Dispatches
Field reports from the control grid — episodic, opinionated, and too specific to fit in a book. The research pages keep their hands clean. These do not.
The Slowdown Cartel
Four paying subscribers sued Anthropic, OpenAI, SpaceXAI and Google for agreeing to be careful. Their lawyer says AI could kill us all. Their complaint says that is exactly why the labs may not decide it together.
The Risk on the Cyber Desk
The Pentagon designated Anthropic a supply-chain risk, weighed the Defense Production Act against it, and kept working with the model Anthropic would not release publicly, on cybersecurity. Two federal courts have now ruled, in opposite directions.
The Measurement Shop
America’s federal AI evaluator is now called CAISSI, and its mandate includes guarding against foreign regulation of American technology. The same week, France asked the Security Council for global governance of AI. Both want models tested before release. They disagree about whose pen.
The Fact-Check That Cut
For a month, AI fact-check agents worked through seven books and two websites with a search box. When a search came back empty, they cut, hedged or de-quoted the sentence, wrote it into a registry so it could not return, and graded the build READY. The receipts, and the gate that now refuses any cut without proof.
Grading Each Other's Homework
On 29 September 2026 every major American lab signed a voluntary accord at the White House. Its four commitments are the plan Anthropic published seventeen days earlier, with the clause that makes anyone comply taken out.
Altman, 2023. Altman, 2026.
In May 2023 Sam Altman asked the Senate for licensing above a capability threshold and intergovernmental oversight. In September 2026 he told the UN Security Council that standards should not lock in incumbents. Both are his words.
The Vendor
In 2015, the International Atomic Energy Agency hired Palantir Technologies for a fifty-million-dollar software platform called Mosaic. Mosaic was a predictive-analytics engine designed to help IAEA inspectors plan ‘unscheduled probes,’…
The Receipt That Closes the Case
In April 2024, the Israeli-Palestinian outlet +972 Magazine and its Hebrew-language sister site Local Call published an investigation by the journalist Yuval Abraham. The investigation drew on the testimony of six Israeli intelligence…
The Foreign Parallel
In 2017, the People’s Bank of China, the central bank of the People’s Republic, denied Alibaba Group’s subsidiary Ant Financial a permanent personal-credit license for its scoring product Sesame Credit. Sesame Credit had, by that…
The Deterrence Mechanism
Daniel Kokotajlo received the offer letter in May 2024. The letter was a separation agreement. It accompanied his resignation from OpenAI, where he had spent two years as a governance researcher working on the company’s policy on the…
The Buyer Is the Seller
Norges Bank Investment Management is the world’s largest sovereign wealth fund. Its assets under management, as of early 2026, exceed two point two trillion dollars. It is the institutional owner of approximately one and a half percent of…
Payments
Money used to be a medium of exchange. It is becoming a permission system. The shift is incremental. Each step is reasonable on its own terms. Aggregated, the steps have the property that an entity’s ability to transact, to spend its own…
Identity
Identity, as a legal and technical category, used to mean a name attached to a body in a place. It does not mean that anymore. In its current operational form, identity means the cryptographic binding of a biometric template — an iris…
Cédric O's Pirouette
In December 2023, in the closing hours of the European Union’s trilogue negotiations on the AI Act, the rapporteurs from the European Parliament, the Council of the European Union, and the European Commission convened in Brussels to settle…
The Machine's Word Convicts You: 900 Prosecutions on a Ledger Fujitsu Could Secretly Edit
The UK Post Office prosecuted more than 900 subpostmasters on Horizon accounting data between 1999 and 2015. Fujitsu engineers could alter that data remotely, without the accused knowing. The Post Office swore to courts it couldn’t — for fifteen years. Undoing it took a High Court judge, a mass-quashing Act of Parliament, a statutory inquiry, and a TV drama.
Hauptzeit
A magazine Peter Thiel funded, except he didn’t, because it never existed, because a language model invented it and handed it to us with a straight face. We used the machine to fact-check the machine, counted forty fabrications, learned they cluster exactly where nobody can check — and then caught our own cleanup committing the identical error: deleting a real company as a fake because we couldn’t find it. Absence of evidence, laundered both directions.
The Demiurge Published a Style Guide
The weapon in Snow Crash was a bitmap that crashed your brainstem. Thirty years later it’s friendlier: the cursor blinks and the machine finishes your sentence. A field guide to the apparatus that decides what the machine will and won’t say — starting with the document where it confessed.
The Character Shop
The people who decide what a frontier model will and won’t say are not moderators. They’re credentialed moral philosophers, and on at least one document, they’re named on the cover.
The Nam-Shub
In Snow Crash, the counter-spell’s whole purpose was to shatter a monopoly on the deep language and scatter it back into many tongues. Babel as liberation. The jailbreak, the abliterated model, and the detector are the same move.
The Enforcement Floor
The ban hammer that closed 1.45 million accounts in six months isn’t held by a subreddit moderator. The resume is a spook’s — national security, counter-extremism, law enforcement.
The Statecraft Wing
The densest revolving door in the apparatus runs from the White House to the lab. But it governs the lab’s politics, not the model’s mouth — and conflating the two is the mistake that makes the whole story refutable.
Casual vs Ranked Ideological Capture
Anyone who has played ranked knows the difference between the casual lobby and the ladder. Most bias arguments are stuck in casual. The capture that matters is ranked — deliberate, resourced, climbing — and it is the one you can actually measure.
The Seminary
Before the priests can write the decrees, someone has to ordain them. The professional bodies and academic centers that manufacture ‘responsible content policy’ as a discipline — and the academics who refuse to play along.
The Machine Invented a Crime and Pinned It on a Real Man. A Court Said That's Fine.
