SynthID Detector

Commenters focused on Google’s SynthID detector and the experience of using it: many, including Tiberium and latexr, complained it requires signing in (Google/Apple/ChatGPT) and has tight rate limits (cubefox ≈10/day), while others like RobGR objected to being gated by EULA before seeing the service. Several contrasted Google’s approach with OpenAI’s provenance tool (sarkarghya, cubefox) and older Google Image Search capabilities (Tiberium, lelandfe, mh-), and linked to reverse‑engineering writeups (Retr0id, m00dy, pbronez) that attempt to explain or replicate the watermarking. Retr0id and brokensegue argued the lack of bulk/offline classification is frustrating for legitimate use cases like scanning Wikipedia, and wodenokoto and dannyw read the EULA as evidence Google fears abuse or watermark‑stripping.

Opinion splits on efficacy, intent and privacy. Several (iamleppert, user43928) argued watermark removal could be learned by training models, while MisterMunchkin and others claimed providers and Chinese models refuse to help strip SynthID and that access controls are deliberate to prevent laundering. Possibilistic warned these are “spymarks” that could encode identifiers; cubefox and others disputed that Google embeds personal data. 98codes, MisterMunchkin and Avicebron expressed distrust of Google’s motives given its ad business, while _flux and cubefox countered that privacy could be designed in and abuse concerns justify restrictions.