TechCrunch reported that Anthropic plans to watermark text generated by its AI models. Predictably, that immediately reopened the familiar debate around detection, provenance, enforcement, and how educators might identify AI-generated work.
So I did what I often do when I hear a technical claim. I tested the surrounding assumptions.
I opened Visual Studio Code, created a folder called “AI Police,” and asked ChatGPT to help me build a Python utility for sanitizing Microsoft Word documents. What started as a simple script quickly became something much more substantial. We added DOCX-only validation, drag-and-drop support, metadata cleanup, custom property removal, comment removal, Unicode normalization, zero-width character cleanup, package relationship repair, integrity checking, JSON audit reports, and automatic storage of sanitized documents.
Then we tested it, broke it, fixed it, tested it again, and improved it.
About 45 minutes later, I looked at the finished result and had to laugh at myself.
The coding exercise was not pointless. In fact, it was useful. I learned more about the internal structure of DOCX files than I expected, and the final tool is legitimate enough that I put it on GitHub for anyone who wants it:
https://github.com/robazprogrammer/docx-sanitizer
At least nobody has to walk away from this post empty handed.
But the larger lesson had very little to do with Python.
I had just spent 45 minutes chasing the AI-detection rabbit down another hole.
Those same 45 minutes could have been spent designing a more difficult assignment. I could have built an exercise where students are required to explain their reasoning orally. I could have designed a problem that forces them to defend a decision, respond to follow-up questions, apply a concept to an unfamiliar situation, or demonstrate that they understand the work they submitted.
I could have spent that time helping students learn how to use AI as a coach and collaborator rather than as a cognitive surrogate.
That seems like a far better use of an educator’s time.
There are legitimate questions around authorship, provenance, disclosure, and academic integrity. Those conversations matter. But educators have finite time, and every hour spent trying to invent a better mousetrap for catching AI use is an hour not spent redesigning learning for a world in which AI already exists.
That is the part I think we keep missing.
The challenge is not simply to determine whether AI touched a piece of writing. The challenge is to design learning experiences where students still have to think, explain, defend, adapt, and demonstrate understanding regardless of which tools they used along the way.
A watermark does not redesign an assessment.
An AI detector does not create deeper learning.
A hidden character does not tell you whether a student understands the material.
Good assessment design can.
If the goal is learning, then perhaps our energy is better spent building assignments where weak thinking is exposed through the work itself, where thoughtful AI use improves the process, and where outsourcing cognition becomes difficult because students still have to demonstrate that the reasoning is actually theirs.
And if you have read this far, you probably could have spent your time more wisely too.
That is exactly the point.