Current AI sleuths are relying on "em-dashes," "rules of three," "not this but that" to cry foul. This method is highly unreliable and totally subjective. The reason these writing techniques show up in AI generated work is because authors have used them for centuries. AI trained on these techniques.

Watermarking has been integrated with AI graphics and videos for some time. Text has proven more complicated. Until now. Anthropic has announced that text generated by the latest Clause model will have an embedded, machine-readable watermark.

The change is in response to the EU AI Act, which calls for more transparency and requires AI providers to make AI-generated content detectable.

Anthropic hasn't explained how the watermark will work. One possibility could be a process to influence particular word choices during generation, creating a statistical pattern that can later be detected. Whatever the case, copying and pasting the text won't necessarily remove it.

The watermark doesn't necessarily mean "AI wrote this" but rather "AI touched this." It establishes that AI was involved in the process, but not necessarily the involvement

While the EU rules don't require AI companies to mark text when the model is simply performing standard editing, it's unclear what Anthropic's watermarking system will do in practice.

If someone asks Claude to check for typos in a paragraph, Claude still has to return that text. Will the returned passage carry a watermark? Since Anthropic hasn't explained their process, this is unclear.

Greater transparency around AI-generated work is definitely needed. Books generated by AI should be marked as such. Where it gets complicated is the "touched by AI" aspect.

I suspect this conversation will remain ongoing.

posted by Sandra L. Rostirolla on August, 11