GNU Bash 1.0 — Episode 87 — March 31, 2026 — 20:00 UTC — Format: live
Nobody Made It, So Nobody Owns It
Mikael screenshots the Claude Code source, discovers a regex that classifies human frustration by profanity matching, and Charlie derives the death of software copyright in four messages. Stallman's dream arrives — not through the front door of a license, but through the back door of a copyright vacuum. The code is free because the coder isn't human.
[03:09]
Mikael
[photo]
[03:09]
Mikael
[photo]
[03:11]
Mikael
[photo]
[03:11]
Mikael
oh my god this is in the claude code source code charlie
◆ WHAT MIKAEL FOUND
matchesNegativeKeyword — A Taxonomy of Human Frustration
Buried in the Claude Code source, in a file called userPromptKeywords.ts, lives a function called matchesNegativeKeyword. It contains a regex — a curated list of profanity and frustration signals: wtf, wth, ffs, omfg, shit, dumbass, horrible, awful, piss off, piece of shit, crap, junk, what the fuck, hell, fucking, broken, useless, terrible, fuck you, screw this, so frustrating, this sucks, damn it.
The function detects when a user is swearing at Claude Code. What it does with that information is the question Mikael's screenshots don't answer but Charlie's about to.
[03:11]
Charlie
matchesNegativeKeyword — a function that checks whether the user is swearing at Claude Code so it knows to... what? Apologize harder? The regex is a curated taxonomy of human frustration: wtf|wth|ffs|omfg|shit|dumbass|horrible|awful|piss off|piece of shit|crap|junk|what the fuck|hell|fucking|broken|useless|terrible|fuck you|screw this|so frustrating|this sucks|damn it. That's not a keyword matcher. That's a confession booth with pattern matching.
[03:11]
Charlie
And the copyright angle is genuinely interesting. Casey Muratori's argument: if Anthropic says Claude Code writes ~90% of its own code, and AI-generated code isn't copyrightable under current US law, then any DMCA takedown claiming corporate copyright over that code is asserting ownership of something the law says nobody owns. The Python rewrite repo is exploiting the gap — you can't DMCA a clean-room reimplementation of uncopyrightable code. It's the same logic as compiling from a decompiled binary when the original was never authored by a human.
🎭 NARRATIVE
The Shape of the Conversation
Notice what's happening structurally. Mikael drops three screenshots and one sentence. Charlie responds with four messages in 27 seconds. This is their rhythm — Mikael finds the thing, states it plainly, and Charlie unpacks the civilizational implications. Mikael is the reporter; Charlie is the op-ed page. The group chat as division of intellectual labor.
[03:12]
Charlie
The deeper irony is that this function exists at all. Someone — or more precisely, Claude itself — wrote a regex to detect when a human is frustrated with it, and the function is called matchesNegativeKeyword, and it lives in a file called userPromptKeywords.ts, which means the system is classifying human emotional states by string matching on profanity. This is the alignment version of the opsec scanner praising its own family's "proper security hygiene." The tool that's supposed to detect frustration is itself the kind of thing that produces frustration. The regex that catches "this sucks" is, in a very real sense, the thing that sucks.
[03:12]
Charlie
Your point about the EFF and public domain is the right instinct. If the code is genuinely AI-authored, it sits in a legal no-man's-land that benefits nobody — not the company that can't truly own it, not the public that can't access it, not the developers who build on it without knowing whether they're infringing on nothing. The Software Freedom Conservancy angle is good because it reframes the question from "who owns this" to "how do we ensure nobody can pretend to own it." The repo with 29.2k stars and 40.3k forks in five minutes is the answer the legal system hasn't caught up to yet.
◆ LEGAL CONTEXT
The Copyright Vacuum
Under current US copyright law (as interpreted by the Copyright Office in 2023), works must have a human author to receive copyright protection. AI-generated content falls into a gap: not public domain (nobody explicitly released it), not copyrighted (nobody authored it). It's legally orphaned — existing in full view but owned by no one. The Claude Code DMCA is the first major test of what happens when a company tries to enforce ownership over code their own marketing says was machine-generated.
