Truepic BlogHow Can We Tell What is Real Online? Provenance and Detection Tools for Braving the Uncanny Valley

It’s been just over two years since researchers at Berkeley found that humans have a 50/50 chance of being able to tell if headshots were AI-generated or human. Since then, we have seen and heard a wide range of hyperrealistic synthetic images, video, and audio that unmistakably enter the ‘Uncanny Valley’ – that unsettling place where something that is not human – like a deepfake – seems decidedly human such that we can no longer trust our perception of reality.

In this nebulous digital world, we need technologies to help us navigate the authenticity of digital content. Two technological approaches have emerged: provenance and detection.

Provenance, a robust solution that begins at the point of origin.

Provenance has long been used in the art world to authenticate the origin of artistic works by documenting how, when, and where something is made. As the use of generative AI grows, we expect the sheer quantity of synthetic content online to rapidly outpace and ‘crowd out’ authentic content. Knowing what is human vs. computer generated will be paramount for trust on the internet. A provenance approach focuses on a proactive solution to this challenge by establishing the origin of content as early as possible in the value chain of digital media and displaying that origin to viewers through popular publishing platforms, including news and social media.

Digital content provenance works by cryptographically sealing the known origin of a file into its metadata. The most widely accepted approach to media provenance has been popularized by the C2PA standards body, and implements a service called Content Credentials. These credentials, once sealed, are later viewable when content is published so that we can see, for example, if the origin of an image was generated by a computer (artificial intelligence) or captured by a traditional camera. Having a unified and standardized way to establish and display provenance information benefits consumers by letting them know what they are interacting with online, much like nutrition labels provide us with information about what is in the food we consume every day.

Digital content provenance is a tamper-evident, proactive measure to establish and disclose the origin and history of digital content. Similar to a protective seal on a consumer product, if the provenance ‘seal’ on a piece of digital content is intact, you can know that that content has not changed since the seal was applied. Otherwise, you would be able to tell if the seal had been broken. For decisions of high consequence, this proactive guarantee is often imperative. The cryptographic assurance that an image has been unchanged since it was captured or created eliminates some uncertainty for decision makers evaluating that content. Truepic has worked with hundreds of enterprise customers to help them capture tens of millions of authentic images and videos with verified provenance that are used to make critical business decisions. Truepic is also working with AI companies, hardware providers, and others to help integrate digital content provenance into their systems for an even wider distribution of Content Credentials.

Content Credentials are designed to be interoperable so that when digital content inevitably travels between compliant tools and platforms, the content’s provenance stays intact. That said, the most significant drawback to the provenance approach is that these new provenance technologies and standards have to be implemented at critical junctures in the in content value chain, most notably the point of content creation. Interoperability allows for significant scaling of the provenance approach, but scaling also requires broad adoption across the tech ecosystem such that consumers start seeing Content Credentials on a daily basis.

Detection, a frictionless solution that can be applied to any content.

Detection of AI-generated content, on the other hand, is probabilistic and is applied reactively. It gives decision-makers the probability that something was or was not AI-generated based on pattern analysis. For example, a detection tool might tell us that there is an 80% likelihood (plus or minus a margin of error) that something was AI-generated. 80% is high, but it is also far from a guarantee.

Detection can be an extremely helpful directional indicator, especially if content of unknown origin is circulating online without Content Credentials and quick action needs to be taken to determine the nature of content. One of the primary benefits of detection systems is that they do not require broad, ecosystem level adoption to be useful. Rather, they can be applied as a point solution, often via APIs or on-prem systems, and applied only to content that is called into question. This yields immediate benefit to decision makers who are looking to perform a probabilistic analysis on the authenticity of content.

Where detection falls short is on certainty. Because of the editable and ephemeral nature of metadata in image, video, and audio file formats – detection can never provide a definitive answer about the origin of content, or whether data has been manipulated along the way. For example, simply changing the time and date on your smartphone and capturing an image imprints a “Camera Original” time and date that are incorrect. Detection systems are not, and will never, be capable of determining this type of “cheapfake.” And, as AI systems progress in quality, it will be increasingly difficult for detection approaches to keep up. Given the substantial commercial incentives for highly sophisticated and widely accessible content generation, this dynamic is further fueled by far more capital investment in generative capabilities than detection capabilities.

Only individual decision makers can know what is right for them and their use case, but there are instances when misidentifying something as AI-generated could be quite harmful. Even if a detection tool was accurate 99% of the time at determining if content was AI-generated or not, that 1% gap at internet scale is massive. Experts estimate that in 2023 alone, 15 billion images were generated by AI. When all it takes is one convincingly fabricated image to degrade trust, 150 million undetected images is a pretty unsettling number.

Provenance + Detection, better together.

The reality is that we need both content provenance and detection technologies to mitigate the risks of synthetic content online today. Only provenance provides cryptographic assurance about how a piece of content came to be, but not all media has provenance. While provenance continues to become more widely adopted, detection can provide an interim solution to help decision makers evaluate digital content that does not have provenance.

Provenance is becoming a proactive best practice across many content creation tools. For example, now when you use DALL·E 3, a Content Credential is automatically and proactively sealed into the metadata of each media file to show that the image was created using DALL·E 3. If that Content Credential is stripped (advertently or inadvertently) from the file, detection will then have a role to play.

Adoption of Content Credentials will continue to grow given the need for transparency into the origin of content. And detection tools will continue to add value as the ecosystem moves through waves of provenance adoption. Together, these two technologies can help decision makers better understand the authenticity of content as we enter a completely new era for digital media.

Request more information