After a three-year debate on how to distinguish AI-generated content from human writing, a watershed moment may finally be upon us.
This past week, Anthropic, the maker of Claude, confirmed that its new models will embed invisible, machine-readable watermarks into all generated text.
The announcement was triggered by Article 50(2) of the newly enacted European Union Artificial Intelligence Act, which requires AI companies to include the mark as part of their transparency obligations.
The law responds to the phenomenal leap in AI capabilities that has made it virtually impossible to separate real human-made content from machine-generated synthetic media. Europe has taken the lead in imposing an enforceable duty for content tracing, but the impact of this new law will not stop in Europe.
Instead, it will follow an increasingly familiar trend known as the ‘Brussels effect’, where rules designed for the European market inadvertently become global standards. Anthropic announced earlier in the week that, indeed, the watermark would apply to all Claude-generated content throughout the world. Like many other tech giants before it, the company is defaulting to a business logic that says it is easier and cheaper to use one high standard everywhere than to make different products for different regions.
But business aside, the law raises fundamental questions—especially for Africa—that must be scrutinised before its provisions are celebrated. First, African newsrooms, universities, and regulators did not ask for this system or participate in its development, yet they will inevitably adopt it the moment it takes effect.
This directly implies that the watermark could be applied to content and language types that were never part of the negotiations or technical calculus that led to the creation of the law and mark. Thus, Kenyan code-switched academic prose, Sheng-influenced English and scholarship translated from Kiswahili or Luganda will be subject to a detection logic that was not designed with them in mind.
Yet, once operational, a university in Kampala or Mombasa may enthusiastically apply the new AI detector with the same assumed reliability despite the obvious misalignment.
The second issue relates to the exact nature of the problem this new system is seeking to address. It is important to acknowledge that developing a system that detects and governs work done by AI tools is an attractive proposition for many institutions.
For instance, universities are already grappling with growing questions about the authorship of a considerable number of essays submitted by their students. Newsrooms are under pressure to assure readers that stories published on their platforms are created by humans. Similarly, publishers are concerned about the prospect of an unprecedented number of plagiarism claims arising from AI-generated content being passed off as human-made.
However, at least using extant technology, watermarking promises more than it can deliver. As it is currently designed, the mark can only determine with certainty whether a given text has been touched by an AI model. While this is an important starting point, it does not go far enough in answering the simple question that various institutions would like answered definitively: did a human actually write this unassisted, or was it produced by a highly capable AI tool?
Right now, a watermark can appear in a human-written text simply because the author proofread, translated, or edited it themselves. Conversely, the mark might not show up at all on AI-generated text if someone heavily edited or paraphrased it before publishing.
Anthropic itself has been unusually candid about this certainty gap, cautioning that neither the presence nor the absence of a mark should be read as proof of anything in particular. That caution would be reassuring if it weren’t so difficult to sustain in practice. A university disciplinary panel investigating a plagiarism claim, for instance, craves a binary answer: the student either engaged in academic misconduct or they did not.
Therefore, a machine-readable signal, regardless of its makers’ cautionary notes, may still be used to make consequential decisions. In the past, similar fuzziness has not prevented enthusiastic enforcers from unduly leaning on such technologies as lie detectors and breathalysers, which suffer from the same certainty gap.
This must not be seen as a case against watermarking or any future regime that would seek to address transparency in content creation. Instead, the moment calls for a more thoughtful reflection on what the real problem was in the first place. Could it be that the question is not if AI touched something, but rather whether the thinking in the text was honest and in good faith? Unfortunately, no watermark may ever be able to answer that question for us.
Instead, we have to embark on the more difficult but potentially durable work of rebuilding defensible editorial and academic judgement around AI-assisted text. This includes stringent disclosure norms that normalize declaring AI-assisted work, alongside multi-layered review processes.
Watermarks can be one useful signal here, but they are not the end-all-be-all of content tracing. These measures are not mutually exclusive. A watermark, a widely embraced disclosure system, and human judgment are much stronger when used together.
Dr Ageyo is the Editor in Chief of the Nation Media Group. He holds a PhD in media studies with a focus on science and environment communication. [email protected]
