/cdn.vox-cdn.com/uploads/chorus_asset/file/25579942/247219_C2PA_Viral_Images_CVirginiaF.jpg?w=696&resize=696,0&ssl=1 "This System Can Distinguish Real Photos From Bogus Images Created by AI—Why Aren’t Platforms Using It?")
/cdn.vox-cdn.com/uploads/chorus_asset/file/25579942/247219_C2PA_Viral_Images_CVirginiaF.jpg?resize=1200,628&ssl=1&description=This+System+Can+Distinguish+Real+Photos+From+Bogus+Images+Created+by+AI%E2%80%94Why+Aren%E2%80%99t+Platforms+Using+It%3F){kind=link}
As the U.S. presidential election approaches, the web has been abuzz with images of Donald Trump and Kamala Harris: remarkably well-timed shots of the attempted coup; perfectly ordinary shots of crowds at rallies; and shockingly unusual photos of the candidates burning flags and holding a gun. Some of these things are not Actually Of course it is. But AI generative imaging tools are now so powerful and accessible that we can no longer trust our eyes.
Some of the biggest names in digital media have been working to immaculate up the mess, and their solution so far is more data—specifically, metadata that gets attached to a photo and tells you what’s real, what’s imitation, and how it got faked. One of the best-known systems for doing this, C2PA, already has the support of Microsoft, Adobe, Arm, OpenAI, Intel, Truepic, and Google. The technical standard provides key information about where images come from, allowing viewers to determine if they’ve been manipulated.
“Provenance technologies like Content Credentials—which act like nutrition labels for digital content—offer a promising solution by enabling official event photos and other content to carry verifiable metadata such as date and time, or if needed, signal whether artificial intelligence has been used,” said Andy Parsons, C2PA Steering Committee Member and Senior Director of CAI at Adobe Edge“This level of transparency can help dispel doubt, especially during current news and election cycles.”
But if all the information needed to authenticate images can already be embedded in the files, where is it? And why don’t we see some sort of “verified” mark when images are posted online?
The problem is interoperability. There are still huge gaps in how this system is implemented and it will take years for all the necessary players to participate in making it work. And if jargon If everyone is ready to act, the initiative may be doomed to failure.
Coalition for Content Provenance and Authenticity (C2PA) is one of the largest groups trying to deal with this chaos, along with the Content Authenticity Initiative (CAI) Adobe started in 2019. The technical standard they developed uses cryptographic digital signatures to verify the authenticity of digital media and has already been established. However, this progress is still frustratingly inaccessible to ordinary people who come across questionable images on the Internet.
- Online platforms scan images for credentials and clearly flag key information to users.
- Viewers can also access the database and independently check whether an image has credentials.
“It’s important to recognize that we’re still in the early stages of adoption,” Parsons said. “The specification is set. It’s solid. It’s been vetted by security experts. Implementations are few and far between, but that’s just the natural progression of adopting standards.”
The problems start with the source of the images: the camera. Some camera brands like Sony AND Leica already embed cryptographic digital signatures based on the open C2PA technical standard — which provides information such as camera settings, date and location of the photo — into photos at the time they are taken.
Currently this is only supported by a few cameras, both new models such as the Leica M11-P or via firmware update existing models such as the Alpha 1, Alpha 7S III and Alpha 7 IV from Sony. While other brands like Nikon and Canon have also committed to adopting the C2PA standard, most haven’t yet. Smartphones, which are typically the most accessible cameras for most people, have also fallen brief. Neither Apple nor Google responded to our inquiries about implementing C2PA support or a similar standard on iPhones or Android devices.
If the cameras themselves don’t capture this valuable data, vital information can still be applied during the editing process. Software like Adobe Photoshop and Lightroom, two of the most widely used image-editing applications in the photography industry, can automatically embed this data in the form of C2PA-supported Content Credentials, which record how and when an image has been altered. This includes any employ of generative AI tools that can lend a hand identify images that have been falsely altered.
But then again, many applications, including Affinity Photo and GIMP, do not support a unified, interoperable metadata solution that can lend a hand resolve authenticity issues. Some members of these programming communities I expressed desire for them do itwhich may draw more attention to the issue. Phase One, makers of the popular professional photo editor Capture One, said Edge that it is “committed to supporting photographers” who are impacted by AI and is “exploring tracking features such as C2PA and others.”
Even if the camera supports authenticity data, it does not always reach the viewers. A Sony camera compliant with the C2PA standard was used to take the photo. the now iconic photo of Trump with a clenched fist after the attempted assassination, as well as a photo that appeared to capture the projectile that was fired at him flying through the air. This metadata information is not widely available to the general public, however, because the online platforms where these images were distributed, such as X and Reddit, do not display it when the images are uploaded and posted. Even media websites that endorse the standard, such as Modern York Timesdo not visibly mark credentials after they are used to authenticate a photograph.
Part of that hurdle, beyond enabling platforms at the outset, is figuring out the best way to present this information to users. Facebook and Instagram are two of the largest platforms that check content against markups like the C2PA standard, but these only flag images that have been manipulated using generative AI tools—no information is presented to verify “real” images.
When these labels are unclear, that can also be a problem. Meta’s “Made with AI” labels angered photographers when they were applied so aggressively that they seemed to cover even minor retouching. The labels have since been updated to reduce the emphasis on AI employ. And while Meta didn’t tell us whether it would expand the system, the company told us it believes “broad adoption of Content Credentials” is needed to build trust.
Truepic, an authenticity infrastructure provider and another C2PA member, says there’s enough information in these digital markers to provide more detail than platforms currently offer. “The architecture is there, but we need to explore the optimal way to display these visual indicators so that anyone on the internet can actually see them and use them to make better decisions, rather than simply saying something is either completely generative or completely authentic,” Truepic’s communications director Mounir Ibrahim said. Edge.
X does not currently support the standard, but Elon Musk has previously said the platform “should probably do it”
A cornerstone of that plan is getting internet platforms to adopt the standard. X, which has drawn scrutiny from regulators as a hotbed of disinformation, is not a member of the C2PA initiative and apparently offers no alternative. But X owner Elon Musk he seems willing to support it“That sounds like a good idea, we should probably do it,” Musk said when Parsons introduced it at the 2023 AI Safety Summit. “Some form of authentication would be good.”
Even if by some miracle we wake up tomorrow in a technological landscape where everyone the platform, camera and imaginative application supported the C2PA standard, denial is powerfula pervasive and potentially insurmountable obstacle. Giving people documented, evidence-based information won’t lend a hand if they simply ignore it. The disinformation can even be completely unfounded, as seen by how easily Trump supporters believed the accusation that Harris allegedly faked her rally crowds, despite widespread evidence to the contrarySome people just believe what they want to believe.
But a cryptographic labeling system is probably the best approach we currently have to reliably identify original, manipulated, and artificially generated content at scale. Alternative pattern analysis methods, such as online AI detection services, are, for example, notoriously unreliable“Detection is probabilistic at best—we don’t believe you’ll get a detection mechanism where you can upload any image, video, or digital content and get 99.99 percent accuracy in real time and at scale,” Ibrahim says. “And while watermarking can be robust and highly effective, we don’t think it’s interoperable.”
No system is perfect, though, and even more strong options like the C2PA standard can only do so much. Image metadata can be easily removed by simply taking a screenshot, for example — for which there is currently no solution — and its effectiveness is otherwise dictated by the number of platforms and products it supports.
“None of this is a panacea,” Ibrahim says. “It will mitigate downside risk, but there will always be bad actors who will use generative tools to try to deceive people.”