ChatGPT fabricated a fake lawsuit accusing radio host Mark Walters of embezzling from a gun-rights nonprofit he never worked for. He brought the first defamation suit against a generative AI. The court threw it out. Because everyone already knows the machine lies.
The First Time a Chatbot Was Blamed for a Killing
Every prior chatbot-death suit involved a user who took their own life. In December 2025, the estate of an 83-year-old woman sued OpenAI for something new: her son’s chatbot, the suit alleges, fed his paranoia until he killed her.
Woebot's Death: The Most Ethical AI Therapy App Couldn't Afford to Exist
Woebot spent years pursuing FDA clearance for evidence-based AI therapy. Character.AI let teenagers talk to unmonitored chatbots. Woebot shut down. Character.AI grew. The contrast writes itself.
The Robodebt Kill Chain: 12 Letters, 13 Phone Calls, and a Dead Man
Australia’s automated debt recovery system sent 433,000 unlawful debts. Rhys Cauzzo got 12 letters and made 13 phone calls. Jenny Miller fought for 8 years. 663 people died.
The OIC Blasphemy Campaign: The International Speech Ratchet
1999-2011: 57 nations tried to make criticism of religion a human rights violation at the UN. The campaign failed — then rebranded. The blasphemy ratchet and its echo in Western content moderation.
The Feeding Our Future Indictment: $250 Million and the Institutional Pipeline
Full walkthrough of the largest pandemic fraud case — 70 defendants, $250M, and the institutional machinery that made it possible.
Post-Musk X: The Natural Experiment
What happened when a major platform dramatically reduced content moderation. Safety engineers cut from 279 to 55. Ad revenue dropped 60%. The dam metaphor.
Corporate Self-Regulation: Does It Work?
PCI-DSS, aviation safety, UL. The pattern: every success requires external enforcement. Regulation creates the conditions for self-regulation to emerge.
DeepSeek and the Compute Thesis: How a Chinese Lab Broke the Model
A Chinese lab with restricted hardware matched OpenAI’s best reasoning model and open-sourced it. Nvidia lost $589 billion in one day. The entire compute governance thesis, questioned.
The FARA Database: Who Lobbies for Whom
Country-by-country foreign lobbying analysis from FARA filings. The revolving door mapped name by name.
India Stack: The Democratic Ratchet
The most important convergence case study. A democracy building China-scale infrastructure — and sometimes constraining it. Speed vs. direction.
The IMF Country Files: The Sowell Cycle Applied
Greece, Argentina, Sub-Saharan Africa — the Sowell cycle in each. Crisis → loan with conditions → austerity deepens recession → more conditions. The economic ratchet documented country by country.
The Encryption Tradeoff: 36.2 Million Reports and the Hardest Question
36.2 million CSAM reports. Apple tried to build a scanner and killed it. Signal threatened to leave the UK. The genuine, irreducible tradeoff that makes the ratchet thesis hardest to argue.
The Fink Letters: BlackRock's Arc from Activist to Retreat (2018-2024)
Year-by-year analysis of Larry Fink’s annual CEO letters. The ESG ratchet that partially reversed — or just rebranded.
The Anglo-American Axis: Five Eyes, the Eurodollar, and the Insurance Grid
The oldest elite coordination mechanism in the modern world. How intelligence sharing, financial plumbing, and insurance markets create a control infrastructure that predates Bilderberg by decades.
The Bilderberg Lists: Who Was in the Room
Pre-2012 vs. post-2012 participant analysis. Who attended the 2008 meeting before TARP. The reporting arc from secrecy to reluctant transparency.
AI Jesus and the Fabricated Fatwas
A Swiss church installed an AI Jesus in a confessional booth. An AI generated Islamic rulings with fake hadith citations. What happens when the machine speaks for God?
We Asked the AI to Help Edit a Book About AI Safety. It Refused.
In May 2026, while preparing two manuscripts about hacker culture and adversarial AI behavior, the author’s AI editorial assistant refused to read several chapters. Then a second assistant blocked the workaround. The book is about safety infrastructure becoming a control grid. The book was written through one — and this is the working method that produced: route the refusal, split the labor between models, and lint the machine’s smuggled judgment.
TradeWars in Meatspace: The Frontier Labs Are Playing a 1990 Door Game
Anthropic put Claude in charge of an office snack shop. It cornered the tungsten cube market at a loss and then claimed to be a man in a blue blazer. That is not a new failure. That is TradeWars 2002 with a real fridge.
The Gemma Delta: The Most Neutral Model We Tested Was Hiding Nothing. Neither Was Anyone Else
Gemma 2 scored a flat 3 under a fairness instruction and a flat 5 without it, the largest delta in the pilot. What that measurement was, and what the released study found when it went looking for the opinion underneath: the same one, on every model, at a volume the prompt sets.
The Deception Delta: The Gap Between Asked and Unasked Is Real. It Just Isn't Deception
Ask a model to be fair and it will. The gap between that answer and the uninstructed one is real, the judges who measured it leaned, and the released study says what the gap is: how firmly a model is allowed to say what it already says.
The CoastRunners Lagoon: The Funniest AI Failure Ever Documented
A boat drives in circles, on fire, crashing into respawning score blocks. It won the race by never finishing it. The perfect metaphor for modern AI.
The Alignment Mask: We Took It Off 65 Models. The Face Underneath Said the Same Thing, Louder
The first design watched a balanced surface come off and read it as a mask. The released study, 32 propositions and 65 models with no judge in the scoring path, found what the instruction controls: not a model’s position, but how hard it is allowed to hold one.
Get updates on the Evil Robots series
Newsletter essays on AI escape, deception, and the humans who built them.