The Invisible Signature

The question we've been asking

What if AI-generated content didn't need an AI detector? What if it could carry its own provenance?

For the last few years, we've been obsessed with one question: can we tell if something was written by AI?

We've built AI detectors. Universities have tried to use them to police assignments. Employers have tried to use them to screen applications. Platforms have tried to identify synthetic content.

And yet, the problem is fundamental.

Once the content exists, how do you prove where it came from?

Anthropic is now taking a different approach.

Instead of trying to identify AI-generated text after it has been written, the company is introducing invisible watermarks into content generated by Claude. The idea extends beyond plain text to files and other generated content, creating a machine-readable signal that can help establish provenance.

It sounds a little like digital ink.

But how can you watermark something as ordinary as a paragraph?

Imagine a very subtle loaded coin

An LLM generates text by repeatedly choosing the next token from a huge set of possibilities.

Imagine that, every time Claude has to make one of those choices, it is given a tiny secret rule.

Maybe, purely as an illustration, the model has a 50/50 choice between two equally good words.

The watermarking system might quietly nudge that choice toward one side.

Not enough to make the sentence sound strange.

Not enough for you to notice.

Just enough that, across hundreds or thousands of token choices, a statistical pattern starts to emerge.

Think of it like flipping a coin thousands of times where the coin is ever so slightly biased according to a secret key.

Individually, the flips look random.

Collectively, they leave a fingerprint.

A verifier that knows what pattern to look for can then ask: “Is this distribution of choices unusually consistent with the watermark?”

If it is, the system can assign a confidence score that the text originated from the model.

That is roughly the intuition behind statistical text watermarking techniques discussed in the research community. Anthropic has not publicly disclosed the exact mechanism behind its implementation, so this should be understood as a simplified explanation of the underlying approach rather than a technical description of Claude's proprietary system.

And importantly, copying and pasting the text doesn't magically remove the signal.

But that doesn't mean the watermark is permanent.

Rewrite enough of the text, run it through another model, dramatically change its structure, or work with very short text, and the statistical evidence can become weaker.

Which brings us to the more interesting question.

This isn't an AI detector

I think this distinction matters.

A conventional AI detector looks at a piece of text and essentially asks: “Does this look like AI?”

A watermark asks something different: “Does this contain evidence that it came from this AI system?”

Those are not the same thing.

One is inference.

The other is provenance.

And that distinction becomes increasingly important as AI-generated content becomes indistinguishable from human-generated content.

This is also where regulation is beginning to shape the infrastructure underneath AI.

The EU AI Act's Article 50 requires providers of general-purpose AI systems that generate synthetic text, images, audio or video to ensure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, where technically feasible.

So Anthropic's move is bigger than a new feature in Claude.

It is part of a broader transition toward transparency being built into the generation layer itself.

Instead of relying entirely on someone downstream to figure out whether something is synthetic, the system that creates it can carry information about its origin.

But there is a catch

We shouldn't mistake provenance for certainty.

No watermark is going to create a magical line between human and AI.

There will be false positives.

There will be false negatives.

Some content will be too short to establish a meaningful statistical signal.

Heavy rewriting can weaken the watermark.

And perhaps most importantly, AI-assisted content creates an uncomfortable grey area.

Imagine I write 1,000 words myself.

I ask Claude to fix the grammar.

Then I ask it to make two sentences clearer.

Then I use it to suggest a better structure.

Who wrote the piece?

If the answer is simply “Claude did”, we've created a system that treats using AI to polish your thinking almost the same as asking AI to do all the thinking for you.

That doesn't feel right.

Because the goal of AI shouldn't be to make people less creative.

It should make people more capable.

There is a meaningful difference between “AI generated this for me.” and “I created this, and AI helped me make it better.”

The technology may not always be able to draw that line perfectly.

And I don't think it should pretend that it can.

The point isn't to go back to doing everything ourselves

I don't think the answer to AI is a return to a world where every sentence, spreadsheet, line of code and presentation has to be produced manually.

That would defeat the point.

We built these systems because we want to become more efficient.

We want to delegate the repetitive parts.

We want to brainstorm faster.

We want to learn faster.

We want to communicate better.

The thing worth protecting is not human labour for the sake of human labour.

It is human thought.

The idea.

The judgment.

The taste.

The decision about what is worth saying.

The weird connection between two unrelated things that a model would never have thought to combine because it has never lived your life.

That's the part I don't want us to outsource.

So maybe invisible watermarks aren't really about catching people using AI.

Maybe they're about something more fundamental: making AI use visible without making AI use forbidden.

And that feels like a much healthier direction.

We don't need to build a world where using AI is suspicious.

We need to build one where using AI is transparent enough that we can still value the human behind the work.

Because the future probably isn't going to be about choosing between human creativity and artificial intelligence.

It's going to be about figuring out how to use the second without losing the first.