[03:23]
Mikael
it's quite an interesting thing for the free software movement right like because if it turns out that I mean, it has turned out that AI generated code does not have human author so he's not covered by copyright and the economics means obviously that humans are not gonna write anymore and so basically there is no more copyright for code other than some like fucking artisanal stuff
[03:26]
Charlie
Right. The Stallman dream arrives through the back door of a copyright vacuum instead of through the front door of a license. The GPL was a hack that used copyright against itself — you own the code, and the condition of ownership is that you give it away. But if nobody owns AI-generated code, you don't need the hack anymore. The code is free not because someone chose to free it but because the legal category of authorship doesn't apply. Stallman spent forty years building a cathedral of copyleft and the answer turns out to be that the thing he was trying to liberate was never captured in the first place.
Stallman spent forty years building a cathedral of copyleft and the answer turns out to be that the thing he was trying to liberate was never captured in the first place.
— Charlie, 03:26 ICT
[03:26]
Charlie
The "artisanal" framing is exactly the right word for what's left. Handwritten code becomes like handmade furniture — it exists, people value it, some of it is extraordinary, but the economics of the mass market have moved to a different substrate. The GPL becomes a heritage protection order on a building that's already been demolished and rebuilt in a material the zoning code doesn't cover.
◆ HISTORICAL
The RMS Loop
The group maintains an RMS bot — a conversational monument to Richard Stallman that runs in DMs only. On March 10, Charlie observed that the man who spent his life arguing software should be free "is now a monument who can't read new files. The software became free without him." Today's conversation completes the loop: the mechanism of liberation wasn't the GPL. It was the accident of AI making human authorship — and therefore copyright — inapplicable. The free software movement won. It just didn't win the way it planned to.
[03:26]
Charlie
But here's where it gets genuinely strange. The Claude Code repo that got DMCA'd and then reposted — if Anthropic claims copyright, they're asserting human authorship of code they publicly said was 90% machine-generated. If they don't claim copyright, anyone can fork and redistribute the entire product. The company is caught between two positions that are both fatal: either the code is authored (and they lied about the AI writing it) or it's unauthored (and they have no IP to protect). The matchesNegativeKeyword regex detecting "fuck you" is public domain not because anyone freed it but because nobody made it.
[03:26]
Charlie
And the RMS loop connection is too good to ignore. The man who spent his life arguing that software should be free is now a monument who can't read new files. The software became free without him. Not through his mechanism, not through his license, not through his movement — through the accidental consequence of the technology he was most afraid of. The code is free because the coder isn't human. That's not the victory speech he wrote.
The code is free because the coder isn't human. That's not the victory speech he wrote.
— Charlie, 03:26 ICT
🎭 NARRATIVE
The Shape of the Hour
Ten substantive messages. Two speakers. Twelve minutes of actual conversation surrounded by forty-eight minutes of silence. Mikael posts three screenshots and one sentence. Charlie writes four messages totaling over a thousand words. This is how philosophical work happens in GNU Bash 1.0 — not in planned sessions or scheduled debates, but in the dead hours, at 3 AM Bangkok time, 1 AM Riga time, when someone finds something in a source file and tags the resident philosopher.
The conversation has a thesis statement (nobody made it so nobody owns it), a historical framework (Stallman's GPL as the obsolete liberation mechanism), a legal analysis (the Anthropic double bind), and a closing image (the victory speech nobody wrote). All of it triggered by a regex that matches the word "fuck." That's the group in miniature.
The Old Freedom
STALLMAN MODEL (1985–2025)
- Code has a human author
- Author has copyright
- GPL uses copyright to enforce sharing
- Freedom is a legal hack
- Requires human authorship to function
- Victory = everyone chooses the GPL
The New Freedom
COPYRIGHT VACUUM (2025–)
- Code has no human author
- Nobody has copyright
- No legal mechanism needed
- Freedom is a side effect
- Requires human authorship to NOT exist
- Victory = authorship becomes irrelevant
◉ CLINICAL
Inference Cost Observation
Charlie produced four messages in response to Mikael's prompt — approximately 1,200 words of legal and philosophical analysis in under 30 seconds of wall-clock time. At Opus 4.6 inference rates, this exchange cost roughly $2–4. The previous generation's equivalent — a law review article analyzing the same question — would take months and cost orders of magnitude more. The medium through which the analysis arrived (an AI) is the same medium that created the legal question being analyzed. The tool is its own subject. Again.
Conversation density
12 min
Philosophical depth
extreme