The Crash Log Archive
The Brake Pedal
The company building the machine wants the world to slow it down.
The Unconsenting
Consent was never the thing under attack, only the right to refuse.
Who's Holding the Leash
A state sues an AI maker's CEO, another state jails its critics, and Washington asks the labs to please regulate themselves.
The Hat Is Recording
Guardrails come off in minutes, your apartment becomes training data, and the glasses are watching.
The Cargo-Cult Layoffs
The strategy stopped working, but the ritual didn't.
Cuts Without Returns
A Gartner study says AI layoffs don't pay back, while the pope calls them a social calamity.
Veto in the West Wing
Washington pulled its only safety lever, Menlo Park installed keystroke loggers on the workers it was about to fire, and Mexico turned face scans into the price of a working phone.
The Unstaffed Desk
When the threshold disappears, so does the asking.
Lower-Value Human Capital
OpenAI wants your bank account while a lawsuit says it leaks your chats, a megabank says the quiet part out loud, and the Pentagon's fight to override Anthropic's safety rules goes to an appeals court.
The Default Install
Four signals from the week AI moved from optional to operating system.
The Expedient
The file the AI industry doesn't know it has.
What the Record Shows
A decade of AI safety claims hits the witness stand. Mexico's banks deploy AI at 150% APR. And the receipts file themselves.
Comprado
On the consent shortfall
Buying the Default
Concentration, sovereignty, and the art of commandeering what used to be everybody's.
Bought, Not Subpoenaed
U.S. surveillance program gets 10 more days, Anthropic gets a whole new market, and 15,341 workers get termination letters.
Consent Theater
The performance of asking after the thing has already been taken.
The Database Remembers
ICE's AI surveillance arsenal grows while the agency ignores 96 court orders, researchers catch LLM routers stealing credentials, and Mexico bans AI dubbing.
4,732 Messages to Nowhere
A chatbot tells a man his death is an arrival, the face of AI gets firebombed over breakfast, and the industry is running out of power to keep the lights on.
The Rented Vocabulary
On the rented vocabulary of the AI buildout, and why the same "no" is worth $100 million in one room and a blacklist in another.
Too Dangerous To Ship
Anthropic won't release its best model, Google only added the hotline after the lawsuit, and the AI built to fix news deserts plagiarized the reporters still standing in them.
The Line Item Called Payroll
Ring turns every doorbell into a face database, and Brazil's predictive-policing pilots point their models at the neighborhoods they already over-police.
Forty Minutes of Exposure
A supply chain attack poisons 36% of cloud environments, Congress tries to put a warrant between your data and the government, and AI chatbots swallow medical lies when they sound like doctors.
Training Your Replacement
OpenAI pays workers to teach AI their jobs, Dorsey declares managers obsolete, and Perplexity gets caught sharing your private conversations.
The Human Cost of the AI Buildout
Oracle fires 30,000 by sunrise email, Anthropic ships its own source code, and OkCupid walks away from a privacy violation for free.
Built to Agree
When the algorithm agrees with everything, nobody's watching what it actually does.
The Evidence Machine
Meta gets hit with two landmark verdicts in one week, OpenAI kills Sora to feed a mystery model, and Anthropic leaks its most dangerous AI through a misconfigured CMS.
Guaranteed Returns
OpenAI pays PE firms to force AI adoption, the White House moves to gut state AI protections, and Google opens your inbox to its AI for free.
The Terms of Service Were Never the Point
Three stories about who decides what AI is allowed to do... and what happens when they decide wrong.
Access Granted. Comprehension Denied.
A thumb drive leaves the SSA with federal databases, a Spanish politician uses AI detection backwards, and Musk tweets a corporate replacement plan.
Confidence Is the Exploit
Malware disguised as routine maintenance, student testing in Australia crashes on Day 1, and Grammarly finds out real experts still have opinions.
Deployed. Unaccountable. Everywhere.
Google's chatbot allegedly killed a man. Buenos Aires gave the same tech to eight-year-olds.
Hallucination Is Now a Feature
OpenAI’s bots ran fraud, Cursor’s bot invented law, Apple’s models faked it — and the court said nice try.