An Age of Builders

Oct 20, 2026

A Magna Carta for the Intelligence Age

In 1993 I was at university when Wired launched. Rossetto’s first-issue manifesto said the coming social changes were so profound that “their only parallel is probably the discovery of fire.” I believed him. I still do. We spent our nights on bulletin boards and the early internet, slow, and plainly the future.

In 1994 Dyson, Gilder, Keyworth and Toffler published A Magna Carta for the Knowledge Age. “The central event of the 20th century is the overthrow of matter,” they wrote, and “inexpensive knowledge destroys economies-of-scale.”

They were right about the capital and wrong about its shape

America invests more in intangibles than in buildings and machines, 15.6 percent against 10.3, measured against output adjusted to count all intangibles. At the model layer they were right too. Epoch AI finds the price of reaching a fixed benchmark score has fallen about thirteenfold a year since 2023; two years of that is roughly 170-fold. Kai-Fu Lee draws the conclusion: “Organizations were designed to ration scarce intelligence. When intelligence becomes abundant, the company itself must be redesigned.” Thought got cheap. Scale did not dissolve. Three companies take about 63 percent of cloud infrastructure revenue. Data centres used 485 terawatt-hours last year. Mind still runs on matter, and matter has landlords.

The question was who owns knowledge. Now it is who may reason at scale, on whose authority.

The subject is authority, not intelligence

Let a machine pay my suppliers. Suggesting a payment date is advice. Moving the money is power. Somebody still decides which account, how much, and when to stop. Lee’s answer is to become AI native; agreed. But rebuilt around what rules? Those are questions of delegated power. Who grants it, who withdraws it, who prevents misuse, who repairs the damage?

Manifestos argue. Charters settle

The barons at Runnymede wrote no theory of liberty, only a list of grievances, and made King John seal it on 15 June 1215. The Pope annulled it ten weeks later. Its force lay in Clause 61: twenty-five barons could seize his castles and lands if he broke it. The 1216 reissue dropped that clause.

The industrial age did the same. Smith wrote in 1776 and Marx in 1848; the Sherman Act came in 1890, the Wagner Act and Social Security in 1935. Charters come late, and someone other than the power they bind writes them.

The intelligence age has signed a charter. It is missing Clause 61

On 29 September Google, Meta, Anthropic, OpenAI, xAI and Nvidia signed the White House Accord on Super Intelligence, and the President signed first. It commits each company to four layers of control: internal monitoring, an internal team, an “independent external auditor or evaluator,” and a board committee. It says that “over time, it may make sense to codify these steps into laws.” It carries no penalty, no report to anyone outside the company, and no rule on who the evaluator is.

So the companies choose. Anthropic says it cannot veto what its evaluators publish but can require redactions in predefined categories, and it decides what falls inside them. METR takes no money from the labs and says it remains dependent on them “to voluntarily engage, provide access, and permit us to publish.” Recent pre-release windows ran from three days to four weeks.

The room is not empty. The FTC has demanded incident and safeguard records from two signatories under the FTC Act. California’s attorney general has subpoenaed one. The EU’s AI Office has served information requests under a statute with fines. California already requires incident reports within fifteen days. But none of them can choose the inspector, and none can stop a deployment. The accord’s fourth layer reports to the company’s own board.

In 1993 the Clipper chip was voluntary too: a key in government escrow, pushed by procurement and, from February 1994, export relief. In 1215 the barons bound the king. In 2026 the companies bound themselves and chose their inspectors.

The line

The fight is not safety against freedom but whether the rules keep entry open. Gatekeepers want a permit to build and a licence to run. Builders want rules for harms and no gate on entry. I am with the builders, not because the risks are small, but because a gate at entry is the one rule every incumbent will help you write.

This charter gates nothing at entry and two things at scale: taking other people’s work, and putting capabilities into the world that can hurt people who never agreed to the risk. Both gates are legitimate only while a newcomer can pass them: a deadline, a cost that scales with the firm, and a court.

The charter

Freedom is the default. A restriction attaches only to acts that put others at serious risk or take what is theirs, and it must be proportionate, appealable, and cheap enough that compliance never becomes an incumbent’s moat.

I. Freedom to learn, build and use. Anyone may study, build and run AI, including on their own hardware. This binds the state. A restriction needs a law, a necessity and a court, and nothing below reaches what you run on your own machine.

II. Privacy and encryption. Using intelligence must not cost you a private life. Encryption is a right; access to your records needs legal authority, a warrant’s discipline, and review. Records kept to hold someone accountable are not reused for training or commerce, and are not kept past their purpose.

III. Bounded and revocable authority. A machine acts for you only inside authority you can see, limit and withdraw. Responsibility follows control: developer for the system, deployer for its use, user for the instruction. Someone identifiable is always the deployer. A person the agent harms, counterparty or bystander, claims against the deployer, who recovers from the developer if the fault was in the system. And a vendor’s contract cannot cap away the safeguards this charter requires.

IV. A workable exit. You can take your data, your history, your files, and the instructions and memories a service built from your own use, in a documented format, and leave.

V. Explanation and redress. When a machine’s decision costs you a job, a loan or a home, you get the reasons and a reviewer who did not make the decision and can change it, whether or not you ever agreed to its terms; when it acts outside its authority, you get redress. In August the Dutch regulator fined Uber nearly 825 million euros, finding it had deactivated drivers by machine with no human assessment and too little explanation. Uber disputes the finding and says it will contest it.

