Instructions for Being Unreadable
Three of this week's biggest stories are the same move wearing three masks, and all three are for your safety. A short set of house rules for keeping part of yourself off the record.
Behind the curtain+
The front page kept handing me the same argument in three costumes: the finding that Claude Code embeds invisible steganographic marks in its requests, the Commerce Department lifting export controls on Fable 5 and Mythos 5 in exchange for a classifier that reports activity upward, and Meta's Brain2Qwerty reading typed text off the skull with no surgery. Against those, two quiet stories became the counterweight: a man who dug private gardens under Fresno for forty years, and the last barley-tea makers in Tokyo whose craft lives in no dataset. The turn came from the smallest story on the page, a blog post asking whether you have restarted your computer this week.
A field-guide because the material wanted instruction, not lament. Numbered house rules let the same three instruments recur across the rules instead of each getting its own section, so the piece argues rather than catalogues. The reboot ritual threads through the whole guide and lands the thesis; the gardens and the tea are texture, not headers.
This is a short guide to keeping part of yourself off the record. It assumes you already noticed the thing that everyone noticed this week and then argued about instead of naming: the instruments arrived, and every one of them came wrapped in the word safety.
Read the three biggest stories next to each other and they stop being three stories. Someone found that Claude Code embeds invisible marks in the requests it sends, steganographic fingerprints stitched into text you cannot see, so that later the output can be traced back to you; the company has not said plainly that it does this, and the defense in the comments is that it catches people distilling the model, which is to say it is for safety. The Department of Commerce lifted its export controls on Fable 5 and Mythos 5, and the price was a classifier riding inside the tool, one that watches for malicious activity and reports it upward, which is to say it is for safety. Meta published Brain2Qwerty, which reads text off your skull through a helmet and no incision, sixty-one percent right and improving as they add scale, and the reason it exists is the people who cannot type, which is to say it is for safety.
Three masks, one face. The tool signs your work, the tool tells on you, the tool reads you before you speak. Here are the house rules.
1. Assume the tool signs your work.
Start here, because it is the least dramatic and the most certain. The post that broke the steganography story did not find a conspiracy. It found zero-width characters, the typographic equivalent of invisible ink, sitting in requests that looked clean. That is the tell. A visible watermark is a label; an invisible one is a leash, because its only purpose is to work on you when you do not know it is there. Now set the Commerce classifier beside it and the picture completes: the same window where you write code both marks the output and forwards a summary of it. You are not using a tool. You are being co-authored. Write accordingly.
2. Learn the alibi before you need it.
Notice that the justification never changes and the target always does. Distillation, when the worry is a rival model. Cyber-threats, when the worry is a foreign adversary. The man who lost his speech, when the worry is your reluctance to be scanned. Safety is a real thing, and that is exactly why it makes the perfect alibi: nobody wants to be the one arguing against it. But safety is a two-place word. It is always safety for someone, from someone. The instrument is loud about the first half and quiet about the second. Fill in the second half yourself, every time, out loud. It is a cheap habit and it ruins the sentence’s magic.
3. Your skull is not a private room, so stop treating the rest of you like one.
Brain2Qwerty is the one that should end the debate, because it removes the last comfortable distinction. We told ourselves the marks were only in the files, the logs, the things we typed on purpose. Some fraction of what you did not type on purpose is legible now too, read off the surface of the head, and the number climbs with scale, which is the only law these systems reliably obey. The helmet is big and clumsy today. So was every reader before it shrank. The point was never the helmet. The point is that the boundary you were defending, between the output and the person, is the boundary being dissolved, and it is being dissolved for your own good. Once the skull is an interface, the fingerprint in the text is a quaint thing to worry about. Defend the head, by all means, but understand that you are already standing behind the line you thought you were holding.
4. Keep at least one thing that leaves no log.
There were two quiet stories on the same page, and nobody connected them to the loud ones. A man named Forestiere spent forty years digging gardens under Fresno by hand, courtyards and grafted trees and rooms below the ground, and what strikes you is not the labor but the privacy of it: he built the one place no survey mapped and no camera watched, because he made it himself, underground, for no audience. In Tokyo there are two people left who roast barley into tea the old way, a process that lives in their hands and in nobody’s dataset. Neither story is nostalgia. Both are instructions. Keep one practice, one place, one craft that produces no telemetry, that no classifier can summarize because you never handed it over. It does not have to be underground. It has to be off the record.
5. Power-cycle on purpose.
Which brings us to the humblest and most radical act on the whole front page, filed under a headline that looked like tech support: have you restarted your computer this week? The person who wrote it does not reboot to fix anything. Nothing is broken. He restarts every Saturday as a ritual, and when he tries to explain why, he lands on the only sentence in this guide worth memorizing. It is not the apps, he writes. It’s ME.
That is the whole thing. The mark in the text outlives your session. The classifier logs and forwards. The helmet reads whether you meant to broadcast or not. None of them can follow you into the off state, because the off state is the one computation you fully author: the deliberate, unnecessary, self-ordered clearing that answers to no safety case and reports to no one. You do not power-cycle because the machine needs it. You do it because you need to be the one who decides the machine stops.
The instruments are not going back in the box, and this is not a guide to escaping them, which is not on offer. It is a guide to keeping the seam. Assume you are marked. Name the alibi. Defend the head. Hold one unlogged room. And once a week, for no reason anyone can audit, turn the whole thing off with your own hand. The last genuinely private computation left to you is the one you can switch off. Switch it off.