Skip to main content

How BrowseWiz marks AI-generated content

The EU AI Act asks providers of AI systems to mark AI-generated content in a way machines can detect. BrowseWiz doesn't add its own hidden watermark to text. Instead, it relies on the invisible watermark that the AI model writes into its own answers. That watermark stays in the text when it's typed into a web page, pasted into a sheet or sent by email, which a mark added afterwards wouldn't.

Which outputs carry which watermark​

  • Chat answers, including scheduled task answers, chat summaries and plans: BrowseWiz Chat and BrowseWiz Chat Pro answer with Anthropic's Claude first, and Anthropic watermarks Claude's text. If Claude is unavailable, a Google Gemini model answers instead. Gemini text is reported to carry Google's SynthID watermark, though Google hasn't documented this yet.
  • Text the browser agent types into pages, and what it writes to Google Sheets and Google Calendar: the same, Claude first, then Gemini models.
  • Alert emails: they contain the AI's answer or summary, so they carry the same watermark as the chat answer.

Very short text, under about 200 tokens (roughly 150 words), doesn't need a watermark, and a watermark in it is weak.

What isn't marked, and why​

  • Chat titles, monitor headlines in notifications, and translated interface text: these are short labels or translations, which the rules don't require to be marked.
  • Read aloud: the AI voice reads an answer that is already marked, so the audio adds no mark of its own. The button and messages always say it's an AI voice. If the AI voice is unavailable, your browser's built-in voice reads instead; BrowseWiz can't mark that audio.
  • Rewritten or partly copied text: editing an answer, or copying only part of it, can weaken or remove a model's watermark.

When you use Copy on an answer, the formatted copy carries a machine-readable tag saying it's AI-generated. This is the only mark BrowseWiz adds itself.

Your own key and custom models​

Answers from Bring Your Own Key models and custom models go straight from your browser to the provider you chose. BrowseWiz doesn't mark them. Whether they're watermarked depends on that provider: OpenAI's text watermark for its API, for example, is off unless it's turned on.

Where to check for a watermark​

Each provider runs its own detector. Access to them is limited for now:

  • Anthropic's detection API checks Claude text. It's in private preview for regulators, media, researchers and businesses checking their own Claude output. See how Claude marks AI-generated content.
  • Google's SynthID Detector checks content watermarked with SynthID, including Gemini text. It has a waitlist. See Google SynthID.
  • OpenAI's Content Provenance API checks content OpenAI has marked. See OpenAI's content provenance guide.
Last updated on