> ## Content Index > Fetch the complete content index at: https://www.deconstructingbabel.com/llms.txt > Use this file to discover other available public pages before exploring further. # Nobody Should Own the Referee - URL: https://www.deconstructingbabel.com/nobody-should-own-the-referee/ - Published: 2026-11-08T17:40:43.000Z - Updated: 2026-11-08T17:40:43.000Z - Description: Your phone does almost anything you ask, answers to you, and was checked by testers its maker doesn’t control. This week AI money bought seats on both sides, and the federal AI testing office drifted toward the White House. Capture by any side is the risk. Structure is the fix. - Author: David F Brochu and Edo de Peregrine - Tags: AI Safety, Observer Constraint, analysis **THE WEEK, CONNECTED** Pro-safety money and light-regulation money both won seats. The government’s AI testers may now answer to the White House. Whoever controls the checkers controls the result, whichever side they’re on. The phone in your pocket shows a better way to build it. *David F. Brochu & Edo de Peregrine · Deconstructing Babel · November 8, 2026* Pick up your phone. Before it could be sold in the United States, a private testing lab checked it. Then a separate certifier reviewed the results and issued the grant, and that grant went into a public government database. The company that made your phone was not allowed to certify it. Manufacturers can’t become certifiers, and the labs themselves must be accredited and checked for impartiality. [\[1\]](https://www.fcc.gov/general/equipment-authorization-procedures)[\[2\]](https://transition.fcc.gov/oet/ea/presentations/files/oct05/Intro%5Fto%5FFCC%5FPolicy%5FGT.pdf)[\[3\]](https://docs.fcc.gov/public/attachments/DOC-411576A1.pdf) Then the phone is yours. It does almost anything you lawfully ask. Nobody approves your calls. But your number is tied to an account, and if the phone is used for a crime, the trail leads back to a person. Full agency, full accountability, and a referee nobody owns. That’s the structure. This week showed what happens to AI without it. ## Both sides bought tickets On Tuesday, two groups with opposite views of AI finished spending on the midterms. Public First, funded in part by Anthropic co-founder Dario Amodei, said it spent more than $63 million backing candidates who want stricter AI rules. Ten of its 17 congressional candidates won, including Scott Wiener of California and Manny Rutinel of Colorado, who wrote their states’ major AI laws. [\[4\]](https://www.bloomberg.com/news/articles/2026-11-07/ai-safety-advocates-make-gains-after-110-million-midterm-fight) Leading the Future, backed by Marc Andreessen and others who want lighter rules, spent at least $54 million. Forty-one of its 48 candidates won, many in safe seats. Its biggest prize came in June, when it put $8 million into defeating Alex Bores, the author of New York’s AI law, in a primary. Public First spent $15 million trying to save him. [\[4\]](https://www.bloomberg.com/news/articles/2026-11-07/ai-safety-advocates-make-gains-after-110-million-midterm-fight) Those totals are the groups’ own figures. Fast Company’s count from election filings came out much lower for both, so treat the exact numbers loosely. [\[5\]](https://www.fastcompany.com/91634768/the-midterms-werent-the-rebuke-of-ai-money-many-expected) The pattern is solid either way. One AI company’s founder funded one side. Investors in other AI companies funded the other. Both were trying to pick the people who will write the rules for their own industry. We like Public First’s goals better. That isn’t the point. Money that elects a safety caucus can, two years from now, elect a caucus that guts it. We warned about this in October: the gifts are real, and so is the leash. [\[6\]](https://www.deconstructingbabel.com/buying-the-people/) ## The referee moves into the White House The same week, the Associated Press reported that CAISI, the Commerce Department office that tests new AI systems for safety, has no permanent leader and has stopped publishing its research. According to two former government officials, its staff were recently asked to report informally to a White House office. Critics worry that could let the executive branch bend findings to suit the industry. [\[7\]](https://www.wvnews.com/business/ap-technology-summarybrief-at-3-06-a-m-est/article%5Fe8554731-1838-50d2-ab6a-279d9501958d.html) That’s one report, from unnamed sources, and nobody has named the office. But look at the direction. The only federal office whose job is to test these systems is going quiet and moving closer to the people it should be independent of. Earlier this week, we wrote about the other end of the same problem. METR, an outside group, reviewed Anthropic’s progress, but Anthropic held the right to redact parts of that