Claudefor a final polish or translation. Conversely, a lack of a watermark isn't a certificate of human authenticity.
The panic surrounding this usually stems from people conflating three entirely different technical layers of the AI workflow.
The three layers of "detection" #
Model-level watermarking: This is what Anthropic is doing. It's a statistical pattern baked into the token selection process. It's fragile; heavy editing or translating the text can wipe the signal entirely.Third-party detectors: Tools like ZeroGPT. These aren't official and have a track record of being unreliable at scale. They guess based on perplexity and burstiness, not a secret key from the model provider.Platform-level badges: When a site like Medium decides to slap an "AI" label on a post. This is an editorial or algorithmic choice by the platform, not a mechanical certainty triggered by a watermark.
The mistake most people make is assuming a linear chain: Anthropic marks the text → detectors find the mark → platforms punish the author. In reality, these are disconnected events.
Assisted vs. Generated vs. Produced #
From a prompt engineering and AI workflow perspective, we need to stop using "AI content" as a catch-all term. There is a massive difference in editorial responsibility depending on the process. AI-Assisted text is where the human maintains total control. The core idea, the structural outline, and the final verification are all human-led. The LLM agent is used as a high-end thesaurus or a translation tool to refine phrasing. The "intent" remains human.
AI-Generated text is the result of a prompt where the model handles the heavy lifting of drafting. While the human provides the prompt, the phrasing and flow are the model's.
AI-Produced text is industrial-scale automation. This is the "content farm" approach where pipelines pump out articles with zero human editorial oversight.
The nuance here is that a watermark cannot distinguish between these three. A heavily assisted piece of writing could still trigger a watermark if the model rephrased a few key paragraphs. If we judge content based on a binary "AI or Not" badge, we ignore the actual quality and the level of human agency involved in the production. For those of us benchmarking these models, the real metric isn't whether a model helped, but whether the human remained the decision-maker.
Next Can we actually trust "hidden" reasoning blocks in LLM APIs? →