Last week, Anthropic announced that text generated by its Claude models will now carry an invisible watermark. Within hours, certain corners of education X were celebrating: “We’ve got ’em now, Jack!” While I understand the impulse, the schadenfreude revealed more about the educators than the students they thought they'd finally gotten the upper hand on.
Of course, Anthropic did not build its watermarking technology for educators. The watermark exists to satisfy the EU’s Artificial Intelligence Act — another attempt by EU regulators to dictate technological progress through misguided regulation. It works by subtly altering the model’s word choices, creating a pattern detectable with a key. Anthropic itself urges skepticism: a detected mark means Claude processed a text, which could mean drafting, proofreading, translation, or summarization. Nor does an unmarked text mean a human wrote it. Most critically, marked or not, a text tells us little about our students’ thinking.
So what, exactly, were these educators celebrating? The truth is they were hoping for deterrence — that their students, under threat of shaming or a bad grade — would be scared back to a world where the word processor is the only technology in the writing process. They appeared less concerned with the edge cases: the student who drafts in Spanish and translates, the one who has AI proofread his grammar, the one who argues her thesis out loud with a chatbot — while their most tech-savvy peers find ways around the watermarking key. A pedagogy rooted in instilling fear does little to engender learning. Rather, it pushes students to choose between missing the opportunity to develop skill with the technologies shaping their world, or getting better at evasion — both poor uses of their time and energy.
Responding to the announcement, the technology analyst Ben Thompson called back to a 2022 piece that forecasted this change. Thompson distinguishes between having an idea, substantiating it, and distributing it — activities “bundled” and “unbundled” over the centuries. Educators should recognize this history. To learn from a great teacher once meant sitting in the room while she taught. Socrates never wrote a word, and to learn from him you had to stand in the agora while he questioned. It was Plato who wrote the ideas down, and the dialogues outlived them both. The written word allowed lessons to outlive the lecture. The printing press allowed ideas to enter every classroom. The internet put the lecture in every pocket. Writing, print, and the internet each pried apart a piece of that bundle; what AI unbundles is the last pairing, the idea from the physical labor of substantiating it. To put this in education terms: writing is thinking — or rather, the writing process is thinking. Framing a problem, deciding what matters among sources, revising an argument, deciding something is worth arguing about — that is the thinking, and thankfully (for now) it remains distinctly human. What AI can take on is the stringing together of subjects and verbs. But all of this, as Thompson argues, is downstream from the prompt. That is, it is downstream from the thinking.
This is why the cheers deserve to be named for what they are: a kind of laziness. The worry is legitimate — polished prose really did once serve as rough evidence of thought, and tools like Claude and GPT really have broken that proxy. But faced with this shift, some educators are refusing to reimagine and redesign their pedagogy for the demands of the day. They are now relying on the chilling effect of a detector so their syllabi and assignments can remain evergreen. While accusing students of outsourcing their thinking, it is these same educators who are outsourcing a critical piece of their pedagogy to AI.
Writing in the National Catholic Register this spring, Santiago Schnell brings us to where educators should be focused. Learning, he writes, “is not the production of acceptable performances but the formation of a person capable of truth, judgment and responsibility.” Schnell calls on educators to look past prohibition and toward pedagogical redesign: oral defenses, on-demand essays, seminars built around live questions. Educators might also task students with sharing drafts or AI chat logs, making their thinking visible. If we want to know whether students are thinking, we will have to build assessments that look at the thinking, not the outputs. To the extent that your assessment strategy has been anchored in word and page counts — that in this moment you are fetishizing the typing of sentences — AI is exposing a vulnerability in your practice. You may still welcome the AI detectors, but they won’t tell you whether your students are thinking, because the construction of a sentence is no longer evidence of the mind.
JARED FRANCIS is a doctoral student in the Graduate School of Education’s Mid-Career Doctoral Program from New York. His email is jrf1@upenn.edu.