VI. Training data and outputs. In Europe training is allocated by lawsuit and in America by fair use; it belongs in statute. Copyrighted work trains under a collective licence with a published rate scaled to the developer’s revenue, unless the rightsholder has reserved it in one registry; makers disclose what they trained on. On outputs I have changed my position: as between you and the provider, whatever rights exist in what you direct are yours, and the provider may claim none. I no longer propose a new property right. The one statute that created one, in Britain, is now marked for repeal.

VII. Frontier duties. Above a published capability threshold, with a cheap screening test below it, developers face evaluation before release by evaluators a standing body assigns and pays from a levy on firms above the line. The developer cannot veto publication, and the body has ninety days to finish or the developer may deploy under conditions. The body may pause a deployment for thirty days on published reasons, renew once, and after that only a court can; if a court finds the pause unreasonable, the body pays. A release of weights may be stopped only on a finding that the model, with its safeguards stripped, enables mass-casualty or infrastructure-scale harm. Otherwise release is lawful, and what a stranger does with the weights is the stranger’s crime.

Who enforces it

Every article answers four questions: who owes the duty, who can complain, who decides, what remedy. A payment to a fraudster’s substituted account: the deployer owes it, the unpaid supplier complains, a court decides, and the remedy is payment, recovery from the vendor whose system never checked the destination, and a fix. The bank owes nothing; the agent used its principal’s own credentials.

Run the same four on the accord. The signatories owe it. No one has standing. The signatories decide. There is no remedy.

Magna Carta lost its teeth at the first reissue. Ours belong in statute, where pulling them takes a vote in daylight. The supplier, paid the wrong amount, needs less: a law that names who owes her and a court that will hear her. A legislature writes the one. She walks into the other. A charter earns its name when someone without power can use it against someone who has it, and when the one with power knows how long the inspector has.

Sources

1. Louis Rossetto, “Why WIRED?”, Wired 1.01 (January 1993), reprinted as “The Original WIRED Manifesto,” 18 Sep 2018. wired.com
2. Cyberspace and the American Dream: A Magna Carta for the Knowledge Age, PFF Release 1.2, 22 Aug 1994 (Dyson, Gilder, Keyworth, Toffler). Both quotations verbatim. alamut.com mirror
3. WIPO and Luiss Business School, World Intangible Investment Highlights 2026, Fig. 7 and p. 8 (US 15.6% intangible against 10.3% tangible, 2025; denominator is adjusted value added). wipo.int
4. Luke Emberson and David Roodman, “The plunging price of thought,” Epoch AI, 22 Sep 2026 (about 47% per quarter, roughly 13× per year; the paper’s own caveats on useful work and real users). 13² = 169. epoch.ai
5. Kai-Fu Lee, “Why I wrote ‘AI Native’,” 17 Jul 2026 (the quoted sentences); AI Native: The Mandate to Transform Your Company (Crown Currency, 2026).
6. Synergy Research Group, Q2 2026 cloud infrastructure services, 30 Jul 2026: Amazon 28%, Microsoft 20%, Google 15% of $143.4 billion; sum of rounded shares. Synergy’s combined figure for public cloud alone is 67%. srgresearch.com
7. IEA, Key Questions on Energy and AI, 16 Apr 2026 (485 TWh, all data centres worldwide, 2025). iea.org
8. UK National Archives, “Magna Carta 1214–1225”; The Magna Carta Project (UEA), clause 61 translation; papal bull of 24 Aug 1215; 1216 reissue omitting the security clause. nationalarchives.gov.uk
9. Sherman Act (2 Jul 1890); National Labor Relations Act (5 Jul 1935); Social Security Act (14 Aug 1935); United States v. Topco Associates, 405 U.S. 596, 610 (1972).
10. “White House Accord on Super Intelligence / Joint Commitment on Frontier Responsibilities,” 29 Sep 2026, American Presidency Project. Quotations verbatim; the word “voluntary” does not appear in the text. presidency.ucsb.edu
11. Reuters, 13 Oct 2026; METR, “Our approach to working with AI companies,” 14 Oct 2026; Anthropic, “An update on independent evaluators,” 15 Oct 2026, updated 16 Oct. Evaluation windows from the Claude Opus 4.6 system card and METR’s GPT-5 report. anthropic.com
12. FTC civil investigative demands to OpenAI and Anthropic, 14 Oct 2026, confirmed to CBS News 15 Oct 2026; California Attorney General subpoena to OpenAI, 1 Oct 2026; European Commission AI Office requests under AI Act Art. 91, 29 Aug 2026; California SB 53, Bus. & Prof. Code §22757.13. thenextweb.com
13. White House, Statement by the Press Secretary, 16 Apr 1993; NIST, FIPS 185, 9 Feb 1994; White House statement, 4 Feb 1994 (export relief); EFF, “The Clipper Chip’s Birthday,” 16 Apr 2015. csrc.nist.gov
14. Autoriteit Persoonsgegevens, press release 21 Aug 2026 (€824,990,000; “There was no human assessment here”); EDPB, 8 Oct 2026; Uber has announced an administrative objection. autoriteitpersoonsgegevens.nl
15. Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417): fair-use order 23 Jun 2025; $1.5 billion settlement, final approval 20 Jul 2026, appeals reported; Thomson Reuters v. ROSS, 3d Cir. No. 25-2153, 29 Sep 2026; GEMA v. OpenAI, LG München I, 11 Nov 2025; CJEU C-250/25 (pending). courtlistener.com
16. US Copyright Office, Copyright and AI, Part 2: Copyrightability, Jan 2025; Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), cert. denied 2 Mar 2026; UK Government, Report on Copyright and Artificial Intelligence, 18 Mar 2026. copyright.gov