Money, power and the institutions that connect them.
FULL INVESTIGATION · 1,277 subjects · 2,054 relationships
WHAT THE RESEARCH SHOWS · 19 SEPTEMBER 2026
Who sets the rules. Who holds the switches.
AI campaigns, online identity checks and payment restrictions are different fronts in a struggle over who can build, speak and trade—and on whose terms.
Follow the money and the decisions, and a recurring arrangement comes into view. Wealthy funders sustain policy organizations. Campaigners turn their proposals into public pressure. Governments write duties and prohibitions. Companies translate those rules, alongside their own priorities, into the systems people actually encounter.
The power is spread across these institutions. The fight is over who gets to direct the machinery—and who becomes dependent on it. Red Threads maps that fight. Here is our current reading of the people and organizations doing consequential work.
The AI campaign: from a warning to a prohibition
This investigation began with Sabine Hossenfelder’s account of an offer to promote a message about AI catastrophe. ControlAI has since identified itself as the organization behind the approach; the two accounts differ over the editorial conditions. That small encounter led into a much larger operation combining creator outreach, proposed legislation and access to lawmakers.
Connor Leahy supplies a prominent public argument for stopping superintelligence. Andrea Miotti leads ControlAI’s policy and political work; Max Winga works on creator outreach. MIRI, co-founded by Eliezer Yudkowsky, has shifted resources toward communications and governance. Max Tegmark’s Future of Life Institute combines scientific authority, grantmaking and policy advocacy. PauseAI explicitly describes building public pressure alongside policy organizations.
The desired outcome is binding restriction of advanced AI development. Bernie Sanders and Greg Casar’s announced proposal gives that ambition concrete form: a superintelligence ban, a development pause, a new agency and serious penalties. ControlAI says it consulted on the framework.
The supporting money matters. Skype co-founder Jaan Tallinn reports £2.5 million in historical donations to ControlAI’s UK company. His network spans research, grant allocation, media and AI investment. Good Ventures and Coefficient Giving sustain a wider field of research and policy work. Funding determines which projects can hire, publish and keep campaigning.
The institutions now taking shape
The latest research follows the practical decisions behind this activity. Lightcone Commons organizes donor and evaluator choices. Resolution combines conditional philanthropic funding with laboratory-supplied computing credits. Epoch AI turns funded research into measurements and advice used by public and private institutions. Follow the funding into the work →
Two newer arrangements sharpen the picture. Accenture, already a commercial Claude partner, will supply paid embedded evaluation through Faculty. The Pro-Human AI Coalition publicly organizes shared policy, technical and communications work. These are different ways of assembling influence: access inside the laboratory, and organized capacity outside it. See the evaluator relationship or explore the coalition.
Three concrete connections in the mapSelect a name to read about the actor; select the connecting text to inspect the record.
An older, extensive branch runs through Oak Foundation, 5Rights and WeProtect. Oak finances organizational capacity. Beeban Kidron and 5Rights champion children’s design protections; their model-legislation work helped carry the approach into California. WeProtect connects governments, companies and civil society around responses to online child sexual exploitation.
Carnegie UK developed a duty-of-care approach cited in Australia’s statutory review. Reset supported the network continuing the UK work, while Reset.Tech Australia pressed for changes to Australian platform duties. Ofcom and Australia’s eSafety Commissioner hold public regulatory powers. Their requirements create implementation work for platforms and a market for verification and compliance providers. The consequential choices are what must be checked, who must check it, and what happens to people who cannot—or will not—comply.
The EU’s September 17 KIDS Act proposal now supplies an actual text: proposed age-access rules, certified credentials, audits and supervisory funding, alongside explicit privacy protections. Read how the proposed system would operate →
The everyday switches: money, visibility and access
Mastercard sets network rules and supplies merchant-risk information. Stripe decides which businesses it will serve. Collective Shout’s campaign targeted payment intermediaries; itch.io’s account describes the consequences for a game marketplace. Pressure applied to an intermediary can reach many people who never dealt with the campaigners.
The same practical question appears in other systems: whose assessment gets acted on? The Internet Watch Foundation produces content assessments. Tech Coalition’s Lantern exchanges safety signals. Nominet can act on domain registrations. In advertising, Check My Ads campaigns over where money flows, while DoubleVerify supplies commercial classifications. The Global Disinformation Index’s historical licensing records show another route from risk ratings into advertising decisions. Their influence comes from connecting a judgment to someone else’s access, income or visibility.
Other powerful actors want AI deployed—and bought
Andreessen Horowitz finances technology companies while pursuing policy changes and electoral influence. Marc Andreessen argues for rapid AI development and also holds a board role at Meta. Palantir, led by Alex Karp, integrates commercial AI into government operations. Microsoft and Amazon provide important parts of the infrastructure.
