Does Gemini API text output carry SynthID watermarking? (gemini-2.5-flash-lite / gemini-3.1-flash-lite — EU AI Act Art 50(2))

Gemini API
1

I run a small EU-based game whose scenes are generated through the Generative

Language API (paid tier, gemini-2.5-flash-lite today, migrating to

gemini-3.1-flash-lite in October). Under the EU AI Act, Article 50(2), I have to

mark AI-generated text in machine-readable form — unless I can rely on the model

provider’s own marking solution (Commission guidelines C(2026) 5054, point 74).

I could not find a definitive statement on whether API text output is watermarked:

- The SynthID overview mentions text watermarking only for "the Gemini app and

web experience" (deepmind.google/technologies/synthid)

- The developer docs present SynthID Text as a logits processor you apply to

your own model’s generation pipeline, not as an API property

(ai.google.dev/responsible/docs/safeguards/synthid)

- The Gemini 2.5 Flash-Lite model card does not mention watermarking at all

- The published verification routes (uploading content to Gemini chat, the

SynthID Detector portal) accept image, video and audio only — no text

(deepmind.google/models/synthid)

- An earlier answer in this forum said Google does not offer a public API to

detect text watermarks

I also inspected a live generateContent response — headers and JSON body carry

no provenance metadata, so my question is about the generated text itself.

My questions:

1. Is text returned by the API for these models SynthID-watermarked, or marked

with any machine-readable provenance signal? Is that documented anywhere I

can cite? Does the behaviour differ between the 2.5 and 3.1 model families?

2. If it IS watermarked: what detection route exists for a downstream provider,

or for people exposed to the content?

3. If it is NOT: is native text watermarking for the API planned?

A definitive answer either way — ideally from Google staff with a pointer to

official documentation — would help every developer here who has to write an

Article 50(2) compliance file. Thank you!

2

Hi there thanks for contributing to community!

Generated text from the API is NOT SynthID-watermarked. There is no machine-readable providence signal. This is only supported for other generated modalities (image, video, etc.)

Native text watermarking is not planned at the moment

3

hello trey, does this also apply to text generated through google ai studio and/or antigravity?

4

Hi, I need to post an important correction to my previous message:
After checking further with the team, it turns out that text generated via the Gemini API IS actually SynthID-watermarked! I apologize for the confusion earlier — the previous statement was incorrect. This also applies to text generated through Google AI Studio and/or antigravity!

5

thanks for checking. was synthid text always applied to api text, or was it introduced at some point? if it was introduced later, can you say approximately when? I’m asking because deepmind’s current documentation doesn’t mention anything. also since we have no means of checking the text itself, it would be useful to know if it was something introduced only recently. thanks in advance

6

Follow-up for downstream providers under EU AI Act Art. 50(2) - thanks for confirming
that Gemini API text output is SynthID-watermarked. Two follow-ups:

  1. Does the same apply to Gemini models served through Vertex AI generateContent
    (location “eu”), e.g. gemini-3.1-flash-lite and gemini-3.8-flash? Does JSON /
    structured output, thinking level or disabled caching change it?
  2. Which detection tool can third parties (providers, market surveillance
    authorities, the public) use to verify SynthID-Text on such output? The SynthID
    Detector portal is waitlist-only.

The Commission guidelines on Art. 50 (C(2026) 5054, points 70, 74, 75 and 78) let
downstream providers rely on model-level marking, but require both marking and
detection to be available.

Thanks a lot!

7

Following up on the confirmation above (19 Aug): does the same hold for gemini-3.8-flash on a paid, billing-enabled AI Studio project (Generative Language API, generateContent, thinkingLevel low, default sampling)? Concretely: (1) is the returned text SynthID-Text watermarked, and since when; (2) is there official documentation we can cite in an EU AI Act Art. 50(2) file, given that the Commission guidelines (C(2026) 5054, point 74) let downstream providers rely on model-level marking; (3) which detection route can market surveillance authorities, and we ourselves, use for text today? Thank you.