/IP succeeded partly by being thin and indifferent to meaning, and reasons are not indifferent to meaning: they carry provenance and authority as well as content. A Reason Commons cannot standardize what communities believe, and should not try. What it can standardize is a small shared grammar—claim, ground, assumption, dispute, prediction, revision—thin enough to stay neutral and thick enough that a claim, its evidence, its objections, and its revisions travel together. The governance stays local. Only the grammar is common.

That would make a reasoning commons something unusual in today's politics: infrastructure for cooperation without compulsory consensus.

Begin the commons with one contribution

The governing principle is simple. Models may propose. Communities govern what they rely upon. Reality remains entitled to prove them wrong. And no model gets to own the inheritance. Every one of those four clauses has a version small enough for one person to carry out this afternoon.

Figure 6 · One person's route inThe dashed step is the machine's guess. Everything solid is a person's decision.

You have been reading diagrams of this kind for several pages. The one showing four locally rational choices converging on an outcome nobody chose is a current reality tree. The one that dissolved the funding-versus-inheritance conflict by naming the assumption underneath it is an evaporating cloud. The one directly above is a transition tree. They belong to a family of reasoning forms that predates AI by decades, and nothing in any of them was new information: every box paraphrases a sentence you had already read, and every arrow marks a claim the prose had already made. That is the point. This essay's own argument is precisely the kind of object a Reason Commons preserves—claims, grounds, an assumption someone can attack, a conflict with a resolution, and a sequence of actions—and it can be inherited, contested, and revised by people who were not present when it was written.

So take the first step rather than the whole institution. Post one claim, observation, concern, or proposed action on Bluesky with #reason-commons. Include the reason you rely on and, if you can, one observation that would change your mind. We read the tag by hand for now and reply with where your contribution appears to connect to reasoning that already exists—and you can tell us we placed it wrong. Nothing enters the commons automatically. People decide what they are willing to rely on, and the source and the disagreement stay visible either way.

If you do not use Bluesky, bring the thought straight to reason-commons.lovable.app and watch where it lands.

The experiment is whether an ordinary sentence can become more than content: whether someone else can attach evidence to it, challenge one assumption inside it without discarding the rest, register a prediction it implies, and carry an improved version forward long after you have stopped paying attention.

Pollock's great insight was that an information economy should not manufacture scarcity where abundance is possible. Ostrom's was that common resources survive through institutions rather than goodwill. Atkins and Wilson add what the age of AI requires: if competitive pressure rewards extraction, asking individual actors to behave better will not be enough, and the environment in which they compete has to change. A Reason Commons applies all three insights to a resource whose importance is only now becoming apparent.

Not information alone. Not intelligence alone. Reasoning.

The ability to inherit what another mind discovered; to see why it believed what it believed; to challenge one piece without rejecting the whole; to put a claim in contact with reality; to keep what survives; and to pass the improved structure onward.

AI makes that inheritance vastly easier to create. It also makes it vastly easier to enclose, overwhelm, manipulate, or forget.

The geopolitical contest over artificial intelligence will continue. The United States and China will pursue their advantages, open and closed models will coexist, companies will rise and fall, and today's frontier systems will eventually seem primitive. None of that has to be settled before we build institutions capable of remembering what humans and machines learn along the way. Because whichever country wins the AI race, and whichever model becomes the most powerful, humanity's accumulated reasoning need not become collateral in that contest.