These actors seek markets, deployment and favorable rules. Governments want capability and authority over its use. Model suppliers want customers while retaining a say in how their products are used. Anthropic’s defense dispute exposes that bargaining: contractual limits, procurement pressure and court intervention all bear on who gets the final decision.
The result is contested. That matters.
Oak also supports digital-rights organizations that challenge restrictions. Digital Freedom Fund finances litigation. The CFPB ordered Block to improve Cash App account-restriction notices and appeals. Courts have constrained government action against Anthropic. These are working sources of opposition and remedy inside the same wider system.
Our working picture is a field of institutions exchanging money, expertise, legitimacy, access and enforcement. Interests align around particular campaigns and collide over others. Many contracts, donor conditions and downstream decisions remain out of view. The useful next step is to follow a specific action far enough to find who made the decision, who supplied its justification, who benefits, and who can change the result.
Take a closer look. Open any linked name above for its current profile. On the network map, select a person or organization, then show its connections; select a line to see the evidence behind it.
Read the original long-form investigation and its four evidence maps
The original editorial argument, retained with its source notes and links.
THE BIG PICTUREAN ARGUMENT ABOUT POWER · WITH THE RECORDS OPEN
The future, by permission.
AI could put extraordinary power into ordinary hands. But the ability to build, publish, earn and participate can still depend on someone else’s approval. Red Threads investigates the institutions in a position to turn that promise into a permission system.
The promise is not another chatbot. It is a person who can do work that once required an institution: investigate a problem, design a tool, produce a film, or start a business. Imagine that becoming cheap enough to be ordinary. The political question is not only what the technology can do. It is who gets to use it without asking.
An AI tool does not, by itself, guarantee a way to get paid, reach an audience or enter a service. Those depend on other systems. A restriction in any one of them can matter more than a freedom granted somewhere else. You may possess the means to make something and still lack a usable way to bring it into the world.
That is why this investigation looks across debates that are usually kept apart: AI restrictions, payment screening, age checks and platform rules. The particulars differ. The recurring question does not: who can turn someone else’s ability to act on or off?
Our argument is that these fights belong in the same picture. As AI threatens to spread practical power, rules around the surrounding infrastructure could keep that power dependent on a much smaller set of institutions. The danger is a future in which people can do more than ever, but exercise that ability only on terms they do not control.
That is the threat we mean by technofeudalism: not a return to castles, but a relationship of dependence. You do not need to own everything someone makes if you can decide whether they are allowed to use the road to market.
WHY READ THESE CASES TOGETHER?
Different disputes. Interlocking dependencies.
A common pattern does not require a single command centre. Governments, companies, funders and campaigners can cooperate on one question and compete on another. Their interests can overlap without becoming identical. The important work is to identify the actual handoffs: who supplies a classification, who turns it into a policy, and who turns that policy into a decision.
There are real harms to prevent. Widely distributed power includes people who will misuse it—not just the people we admire. That makes the choice of remedy consequential. Does a restriction address a specific harm, or make independent activity dependent on continuing approval? Who gets an exception? What happens when the institution is wrong?
The cases below are starting points, not interchangeable examples of one proven plan. Each exposes a different part of the machinery. Follow one far enough to see exactly where the decision is made.
01
THE MONEY THREAD Payment screening & exclusion
DOCUMENTED POLICY / ANALYSIS
The off switch isn’t always where you think.
A business can lose payment access because of a record somewhere else in the system. Stripe’s documentation describes how Mastercard’s database sits upstream of processor screening, and says a listing generally disqualifies a merchant from using , with limited exceptions. [1]
The consequential distinction is between a record and the decision made using it. They can belong to different institutions. Calling the entire result “the blacklist” hides the place where each organisation exercises its own authority.
That is the first thread to pull: not “who is connected to whom?”, but “what does this connection allow one party to do to another?”
This is a piece of the recorded network. Select a name or a labelled line to inspect it without leaving the article. Stripe’s account of the card networks is identified as such in the evidence.Open this case on the map ↗
02
THE CAPABILITY THREAD People, proposals & the right to build
PUBLISHED ROLES / PROPOSED RESTRICTION
FOLLOW THE PERSON. THEN READ THE PROPOSAL.FIG. 02
The arrows carry different meanings. A founding role is historical. An executive role has a stated scope. A proposal is not an enacted law. Open a line to see which record it comes from.Explore the people and the proposal ↗
From building AI to arguing for its limits.
is a useful starting point. Conjecture’s history names him as a co-founder of both and EleutherAI. ControlAI’s published roster lists him as its US executive director. These are distinct roles in distinct organisations. The original record also preserves Control AI US Inc separately. [2][3]
The policy demand is concrete: advocates prohibiting the development of superintelligent AI. Its stated reason is the risk of human extinction. That is not a demand to prohibit every kind of AI; it is a proposal about a defined class of capability. [4]
Now the investigation can ask a sharper question. Where would the line be drawn? Who would decide when it had been crossed? Who could still build, under what supervision, and at what cost? Those details determine whether a safety proposal also concentrates practical power.
