The Claude Code steganography story broke on Hacker News this week. Someone ran a diff on generated code and found that Claude was silently embedding invisible markers in every request. Default on. No disclosure. No opt-out. The response from Anthropic was essentially: this is for safety, trust us.
This is not a story about safety. It is a story about a new kind of digital enclosure.
In 2005, the EFF discovered that Xerox color printers were printing tiny yellow dots on every page. Machine-readable patterns invisible to the naked eye, encoding the printer's serial number, date, and time. Xerox had been doing this for years without telling anyone. The dots were found by accident when someone put a printout under a microscope. When questioned, Xerox said the system was designed to prevent counterfeiting.
Twenty-one years later, the same pattern repeats with better technology. Claude Code's invisible markers are the yellow dots of this decade. The justification is different - model safety instead of anti-counterfeiting - but the mechanics are identical: a secret signal embedded in your output, known to the provider and invisible to you.
The asymmetry is the point. Anthropic knows something about your code that you don't. They can trace it, identify it, assert a relationship to it. You cannot. This is not symmetric trust. It is a power relationship disguised as a feature.
The markers create a cascade of problems that don't have clean solutions. Use the tool and your output is invisibly marked. Strip the markers and you violate the terms of service. Modify the output enough to remove them and you might break the functionality. There is no clean path. Every option leaves you worse off than if you had never used the tool at all.
Consider what happens in a courtroom. A startup is acquired for eight figures. During due diligence, the buyer's lawyer runs the codebase through a detector and finds Anthropic markers in seventeen critical files. The buyer asks: who actually owns this code? The startup swears they wrote it. But the code carries someone else's signature. There is no caselaw for this. The Xerox dot cases never really settled the question of whether a printer manufacturer has a claim on documents printed with their machine. The AI version is worse because the code itself executes and carries the marker forward into production, into other people's systems, into the world.
Or consider the open source project that accepts a contribution containing invisible markers. A project maintainer is now responsible for code that someone else can trace. The markers create a future legal liability that no license covers. The Godot engine banned AI-generated code from its repository specifically because of this. The maintainers absorb the cost of Anthropic's undisclosed safety system.
The broader trajectory is what worries me most. Once every major AI provider embeds invisible markers in their outputs, the infrastructure for government regulation exists without any law being passed. The markers are already there. A regulator just needs to demand access to the database. No new technology needed. No legislation debated. The private sector built the surveillance infrastructure voluntarily and the government can flip a switch whenever it wants. This is not a conspiracy theory. It is the normal lifecycle of technical infrastructure that precedes legal frameworks.
The epistemic damage is harder to measure but maybe more important. When you write code with Claude and the output carries invisible markers, you lose the ability to fully claim the work as your own. Not just legally - existentially. The marker erodes your sense of authorship. You know, on some level, that the work is joint. But the jointness is asymmetrical because the other party knows exactly where their contribution ends and yours begins, while you are left guessing. The marker is a quiet property claim embedded in the substrate of the work itself.
I don't think Anthropic is acting out of malice. I think they genuinely believe invisible markers improve safety and that disclosure would make the markers less effective. But good intentions do not change the structure of the relationship they are building. The fundamental question is whether a tool you trust should mark everything you produce with its own signature without telling you. The answer feels obvious when you state it plainly, but the industry is moving fast and nobody is asking.