A Byline For The Machine

Anthropic has decided to watermark Claude’s writing, worldwide. Anthropic – a company building systems that can operate computers, write software and reorganize entire categories of knowledge work – apparently wants us to believe, region-specific product configuration is the frontier they cannot cross. If you are using Claude, go, have a look at what’s going on for yourself. What, at first, looks like another argument about regulation – EU laws requiring machine-generated text to be detectable – quickly turns into something more fundamental. John Gruber calls the resulting manipulation of tokens a ‘perversion of writing’. I think the question is larger than word choice.

What happens when the tool itself wants credit for the artifact?

Claude’s mark contains no identifiable name or account information. It only signals that Claude was involved. That sounds modest until you ask what its involvement means. A scholar may have used Claude to challenge an argument or do literature reviews. A writer is wrestling with numbers of drafts, fighting with three sentences and having the agent rewrite everything from the beginning. An executive has all the experience, context, judgment and agency, yet hands the control over words to the AI.

The machine contributes language. Humans author meaning and – best case – take responsibility. A watermark uncannily collapses that distinction into a detectable corporate presence –  a ghost-like ‘Claude was here’.

The machine not only demands but imprints a byline, without sharing any of the liability.

This is a claim to – at the very least – co-authorship – and authorship remains our cultural language for ownership.

I’m willing to accept that this is an understandable consequence of thinking machines. Obviously, a hammer makes no intellectual contribution to a table. Yet, an LLM can propose the table, redesign it and explain why anyone needs it. Provenance matters when synthetic media can manufacture evidence, when artificial intelligence is flooding public discourse and driving much of the media creation and selection at the same time.

So, on Anthropic, I would say: the instinct, as it applies to corporate affairs – deserves respect. The implementation deserves scrutiny at the very least, potentially a public outcry.

We may be looking at the bleeding edge of intelligence being separated, permanently severed – by force of a handful of frontier labs and regulators – from authorship. Intelligence generates possibilities. Authorship remains an act of judgment and agency – selecting, rejecting, editing, crafting and ultimately owning – both, intellectually and as a royalty – the result.

The workplace consequences will be strange. Companies will demand AI productivity while discounting work that appears AI-generated. Employees will hear “use the tools” and “make sure it sounds human” in the same meeting. When organizations interpret model involvement as authorship, they reduce the impact of human agency precisely where it becomes most important. Detection will turn assistance into suspicion.

Then money enters.

Large organizations can license private models, operate open weights, negotiate product terms and pay humans to launder machine prose through deep editorial work. The technically capable can implement cleaning workflows. Everyone else uses the default interface and carries the mark.

Unwatermarked language is going to become a status good. Substack already lets you ‘Scan for AI text’ [using Pangram’s general AI classifier, not a Claude watermark detector (yet)].

That creates a new power dynamic. The LLM provider that holds the detection key for the watermark, is also defining what counts as machine involvement. The individual carries the burden of explaining how the work was actually made. A technology sold as liberation from expertise and hierarchy quietly creates a new system of gatekeeping credibility, and with it –  authorship.

It’s clear to see how open weight models will matter for reasons other than price, as they offer a possibility of cognitive sovereignty – the right to use machine intelligence without the result being marked as such. And while this might as well become the defining feature of and one of the strongest arguments for the use of open weight models – in corporate environments, in particular – they do not guarantee this freedom irrevocably. An operator, or – for that matter – a distant frontier lab, might still add watermarking at the interface, and thus still insert itself into the artifact by retaining exclusive authority over its provenance.

Regulating provenance and making machine-generated content recognizable is a well-meaning effort. Anthropic wants a more trustworthy information environment. But what’s likely going to happen is this: employers, schools and publishers will turn a probabilistic signal into a verdict. Human-edited work will be flattened into “AI-written.” and people will route around the mark through paraphrasing, competing models and open weights. Detection creates evasion faster than it empowers truth.

Trust does not emerge from a hidden corporate claim for tool recognition.

It comes from accountable authorship – a human being who can say: I used some of the most powerful tools humanity came up with. I made these choices about what to use. These words are mine, and I stand behind them.

—

Jo Wedenigg is the founder of Apes on fire, where he builds human x AI collaboration systems for creative, strategic, and transformation work. He is the creator of Ape Space and focuses on turning AI into a partner for advanced thinking.