From 1989 to 1999 I was an award-winning conflict photographer for Reuters and The Associated Press news agencies where I repeatedly risked my life documenting violence around the world. During that period, I photographed eight wars on four continents, all major stories at the time. Among other adventures, I was taken prisoner by Saddam Hussein’s army at the end of the First Gulf War, wounded in Sarajevo, and lost over a dozen journalist friends to violent deaths.
My commitment to truth and authenticity runs deep.
In those pre-internet, and mostly pre-digital, days, my images, and those of my photojournalist colleagues were published in newspapers and magazines across the globe and were accepted for what they were; genuine, fact-based photographs of world events. Our work was generally trusted, taken at face value, and rarely manipulated.
Fast forward to today when we are drenched in information constantly delivered to our multiple digital devices. Whether we interact on smartphones, smartwatches, tablets, or laptops, it is increasingly challenging to discern the source and veracity of all the content we consume.
Existent trust models between media outlets, (whether traditional news outlets or social platforms) and their customers are increasingly strained, especially as images and videos can be easily manipulated, separated from their sources, and widely shared, often with misleading context, or no context at all. Add to this an increasingly polarized and agenda-driven traditional and social media landscape and it is very easy for media consumers to get lost and confused.
Additionally, recent advances in generative AI technology (like Dall-E, Stable Diffusion, Midjourney, Firefly and Chat GPT) whereby completely artificial images or text can be generated in response to an initial written prompt, have further confused things.
We recently saw a flurry of “synthetic” AI-generated images of a former US president seemingly wrestling in the street with NY City police officers as he purportedly tried to escape arrest. A startling image of Pope Francis wearing a long white belted puffer coat also made the rounds.
Both sets of images could easily have been real — except they weren’t
For many years the approach to addressing suspect images like those would have involved detection technology whereby files were uploaded to programs looking for telltale signs of manipulation. Things like inconsistent pixel structures, sloppy Photoshopping or impossible combinations of lighting sources are usually visible upon close examination.
While this approach can be useful on an ad hoc basis, especially when combined with other journalistic fact-checking techniques, it is neither scalable nor sustainable. It simply takes too long to forensically examine each image, and, in any case, bad actors will always stay one step ahead of detection software, ultimately rendering it irrelevant.
Instead, the solution lies in establishing the provenance, or origins, of digital files. Proving where, when, and how those files were created, tracking any changes that they might undergo in the editing process and ultimately sharing this information with the consumer is a hugely important way forward in helping us navigate this increasingly confusing media landscape. (Other key areas that require urgent attention are media literacy and policy-driven ethical frameworks)
In late 2019 Adobe (where I now work) launched the Content Authenticity Initiative (CAI) as a large and influential community of media and technology companies (and others) to develop and implement the open standard for provenance. The CAI community has now grown to over 1,200 members and we have made great strides introducing the concept and adoption of provenance far and wide. Our members are working to implement this technology.
Our work is focused on three main areas: capture, edit, and publish.
In the capture area we are working with camera and smartphone manufacturers to integrate the CAI technology into those hardware devices at production. The owner of each “secure capture” device will then be able, if they choose, to opt-in and activate the technology allowing provenance to be established from the moment a photo is taken, or a video or audio file recorded. Additionally, existent metadata like EXIF, XMP and IPTC which have long provided valuable (but unsecured) details like time, date, location, and other technical settings of the device used, will be secured, and made “tamper evident” using cryptography to detect if they are altered.
The next stage of a piece of digital content’s journey typically involves editing it. Here we have integrated the CAI technology into editing programs to include both Adobe software as well as some of our direct business competitors (the CAI is open source, after all). This allows for any changes made to the file (a crop or change in tonality of an image, for example) to be captured and securely stored as additional layers of metadata for later review, thus creating a secure “edit history” of the file in question.
After editing, digital files typically pass into complex content management systems and file sharing technology (email, SMS, social media platforms etc.) where the existent metadata is routinely stripped away. We are working with CMS manufacturers, news publishers and others to retain and display the underlying provenance metadata previously described.
By then sharing this information with media consumers through a simple universal interface (a clickable icon) displayed next to each published asset, we allow the viewer to examine the provenance information contained in each file and help them understand where it came from and what changes might have been made to it.
Like a lighthouse in a storm, this will help cut through the confusion and offer a trustworthy and sustainable solution to help fight mis and disinformation.
All this important metadata can be digitally stored in a variety of ways. It can be embedded into the file itself or housed in “the cloud” with a pointer from the file to the cloud. Or it can be stored using blockchain.
Underpinning all the CAI’s work is the C2PA technical standard. The C2PA (Coalition for Content Provenance and Authenticity) was created by Adobe and Microsoft in early 2021 as a Joint Development Foundation Project within the Linux Foundation. Its members are, for the most part, a subset of CAI members and are tasked with developing technical specifications for establishing content provenance and authenticity. The specifications are informed by scenarios, workflows and requirements gathered from industry experts and partner organizations, including the CAI.
The CAI then takes the C2PA standard and uses it to build open-source provenance tools accessible to everyone. In June of 2022 we released three such tools towards that end, all available on our website or through GitHub.
Beyond news media use we expect provenance to become something foundational to a great many things digital. Every day we see myriad images on e-commerce websites, travel-booking platforms, dating apps etc. and we take great leaps of faith based on those images. The provenance solutions we are building have already attracted the attention of those sectors and many others interested in authenticity.
On the AI front we are working with Generative AI providers to automatically have their content labeled as such using the CAI technology. One such provider, Stability.ai has publicly committed to doing this and other major AI suppliers are working towards this goal. Already, all output from Adobe Firefly, our recent foray into Generative AI, is clearly labeled as having been created by a machine.
Screen capture showing the “Content Credentials” of an AI generated image.
Firefly is in beta right now but there is a clear label viewable alongside downloaded files.
In addition, we have developed and laid the groundwork to implement a secure “Do Not Train” label that creators can use to prevent their content from being used in the massive data training sets that Generative AI providers have used to train their models (Adobe Firefly only uses legally cleared data training sets).
While the road ahead is long, we have made steady progress and will soon see “secure capture” hardware on the market in the form of cameras and smartphones. The CAI technology is already in Photoshop and Lightroom (beta) and we will soon see it in other Adobe and non-Adobe editing tools.
Additionally, we are close to real-time implementation by some major news media organizations (AFP, EPA Images, WSJ, Globe & Mail) and are working with social media platforms towards that same goal. The news organizations who are members of the CAI (including the BBC, the Associated Press, and The Washington Post) are committed to implementing this technology over time through attaching CAI metadata to their content.
We believe that provenance is essential on the path forward and will require everyone to actively participate. Success will be defined by ubiquity, where consumers of digital content come to expect provenance information to be displayed alongside image, video, audio recording and other file types to prove what is real, as opposed to detecting what is false.
This provenance information combined with robust media literacy education, supported by policy (and bolstered by ad hoc detection and fact checking) are the foundations we are building to withstand the onslaught of fast proliferating misleading information.
Though the CAI is only at the beginning of our provenance journey and media implementation examples are works in progress, the stakes are high. We cannot let mis and disinformation erode trust, endanger creative and digital economies and even democracy itself. Our work is a call to action to all those individuals and companies who value authenticity. Passivity is not an option.
Santiago Lyon is an award-winning conflict photographer and former VP for Photography at The Associated Press.
Creative Commons Attribution 4.0 International License (CC-BY 4.0)
