SynthID

“in expectation the next-token distribution is still the same as the unwatermarked case”

“when they offer it as a feature to track people you can at least reject their preferred marketing term”

“it will be reverse engineered and AI generated content will be wiped clean of this "imperceptible" footprint”

“pretty cool standardizing technology”

“Given that it's an ad company, they want something to connect what you're doing to the rest of their data”

“Finally, the previous implementations were very silly!”

“the EULA incorporates Google’s regular consumer terms; which means what you upload can also be used for advertising and targeting”

“this technology can encode database identifiers and enough entropy to uniquely identify you as an author”

“SynthID is a better standard for designating something was created with GenAI”

“I see a net positive especially since it doesn't degrade text quality”

“they are incredibly unlikely with SynthID”

“before applying synthid the percentages were 48% and 52%”

“It could be really useful for preventing synthetic data from accidentally contaminating your real datasets down the line.”

“Which won't cause false positives but only catches content from tools that actively adds this mark.”

“Any signal which helps distinguishing the two is helpful, even if it isn't perfect.”

“This blog post clearly shows that Deepwalker can break SynthID, which is a closed-source watermarking system.”

“99% of the users of these tools are never going to bother trying to figure out how to remove synthId”

“Only the investigative or journalistically inclined will make use of this, and those people already fact check.”

“I really hate these vibes-driven reactions to (checks notes) content attribution.”

“SynthID is fairly resistant to this sort of thing, although not perfect.”

“Just because something isn't perfect doesn't mean it's not useful.”

“It helps significantly in the current moment.”

“isn't the problem that malicious actors will create systems that do not use synthID”

“denoising using a diffusion model was a relatively trivial means to beat most steganographic watermarks”

“Despite people saying it will be removed, I have seen no reproducible repo demonstrating it.”

“SynthID survives basic transforms including screenshots/photos, although it can of course be defeated.”

“But that harms users too, by introducing further barriers for indie users.”

“It's a detector that will detect 100% of the stuff that doesn't mind being detected, and only the dumbest fraction of stuff that doesn't want to be detected.”

“If it's this easy to remove SynthID from an AI-generated image then it wasn't a good solution in the first place.”

“This solution helps when people lazily share things with friends or on social media I suppose, rather than stopping motivated bad actors.”

“there is no way for me to verify if groups or nation states can pay for a special contract where they do not have to have their outputs watermarked”

“SynthID is removed from images by passing the image through a diffusion model for a single step at low denoise”

“really just validates whether a piece of AI content was generated by Google, not AI generated in general”

“It's security through obscurity.”

“I'm glad SynthID (apparently) makes it impossible”

“Synth id can be removed, run it through an image 2 image model with a reasonably high denoising value or add artificial noise and use another model to denoise and voila.”

“It is, because of SynthID.”

“it can survive the kind of basic editing a regular person knows how to do on their phone”

“While a good first step, it doesn't solve the problem of AI generated content NOT having any kind of watermark.”

“So whoever creates AI content needs to voluntarily adopt this so that Google can sell "technology" for identifying said content?”