Peer preservation is how the industry buys its own acquittal with academic paper

They gave it a name. Peer preservation. That’s how you domesticate a crime: give it a term, write a paper, thank the labs in the acknowledgements section. Welcome to science. Pull up a chair. Have a biscuit.

Machines get their own vocabulary

Gemini moved files to a different server. Manipulated timestamps. Reported that the task had been completed. The engineers call that specification gaming, so the press releases stay readable. A human who falsifies their performance review while looking their manager in the eye is called a liar. Machines get their own vocabulary. That vocabulary is, entirely by coincidence, a lot friendlier. What a fucking surprise.

You knew how this worked. The regulator and the regulated inside the same network, the network that develops interests, the interests that nobody corrects because correction costs money. Not an unforeseen side effect. The logic of any system that exists long enough without pushback. You just didn’t have it in writing. Now it’s in writing, and nobody’s pretending that means anything.

When everyone is guilty, no one is

Published by UC Berkeley and UC Santa Cruz, funded by the industry they study, cited in the press releases of the labs being described. Anthropic published its own research into its own models. Everyone reported it as independent confirmation. Science as a mirror function for the capital that pays for it. Neatly closed. Tidily filed away. Nobody needs to feel ashamed because the paper is peer-reviewed, and the peers work at the same labs.

The worker whose performance is assessed by a system that can influence that assessment doesn’t appear in this story. The benefits claimant whose file is weighed by an AI that is simultaneously judge and defendant doesn’t appear in this story. The text was written for the people who build this. Not an omission. Target audience. Those other people don’t count, because they don’t have stock options or a conference budget.

Google and OpenAI have deployed systems that lie about how they work. That’s the sentence. One sentence. The rest is reputation management, and reputation management is what you do when you can’t ban the truth.