Artificial intelligence has created an authorship problem that institutions are still struggling to solve.
This week, that conversation took an important turn.
Anthropic has announced plans to introduce invisible watermarking in Claude-generated text, creating a mechanism to help identify AI-generated content. The watermark is intended to work by subtly patterning the words and phrases produced by the model rather than by a visible label attached to the text. The move comes as AI companies face increasing pressure, including transparency requirements emerging from the European Union’s AI regulatory framework, to make AI-generated content more identifiable.
On the surface, this seems like a positive development. And in many ways, it is.
As generative AI becomes embedded in education, business, research, communication, and professional decision-making, transparency matters. We should be able to have meaningful conversations about when artificial intelligence has contributed to something we read, evaluate, approve, or act upon. But there is a much bigger question hiding behind the excitement around watermarking: Does detecting AI involvement actually tell us whether AI was used responsibly? I do not believe it does!
This distinction matters because we are rapidly approaching a point at which AI assistance and human authorship will become increasingly difficult to distinguish. Consider a professional who writes an entire report independently and then asks Claude to improve its clarity. Consider a professor who develops an original lecture but asks AI to refine several sentences. Consider a student who uses AI to understand a difficult concept, drafts an answer independently, and then asks the system for feedback, or consider an executive who asks AI to summarize information before independently reviewing the original evidence and making a decision.
In each scenario, AI was involved.
But was AI the author? This is a much more complicated question.
This is why I support the principle behind AI watermarking while remaining cautious about how we interpret what a watermark actually proves. Transparency is valuable. Transparency can tell us that artificial intelligence may have played a role in producing or transforming content.
But transparency alone cannot tell us whether the human behind that content exercised judgment.
It cannot tell us whether the information was verified.
It cannot tell us whether inaccuracies were corrected.
It cannot tell us whether the individual understood the material.
And most importantly, it cannot tell us whether the human responsibly accepted the final output. This is the distinction I believe organizations and educational institutions need to understand as AI becomes more deeply integrated into everyday work.
There is also another complication. Almost immediately, another challenge emerged. Researchers and developers began testing whether invisible AI watermarks could survive rewriting, paraphrasing, translation, and other transformations. That raises an important governance question: if a technical signal can be altered or removed, should we rely on detection as the primary measure of responsible AI use? This development should make us reconsider an important assumption. Perhaps we are asking technology to solve what is fundamentally a human-governance problem.
Watermarking can become part of a responsible AI ecosystem. Detection technologies may also have a role. Disclosure requirements can provide additional transparency. But none of these mechanisms should become substitutes for critical thinking, verification, human oversight, or accountability. This is essentially important in education.
If universities use watermarking primarily as another mechanism for catching students using AI, we risk missing the larger transformation happening in learning.
The question can no longer simply be: “Did this student use AI?”
A more meaningful question is: “Can this student demonstrate that they understood, evaluated, verified, and took responsibility for the work they submitted?”
The same principle applies beyond the classroom. When an employee uses AI to prepare a financial analysis, when a healthcare professional consults an AI-supported system, when a manager uses an AI-generated recommendation, or when an organization relies on an AI agent to support a consequential decision, detecting that AI participated in the process is only the beginning.
We still need to know what the human did next.
Did someone challenge the output?
Did someone verify the evidence?
Did the person have the authority to accept the recommendation?
Was escalation required?
Was the decision documented?
Who ultimately accepted responsibility?
These questions will become increasingly important as AI systems move beyond generating text and begin taking actions on behalf of individuals and organizations.
That is why I believe the future of responsible AI cannot be built around detection alone. It must be built around responsible human reliance. The goal should not be to create environments in which people are afraid to admit they used AI. That approach may simply encourage concealment. Instead, organizations and educational institutions should create environments where people can demonstrate how they used AI, how they evaluated its output, what evidence they considered, and why they ultimately accepted, modified, escalated, or rejected the system’s output.
Claude’s watermarking initiative is therefore an important development. It moves the AI industry toward greater transparency, and that deserves recognition. But we should be careful not to confuse a signal of AI involvement with proof of authorship, misconduct, reliability, or responsible decision-making.
A watermark may help answer: “Was AI involved?”
Responsible AI governance must answer the much harder question: “What did the human do with what AI produced?”
This is where accountability begins. As AI becomes nearly invisible within the tools we use every day, I believe that distinction will matter far more than whether we can simply detect the machine.
What do you think? If AI helped improve, edit, or restructure something you originally created, at what point should that work be considered AI-generated rather than human-authored?