You made it. Can you use it? Can anyone else reach it?
IDENTITY & AGE CHECKS
The condition of entry.
Ofcom’s January 2025 guidance described age-checking duties for covered services under the Online Safety Act. It included different methods, including photo-ID matching and age estimation, while stating that privacy and lawful adult access should be protected. [5]
The distinction matters. Establishing an age is not always the same as establishing an identity. The investigation has to follow the actual implementation: what must a person disclose, who receives it, what is retained, and what happens to somebody wrongly refused?
Those questions turn an abstract argument about “safety” or “privacy” into an inspectable decision about access.
YouTube’s violent-content policy distinguishes removal from age restriction. Some material can remain available in an appropriate context while being inaccessible to viewers who are under 18 or signed out. [6]
Something can therefore remain published while access to it is conditional. That is a different outcome from deletion, and it needs a different explanation. A useful inquiry follows the content, the rule applied, the decision and any appeal.
The larger question is how much of a person’s freedom to communicate depends on terms set by the route to an audience.
A restriction is only half the story. An appeal can succeed without every downstream institution changing its decision. A source can be corrected without everyone who relied on it noticing. A right to complain is not the same as a working route back.
In the payment case, Stripe’s documentation sends qualifying MATCH correction requests back to the acquirer that made the listing. Fixing the entry and restoring processing access remain separate questions. [1]
This gives the investigation a demanding test: follow a restriction forward, then follow a correction back. Name every institution that has to act. Look for the final result—not just the form someone can fill out.
The power to say no is incomplete as a description of the system. We also need to know who has the power to undo it.
A more capable public. Or a more capable permission system?
That is the choice this investigation is concerned with. It is not enough to ask whether a technology becomes more powerful. We have to ask whether the people using it become more independent—or more dependent on the institutions that surround it.
The cases here establish specific roles, proposals and access conditions. The broader claim—that these fronts add up to an effort to contain the power AI could distribute—is the argument being investigated. It needs evidence of its own. A shared person is a lead; a documented decision is a mechanism; a claim of coordinated intent requires more.
Read the argument, then leave its chosen route. Open an institution. Follow a relationship. Check the source. The network is there so you can discover where the case becomes stronger, where it becomes weaker, and which missing record would make the difference.
About this edition. This edition includes all 1,003 original subjects, 1,577 relationships, 917 source entries, 38 maps, seven stories and 172 research notes, plus later research bringing the current collection to 1,277 subjects, 2,054 relationships and 1,291 sources. The editorial argument remains separate from the records. Read the full research library ↗
ARTICLES & CONNECTION MAPS
Case studies
Each case separates the source record, the argument and the question that remains open.
Events are spaced for readability, not proportional elapsed time. Select a card to read its story and sources.
Loading dates from the research…
Source documents & connections
A source is attached to a claim—not sprinkled over a conclusion.
THE ARGUMENT MUST SURVIVE CONTACT
Competing explanations
Conviction is stronger when the decisive distinctions remain visible.
The thesis is explicit. The burden of proof is too.
The proposed explanation is that restrictions across capability, identity, distribution, and money can consolidate practical control before AI disperses it. A shared outcome can come from coordination, aligned incentives, or independent decisions. The records must tell those explanations apart.
01
DOCUMENTED MECHANISM
Who can do what to whom?
Named actors, an exact action, a date, an authority or implementation surface, and the supporting record. This is the smallest complete unit of the case.
Compare dependency, concentration, substitutability, asymmetry, and recourse across cases. Similar patterns need comparable evidence; thematic resemblance alone is not a cross-institutional link.
03
STRATEGIC INTENT
Was this outcome sought—and by whom?
Look for scoped plans, shared instructions, contractual rights, meeting records, or choices that distinguish coordination from parallel incentives. Keep deliberate actions separate from a claim about the intent of the entire system.
The published dataset does not establish a coordinated global project. The full investigation must supply that argument’s evidence.
THE STRONGEST RIVAL
Independent efforts to address real harms.
Some restrictions may originate separately, work within limited scopes, and protect people without entrenching incumbents. Test benefits, proportionality, available alternatives, competition, and demonstrated remedy—not merely the stated intention.
EVIDENCE THAT WOULD MATTER
Look for the discriminating record.
Who receives an exemption? Who can operate independently? Does the rule expire? Can a person appeal to an independent body? Does a correction restore access? Which choices would be unlikely under the rival explanation?
Bring a better record.
Prepare a correction or a missing link with the exact claim, source, date, and change it warrants.
YOUR LOCAL EVIDENCE NOTEBOOK
Saved items
Saved in this browser when storage is available. No account. No tracking. No silent uploads.