Claude Can Tell Us AI Was Involved. Now What?

Anthropic has begun introducing invisible watermarks into text generated by new Claude models.

And as an educator, I can't decide how I feel about it.

On the surface, I understand the appeal.

For the past few years, educators have been trying to answer what seems like a relatively simple question:

Did a student use AI to write this?

We've seen AI detectors attempt to answer that question, often with results that are far less definitive than the percentages displayed on the screen might suggest.

Watermarking takes a different approach.

So what is an AI watermark?

In very simple terms, Claude creates a statistical pattern through the words it chooses when generating text.

There aren't hidden characters inserted into the document. You can't highlight the text and find some secret Claude code buried inside.

Instead, Claude's word choices create a pattern that a detection system can potentially recognize later.

In other words, rather than asking whether something looks like AI writing, we may have a signal that Claude was actually involved in producing the text.

From the teacher's perspective, I can immediately see the appeal.

Finally.

Maybe we have something more meaningful than an AI detector making a prediction.

Maybe we have greater transparency.

Maybe educators have another tool for navigating academic integrity in an AI world.

But then I put myself on the other side.

What does this look like from the student's chair?

Imagine you're a student.

You spend hours writing an essay.

The argument is yours. The examples are yours. The experiences are yours. The writing is yours.

Then you open Claude.

You paste in your essay and ask:

"What edits would you suggest to make my argument clearer?"

Claude gives you feedback.

You review the suggestions. You decide which ones you agree with and which ones you don’t. Then Claude generates a revised version incorporating those approved changes.

Now Claude technically has generated the text.

And that version could carry Claude's watermark.

So here's the question:

What exactly have we detected?

AI authorship?

Or AI involvement?

Because those are not the same thing.

"AI was involved" doesn't tell us how AI was involved

This is the part I think schools are going to have to wrestle with.

Consider two students.

One student says:

"Write me a five-paragraph essay about The Great Gatsby."

Another student writes the entire essay and then asks:

"What are some way you help me make my argument clearer?"

Both students used AI.

But educationally, those two uses are completely different.

One potentially outsourced the thinking.

The other may have used AI as a writing coach.

A watermark might help us identify that AI touched the final product.

But it doesn’t tell us how.

Does watermarking solve the detection problem, or just change it?

AI detection is already coming under fire in educational institutions, particularly when detection results become evidence of academic misconduct.

Schools have had to deal with what happens when technology says something was AI-generated and a student says it wasn't.

Watermarking could give us better information, more transparency.

But better information doesn't automatically mean better decisions.

If a watermark tells us AI was involved but cannot tell us how it was involved, could we still end up accusing students of misconduct for AI use that was actually permitted?

Could we still end up with disputes and appeals?

Could institutions continue opening themselves up to legal challenges because a technological signal was treated as proof of misconduct rather than one piece of evidence?

The technology may be changing.

I'm not sure the underlying problem is.

And there's already another wrinkle.

Developers are already building open-source, free tools and websites that claim to remove AI watermarks.

Whether these tools can reliably defeat Claude's watermark is still an open question. But the fact that they're already being built raises another one:

Are we about to recreate the same cat-and-mouse game we've already seen with AI detection?

Maybe we've been asking the wrong question

For the past few years, so much of the conversation in education has centered around:

Did the student use AI?

I'm increasingly convinced that question isn't enough.

Because "using AI" can mean dozens of things.

Brainstorming.

Feedback.

Editing.

Research.

Tutoring.

Revising.

Generating.

and of course Replacing the thinking entirely.

Those shouldn't automatically be treated as the same behavior.

So maybe the question schools need to start asking isn't:

Did you use AI?

Maybe it's:

How did you use AI?

And:

What thinking were you responsible for?

That's a much harder question to answer with a detector or a watermark.

It requires something technology can't give us on its own:

Human judgment.

And that's why I think watermarking is bigger than a conversation about Claude.

It's a policy and a guidance conversation.

If a school's AI policy is primarily built around whether students are allowed to "use AI," we're quickly reaching a point where that isn't going to be enough.

Schools need to define what appropriate AI assistance looks like.

They need to distinguish between AI assistance, AI collaboration, and AI authorship.

Educators need guidance for what to do when those lines aren't clear.

And students need to understand those distinctions before they're accused of crossing them.

The technology is going to keep changing.

Our policies and practices have to be able to change with it.

If your school or district is trying to figure out what responsible AI use actually looks like in practice, this is exactly the work I help schools navigate.

Let's talk.

Need meaningful AI professional development and help with AI integration ?

Email me at [email protected] or visit www.teachingwithmachines.com

We got this!

Want to continue the discussion on Facebook? Join the Teaching with Machines Facebook Community!

Join our GPTs Builders for Education Discord Server!

I love reading your feedback, it helps me design effective solutions for others. I respond to all my emails. Just hit reply to this email. Rooting for you and your students! Love, M.

Beep. Boop.