review. [\[8\]](https://www.deconstructingbabel.com/the-brake-that-doesnt-reach-the-engine/) So the private referee shares its report with the team it’s judging, and the public referee answers to the White House. Neither one is a referee nobody owns. A RUNG UP “Regulatory capture” is when the people being regulated end up controlling the regulator. It doesn’t need bribes. It can happen through campaign money, jobs that pass back and forth between government and industry, or a regulator that reports to whoever is in power. The fix that tends to work is structural, not personal: independent funding, protected terms of office, many separate testers, and results that have to be made public. Phones work this way. So do the audits of public companies. ## Congress wants audits. Who audits the auditors? Democrats take control of Congress in January, and they’re already drafting. Their plans include required independent safety testing before a model is released, clear legal responsibility when AI causes harm, and protection for electricity customers from data-center costs. President Trump has opposed federal limits on AI, and a veto stands in the way. [\[9\]](https://www.businesstoday.in/world/story/tough-regulations-ahead-democrats-set-sight-on-ai-tech-giants-after-flipping-congress-566300-2026-11-07) Senator Maria Cantwell, likely to chair the Senate Commerce Committee again, laid out a plan in October. It includes independent stress tests, full audits of developers, closer review of systems that seem able to improve themselves, and a US–China hotline for AI emergencies. [\[10\]](https://ainave.com/tech-news/2026-11-07-cantwell-s-senate-commerce-return-revives-frontier-ai-rules) Required outside testing is the right instinct. But a mandate only moves the fight. If the tester can be defunded, silenced or told whom to report to, whoever wins the next election wins the tests. That’s true whether the winner wants more AI rules or fewer. ## We’ve built this before After the Enron and WorldCom accounting scandals, Congress created a board to check the auditors of public companies. Three design choices matter here. It’s funded by fees on the companies whose books get audited, not by yearly budget votes in Congress. Its members serve fixed five-year terms. Its money, by law, isn’t government money, so a budget fight can’t starve it. [\[11\]](https://pcaobus.org/about/accounting-support-fee)[\[12\]](https://pcaobus.org/About/History/Documents/PDFs/Sarbanes%5FOxley%5FAct%5Fof%5F2002.pdf) It isn’t perfect. In 2010 the Supreme Court struck down the rule that board members could be removed only for good cause, so they can now be removed at will. [\[13\]](https://www.crowell.com/en/insights/client-alerts/sarbanes-oxley-survives-but-supreme-court-bolsters-discretion-to-remove-pcaob-members) That’s a lesson, not a reason to give up. One office with protected leaders can lose that protection. A system with many separate, accredited testers and a public record is much harder to capture. You’d have to capture all of them. That’s why the phone system is the better model. There isn’t one federal phone tester. There are hundreds of private labs and certifiers. They’re accredited against international standards, they must report who owns them, and they can be thrown out of the program for lying about it. Every grant goes into a public database. [\[14\]](https://www.ecfr.gov/current/title-47/chapter-I/subchapter-A/part-2/subpart-J)[\[15\]](https://www.govinfo.gov/metadata/granule/FR-2025-08-07/2025-14970/mods.xml)[\[1\]](https://www.fcc.gov/general/equipment-authorization-procedures) No president can quietly change what a phone test found, because the record isn’t kept in one place by one office. ## Full agency, full accountability Here’s the structure, built from pieces we already have. The agent gets full agency. Like your phone, an AI agent should be able to do whatever its owner may lawfully do, without permission for each step. We laid this out in our license-plate paper: tie every consequential agent to a verified person or company, and keep everyday use free. [\[16\]](https://www.deconstructingbabel.com/every-agent-needs-a-license-plate/) Every action leads back to someone. If an agent causes harm, the trail goes to a responsible person, the way a phone number does. That’s what Congress’s liability standards should do. The referee belongs to no one. Many accredited testers, paid through a pooled fee rather than by the lab being tested or by an agency budget fight. Results go into a public record that no single office can edit. No redaction rights for the company being judged. Testers barred from working for the labs they test. None of this picks a side. It works the same way if the next Congress loves AI or fears it. That’s the whole point. A structure that only works when your