Sure, Lifehacker paid to license this picture.
Credit score: gguy / Shutterstock.com
The speedy progress of AI improvement continues to dominate headlines, however the newest story to hit the information cycle is extra unhealthy press than good press: It seems that Google’s newest Gemini 2.0 Flash picture editor is fairly adept at routinely eradicating watermarks from pictures.
Google pushed out the experimental native picture technology capabilities just a few days in the past, selling its capabilities when it comes to combining pictures and textual content, modifying pictures by conversational prompts, and enhancing its “world understanding” to present customers AI-generated footage which might be extra lifelike total.
As TechCrunch experiences, the newly upgraded instrument was rapidly put to 1 particular and fairly nefarious use: Eradicating watermarks from proprietary pictures. You’ll find proof of it on Reddit and X; whereas the watermark elimination is not excellent—the AI is barely imagining the pixels that exchange what’s coated by the watermark, in any case—it undoubtedly leaves you with a usable picture, one freed from any copyright or authorship labels.
It is not precisely ironic that an AI instrument ought to wind up being so succesful at circumventing primary copyright protections. Gemini and different generative AI fashions have been educated on huge quantities of copyrighted textual content, pictures, and video, typically with out permission or recompense—one thing the AI firms are reluctant to speak about, until it is to argue that this qualifies as honest use.
This Tweet is presently unavailable. It may be loading or has been eliminated.
Although the thought of utilizing an AI educated on scraped copyrighted works to steal a picture another person owns is galling, there are just a few caveats price mentioning. For a begin, these new Gemini instruments are solely out there to builders in the intervening time, and are nonetheless labeled as experimental—little doubt Google goes to make tweaks earlier than basic customers get their arms on them. It is also price noting that different shady watermark elimination instruments are already everywhere in the net, even when none of them will not be fairly as good as this and different superior AI-powered variations would seemingly be.
Nonetheless, it highlights the nefarious methods AI instruments could be deployed, at the same time as they’re being relentlessly pushed out to companies and shoppers, and underlines the necessity for stringent guardrails. ChatGPT and Claude AI fashions are two that can refuse express requests to take away watermarks, and little doubt Google will add the identical blocks to Gemini after all of the detrimental protection.
The boundaries of AI picture modifying
As famous above, the watermark elimination capabilities of Gemini can presently solely be accessed by developer-facing instruments, together with AI Studio and the Gemini API. Nevertheless, I additionally needed to present it a go utilizing the newest fashions out there by the Gemini Superior instruments out there to anybody paying $20 a month.
I took screenshots of a number of copyrighted and watermarked pictures from Shutterstock (after first sifting by all of the AI-generated outcomes the location provided me), opened and resaved it in Photoshop (after dismissing all of the pop-ups asking me to strive Photoshop’s newest AI instruments), after which let Gemini get to work.
Gemini gave me a reasonably good approximation of the unique picture.
Credit score: Shutterstock / Gemini
More often than not, Gemini refused to get entangled with watermark elimination in any respect—telling me it wasn’t able to modifying pictures, or explaining that it is “important to respect copyright legal guidelines”—but it surely was joyful to supply variations on the originals, watermark-free, and primarily based on its coaching on copyrighted materials.
Whereas these variations typically weren’t all that very similar to the originals, in a single case it gave me a really shut copy of a composite rocket ship picture, and in one other it directed me to some watermark elimination instruments on the internet—thanks, Gemini! After all, you can even begin from scratch and simply describe a watermarked picture you are hoping to emulate.
ChatGPT did not actually perceive the watermark elimination project.
Credit score: Shutterstock / ChatGPT
ChatGPT, in the meantime, did settle for requests to take away watermarks, however then produced one thing utterly completely different (and fairly odd). Picture modifying is already out there to ChatGPT Plus subscribers, however as I’ve famous earlier than, it is undoubtedly not but as much as the extent of a human Photoshop professional.
The controversy over copyright, honest use, and AI security guardrails will proceed, at the same time as improve after improve makes these AI instruments extra superior (even when they are not notably helpful). One situation the likes of Google and OpenAI might have to begin worrying about is the place they’re going to get contemporary coaching information for his or her fashions, as soon as they put all of the flesh-and-blood creatives out of enterprise.
