As a bilingual writer, I have a particular reason to distrust AI detectors.
I have put English writing of mine from decades ago through some of these tools and watched it come back labelled as substantially AI-generated. This was writing produced long before large language models existed, and before anybody had thought of training a machine to imitate a sentence. Apparently, the detector found it suspicious anyway.
Perhaps that is why the idea of watermarking AI-generated text is tempting.
Unlike a detector, which looks at a piece of prose and makes a statistical guess about its likely origin, a watermark would be deliberately embedded in the output of an LLM. In theory, that should mean fewer false accusations against human writers. A watermark ought to be able to reliably identify words chosen by the model, rather than declaring that a bilingual writer’s plain English, regular grammar or slightly unusual rhythm must have come from a chatbot.
I would hope so.
But the word reliably still does a great deal of heavy lifting.
Anthropic’s reported plan to embed invisible watermarks in Claude’s output has started a predictable argument. Supporters say readers, employers and educators should be able to know when they are reading AI-generated work. Critics point out that a watermark may persist after a person has used AI merely to proofread, translate, shorten or tidy text they wrote themselves.
Anthropic’s own reported position is fairly cautious: a watermark could establish only that Claude was likely involved with the content at some point. That seems about right. The difficulty is that a university, employer or client may interpret a positive result rather differently: AI wrote this.
Those are not the same claim.
A person might use an LLM to correct grammar, sharpen a paragraph, suggest a heading or summarise a long report. They may reject most of its suggestions. They may do all the research, form the argument and take full responsibility for the finished work. Yet a watermark could become an indelible mark against the whole document.
That would be a mistake. We do not treat AI assistance in every field this way.
A financial analyst can use AI to summarise figures or help prepare a report. The analyst still has to check the figures, understand the conclusions and answer for them. If the report is wrong, blaming the software will not go very far.
A programmer can use an AI coding agent to generate a function, a test suite, a database schema or a substantial piece of an application. We generally call that efficiency. Again, this is exactly as it should be. The programmer must still decide whether the code is correct, secure, appropriate and maintainable. Code that compiles is not necessarily code that works or does the right thing. Anybody who has spent time in software knows that lesson.
But when the output is prose, we suddenly become much more anxious. The writer is assumed to be passing off somebody else’s thinking. We stop asking whether the work is good, accurate and defensible, and start looking for a technical way to identify the tool.
Why should writing be treated so differently?
There are certainly cases where disclosure matters. A journalist, researcher, lawyer or public official may have obligations that go beyond producing a competent piece of text. A job application may reasonably be expected to show what the applicant can write unaided, although that may become less of a mandatory requirement in the not too distant future. Institutions are entitled to make rules about that.
But rules about disclosure and responsibility are one thing. A permanent technical fingerprint is another.
The education argument is where this becomes most interesting. Universities have a real problem: an unsupervised essay no longer tells us much about what a student can do without assistance. But students have always had ways to get help they were not supposed to receive. Copying. Tutors. Family members. Old assignments. The friend who was “only helping”.
Years before anybody had heard of generative AI, I tutored a first-year university student who was struggling with assembler programming.
He was not unintelligent. The problem was that the lectures had skipped past the foundations: how a CPU works, memory addressing, registers, data encoding, and the awkward but important fact that a byte has no inherent identity. In one context it is data. In another it is an opcode. You cannot know what it means simply by looking at it in isolation.
Over several weeks, we worked through those ideas properly. We also discussed things such as a return stack. The mock CPU used in his course did not provide one, so we talked about how such a convention could be implemented using a register as a stack pointer and a reserved address range as a stack, and why doing so resulted in a better program structure.
When he completed the assessment, he wrote the program himself. But he used some of those supposedly advanced ideas, including a subroutine call-and-return convention. The result was a neatly structured program.
His lecturer assumed somebody else must have written it.
To the university’s credit, he was invited to defend the work orally. He did so without difficulty. He understood every part of it, why it worked and why he had made the choices he had made. He received a high distinction.
That was the right outcome. Not because he had produced the correct and better than expected code, but because he had learned the subject.
Now change one part of the story. Suppose an AI system had generated the entire program for him, guided by his prompting to implement the call stack by convention. Suppose he had studied the generated code, researched the technique, understood its strengths and limitations, and then used that understanding to iterate on the code to arrive at the submitted assignment.
Would he have learned any less?
Possibly he might even have learned more quickly. But he would certainly not have learned anything less.
The proper test is not whether a student has ever seen an answer, received guidance or used a tool. It is whether they understand the work well enough to explain it, adapt it and defend it. A student who pastes an AI answer they cannot explain has learned very little. A student who uses AI as one source of instruction, checks its work and can reason through the result is likely to have learned a great deal.
We already have ways to distinguish the two. Supervised examinations, practical demonstrations, presentations and oral defences are imperfect, but they assess understanding far more directly than forensic inspection of prose or code.
The slide rule was not banned because it made arithmetic easier. Nor was the electronic calculator. Education changed what it tested. Students still had to understand the calculation, recognise an absurd answer and explain their reasoning.
AI should force the same adjustment.
Watermarks may have a legitimate role. They could help model developers identify synthetic material in future training data. They may assist in tracing large-scale automated content. And, if they genuinely and reliably identify model-chosen text rather than merely detecting statistical regularity, they may be less unfair to human writers than the detector industry has been.
There is also a practical cost worth acknowledging. Watermarking works by nudging the model towards some statistically plausible token choices and away from others. Anthropic says any resulting decline in output quality is “imperceptible”. Perhaps it is. But a system that compromises output, even slightly, needs to earn that compromise. Identifying model involvement may be useful in some technical settings; it is a much weaker justification when the result will be treated as proof of authorship or lack of understanding.
But watermarks cannot settle authorship in the meaningful sense.
A watermark cannot tell us who formed the idea, selected the evidence, rejected bad suggestions, checked the facts or took responsibility for the finished work. It cannot tell us whether a student understands a program. It cannot tell us whether a writer can stand behind an argument.
Those are human questions. They need human answers.
The important thing is not whether a tool touched the work. It is whether the person putting their name to it knows what they are doing – and is prepared to answer for it.
Prompted by “3 arguments for and against AI watermarks”, Business Insider via AOL.