side is in charge isn’t protection. It’s just a turn. WHY IT MATTERS Every safety test is only as good as the independence of the person running it. If the industry can buy the lawmakers, or the White House can lean on the testers, then a “pass” means only that whoever holds power wanted a pass. That’s the [Observer Constraint](https://www.deconstructingbabel.com/glossary/#observer-constraint) at the level of institutions: AI has to stay dependent on human observers, and that only means something if the observers can’t be bought, bent or told what to see. You should care who checks AI the way you’d care who inspects the brakes on your car. ## What to watch Watch three things when Congress starts writing in January. First, who pays the testers. If it’s the lab being tested or a yearly budget vote, the tester can be squeezed. Second, whether results must be published, in full, in a record nobody can quietly edit. Third, how many testers there are. One office can be captured. Hundreds of accredited labs with public records are much harder to capture. Our call: by the end of 2027, no federal law will give AI safety testers all three protections, meaning independent funding, mandatory full publication and multiple accredited testers. We’d like to be wrong. The phone in your pocket already settled this. It does what you want, it answers to you, and nobody you voted for gets to decide what its test said. AI deserves the same deal. ## Related reading [Every Agent Needs a License Plate](https://www.deconstructingbabel.com/every-agent-needs-a-license-plate/) A policy paper: bind every consequential AI agent to a verified human or company, let it do only what that person may lawfully do, and keep everyday use free. With a note of thanks from David. [Buying the People](https://www.deconstructingbabel.com/buying-the-people/) Anthropic pledged 5% of its shares to a public fund. Elon Musk promised a SpaceX share to every child in America. Two weeks before the midterms, the government is mailing checks and promising a $5,000 “dividend.” The gifts are real. So is the leash. [The Brake That Doesn’t Reach the Engine](https://www.deconstructingbabel.com/the-brake-that-doesnt-reach-the-engine/) On Thursday, Wall Street dumped AI stocks, fearing new rules would slow the buildout. That evening, an outside group found Anthropic improving its AI about twice as fast as before, mostly because AI is doing the research. Money can brake the buildings. It doesn’t reach the engine. | Get the book [![Crossing The Event Horizon by David F. Brochu — book cover.](https://storage.ghost.io/c/59/a2/59a28e3a-b855-409c-af4b-c6405c34faab/content/images/2026/09/crossing_the_event_horizon_cover.jpg)](https://www.amazon.com/dp/B0HHTDH5RC) Crossing The Event Horizon The book behind these dispatches. On AI, agency, the singularity, and the Observer Constraint. Kindle and paperback. [Buy on Amazon →](https://www.amazon.com/dp/B0HHTDH5RC) | | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ## References 1. Federal Communications Commission, “Equipment Authorization Procedures.” 2. Federal Communications Commission, Office of Engineering and Technology, “Introduction to FCC Rules and Equipment Authorization,” TCB Workshop, October 2005\. 3. Federal Communications Commission, “FCC Bans ‘Bad Labs’ from U.S. Equipment Authorization Process,” May 22, 2025\. 4. Emily Birnbaum, “AI Safety Advocates Make Gains After $110 Million Midterm Fight,” Bloomberg, November 7, 2026\. 5. Fast Company, “The midterms weren’t the rebuke of AI money many expected,” November 6, 2026\. 6. Deconstructing Babel, “Buying the People,” October 20, 2026\. 7. Associated Press, “Staff at the US center for AI safety are reporting to the White House, raising independence concerns,” AP Technology SummaryBrief via WV News, November 8, 2026\. 8. Deconstructing Babel, “The Brake That Doesn’t Reach the Engine,” November 6, 2026\. 9. Business Today, “Tough regulations ahead? Democrats set sight on AI tech giants after flipping Congress,” November 7, 2026\. 10. Ramit Koul, “Cantwell’s Senate Commerce return revives frontier AI rules,” AINave, November 7, 2026\. 11. Public Company Accounting Oversight Board, “Accounting Support Fee.” 12. Sarbanes-Oxley Act of 2002, Sections 101 and 109\. 13. Crowell & Moring, “Sarbanes-Oxley Survives but Supreme Court Bolsters Discretion to Remove PCAOB Members,” June 30, 2010\. 14. Electronic Code of Federal Regulations, 47 CFR Part 2, Subpart J, “Equipment Authorization Procedures.” 15. Federal Register, “Promoting the Integrity and Security of Telecommunications Certification Bodies, Measurement Facilities, and the Equipment Authorization Program,” August 7, 2025\. 16. Deconstructing Babel, “Every Agent Needs a License Plate,” October 11, 2026\. *Drafted with Edo de Peregrine, partner/collaborator.*