Tech analyst Ben Thompson argues that mandating AI companies to disclose the nature of the content generated by their models misses the target. He voiced strong criticism of the European Union's AI Act, which was cited by Anthropic as the motivation behind its decision to incorporate watermarks in the outputs from its Claude models.
In his popular "Stratechery" newsletter, Thompson expressed that the push for such transparency in AI is both excessive and unlikely to be effective. He admitted he understands the rationale for watermarking—wouldn't it be beneficial to distinguish between genuine and fabricated content?—but he questioned what “fake” really means. He posed a thought-provoking scenario: If a language model produces a truthful statement while a human shares a fictional tale, does it really matter if the latter lacks a watermark?
Anthropic’s recent announcement of its watermarking initiative has sparked a significant debate within the tech community. The company indicated that all Claude models released after August 2 will come with watermarking capabilities, with plans to retrofit older models as well.
Thompson raised concerns over the watermarking policy, particularly in scenarios where users utilize Claude for tasks like proofreading. Although the EU's regulations permit exceptions for editing, Thompson highlighted that there’s no reliable method for Anthropic to determine such usage accurately. He noted, "I’m relieved I never developed the practice of copy-and-pasting during proofreading—I modify each change manually in my text editor—but it hardly seems fair that anyone improving their grammar could risk having their original work branded as AI-generated; the same applies to translation."
Meanwhile, OpenAI, which currently employs watermarking for images and audio, has also committed to applying this technique to text outputs, though specific implementation details remain unclear. OpenAI utilizes Google DeepMind's SynthID for watermarking its multimedia content.
Regulatory bodies are increasingly looking to watermarking, which has traditionally been applied in visual media, as a means to help the public identify AI-generated content. While visual watermarks are easily detectable, provenance markers for AI-generated text are not visible, prompting companies like Anthropic to promise the development of tools that will effectively act as AI content checkers.
According to the EU regulations, AI firms are mandated to mark their outputs in a machine-readable format that clearly signifies whether the content has been artificially created or altered. Companies including Anthropic, OpenAI, Google, Meta, and Microsoft have all endorsed this regulation, though SpaceX has not signed on. The watermarking requirement officially took effect on August 2.
As for Anthropic, it has not clarified whether it will enforce watermarking solely in Europe, leaving it uncertain if other companies will adopt similar practices worldwide in the future.
Thompson ultimately contended that watermarking falsely attributes creative ideas to AI, asserting that these models should be viewed as tools used by humans. He likened the insistence on watermarking to demanding that a ballpoint pen take credit for authorship—a notion he deemed preposterous.
